Technological Forecasting & Social Change 174 (2022) 121280 Contents lists available at ScienceDirect Technological Forecasting & Social Change journal homepage: www.elsevier.com/locate/techfore Tracing the evolution of service robotics: Insights from a topic modeling approach☆☆ Ivan Savin *, a, b, c, Ingrid Ott d, e, Chris Konop d a Institute of Environmental Science and Technology, Universitat Autónoma de Barcelona, Spain b Graduate School of Economics and Management, Ural Federal University, Yekaterinburg, Russian Federation c ICTA-ICP Building (Z), UAB Campus, Cerdanyola del Vallés, Barcelona, Spain d Chair for Economic Policy, Karlsruhe Institute of Technology, Germany e Kiel Institute for the World Economy, Kiel, Germany A R T I C L E I N F O A B S T R A C T JEL classification: Taking robotic patents between 1977 and 2017 and building upon the topic modeling technique, we extract their C11 latent topics, analyze how important these topics are over time, and how they are related to each other looking at C15 how often they are recombined in the same patents. This allows us to differentiate between more and less O33 important technological trends in robotics based on their stage of diffusion and position in the space of O34 knowledge represented by a topic graph, where some topics appear isolated while others are highly inter­ Keywords: connected. Furthermore, utilizing external reference texts that characterize service robots from a technical Knowledge diffusion Latent dirichlet allocation perspective, we propose and apply a novel approach to match the constructed topics to service robotics. The Networks matching procedure is based on frequency and exclusivity of words overlapping between the patents and the Patents reference texts. We identify around 20 topics belonging to service robotics. Our results corroborate earlier Topic matching findings, but also provide novel insights on the content and stage of development of application areas in service robotics. With this study we contribute to a better understanding of the highly dynamic field of robotics as well as to new practices of utilizing the topic modeling approach, matching the resulting topics to external classifications and applying to them metrics from graph theory. 1. Introduction While IR still diffuse via intensified application in the manufacturing sector (‘automation deepening’),1 SR continuously capture new domains Robots are increasingly supporting humans both at work and in their (‘automation broadening’). Not always, but often, SR are mobile. Some private life. While the use of industrial robots (IR) has a long standing of them are fully automatic or even autonomous.2 Due to the importance tradition in the manufacturing industries, service robots (SR) are a more of services in value creation and given the prevailing low level of recent phenomenon. Latest advances in artificial intelligence and ma­ automation in this field, future diffusion of SR is expected to have far- chine learning enable robots to sense and respond to their environments reaching implications for overall economic productivity. Due to the so that they can also be used outside secured production environments. recent COVID-19 outbreak, this diffusion is experiencing a further boost ☆☆ Financial support from the Helmholtz Association (HIRG-0069) is gratefully acknowledged. Ivan Savin acknowledges support from the Russian Science Foundation [RSF grant number 19-18-00262]. This work has benefited from presentations at workshops in Kiel, Strasbourg, Karlsruhe, the EMAEE conference in Brighton and the EAEPE conference in Bilbao. All remaining shortcomings are our responsibility. * Corresponding author. E-mail address: Ivan.Savin@uab.cat (I. Savin). 1 We use the term ‘automation deepening’ in the sense of intensified robot density, i.e. an increasing ratio of robots over employees. This is slightly different to Acemoglu and Restrepo (2018) who consider automation deepening as improvements of existing machinery. 2 Following the international ISO standard 8373:2012, a robot is an actuated mechanism programmable in two or more axes, with a degree of autonomy, moving within its environment, to perform intended tasks. Given the current state and sensing, autonomy means the absence of human intervention by performing the task. The classification into industrial robot or service robot is done according to its intended application. The International Federation of Robotics (IFR, IFR (2018a)) bridges this standard to robot investment data broken down by application areas and industries. See Appendix A providing technical definitions of robots, robot systems, and the most important industries in which IR are used. https://doi.org/10.1016/j.techfore.2021.121280 Received 16 March 2021; Received in revised form 3 August 2021; Accepted 13 October 2021 Available online 1 November 2021 0040-1625/© 2021 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 to facilitate more physical distancing in healthcare, logistics, tourism robots in already automated industries (as e.g. the automotive and the and other spheres (Chen et al., 2020; Yang et al., 2020; Zeng et al., electrical industry) or the increased penetration of robots in countries 2020). with hitherto still only few industries being automated. Driven by In this paper we aim to obtain a comprehensive understanding ser­ increased calculation power, a decline in hardware costs, the use of vice robotics and of the associated technologies enabling their evolution. lighter materials and technological progress in complementary tech­ To this end, we use patent data, the most complete description of nologies, such as cloud computing or artificial intelligence (AI), robot technological development. In line with the recombinant growth of use continuously expands into new fields that until today have been knowledge, patent texts typically are composed of many technologies characterized by almost complete absence of automation. (Youn et al., 2015). To decompose these complex documents into However, most of research on robots in economics has still the focus distinct technologies, we take advantage of the topic modeling tech­ on IR, mostly associated with labor market implications, task perspec­ nique developed at the intersection of machine learning (ML) and nat­ tive and addressing questions like the relationship between humans and ural language processing (NLP). In particular, we use not just their short machine being complementary or substitutive (e.g. Acemoglu and abstracts that are too superficial to get an understanding of technologies Restrepo (2019) or Graetz and Michaels (2018)), covering national incorporated in the patent, but full patent descriptions – or more pre­ perspectives (e.g. Dauth et al. (2018), Dauth et al. (2019) for Germany, cisely their non-technical summaries – in the robotic industry between Acemoglu et al. (2020) for France, Bessen et al. (2020) for the 1977 and 2017. Then we extract the latent topics capturing different Netherlands). Agrawal et al. (2019) focus on prediction of labor market technologies used in patents, analyze how important those topics are implications. A more historical dimension has been taken by Atack et al. over time, and how they are related to each other through their (2019) or already Mansfield (1989), who analyze diffusion of robots in co-occurrence in patents. the US and Japan, and by Cheng et al. (2019) for China. Baldwin and In doing so we contribute to the existing literature in several ways. Forlsid (2020) are among the few economists that also explicitly relate First, we apply modern methods of NLP and ML to exploit fine grained their work to services and robots in a globalized context. Autor and technological information included in unstructured textual data of pat­ Dorn (2013) point out that it is crucial to distinguish between service ent documents and identify the optimal number of topics using estab­ and manufacturing occupations. However, explicitly addressing service lished criteria of perplexity, exclusivity and coherence. We end up with robots is still rare, especially in the scientific literature. Recent robot 380 topics, carry out robustness checks for 190 topics and analyze how patent analysis can be found by Analytics (2018) or EPO (2017), while importance of those topics was changing over time. Second, we develop Goeldner et al. (2015) are among the first who analyze patenting activity and apply a novel method for matching topics to SR based on an external in a specific application field of SR, namely care. text corpus provided by IFR that classifies SR technologies into 16 Modern robots are flexible, easy to operate and becoming able to application areas and 49 sub-areas. We use results of this exercise to navigate autonomously, even in unstructured environments. As a classify topics to different SR areas. Third, based on the results of our consequence, aside from using robots within clearly defined environ­ textual analysis we construct a complex graph using cosine similarity ments such as factories, provision of services is increasingly becoming between the topics and identifying significant edges in this network. automated. The associated spread of robot application may be labeled as This approach allows us to overcome the popular practice of analyzing automation broadening providing potential for huge productivity gains topics in isolation. Instead, we can trace robotic transformation from a from automation also beyond the manufacturing industries. system perspective: apply metrics from graph theory, understand the According to the ISO classification, SR perform useful tasks for mutual relationship of discovered topics, distinguish between central humans or equipment excluding industrial applications.3 SR are further (enabling) and peripheral (application) topics, discover communities of differentiated into two types, namely those for private use that are hardware- and software-oriented topics, and how those were changing operated by a lay person and those for professional use that are usually over time. It is important to stress that all these steps are independent operated by a professional or properly trained operator. Due to the from structured patent data (e.g., patent classes, concordances) and multitude of forms and structures as well as application areas of SR, it is expert bias. The entire process from topic identification until matching not easy to delimit them from IR. Robots in logistics are a prominent to SR is data-driven and thus is unbiased by subjective expert judgment. example of such an unclear assignment. They are used in non- Another important strength of our approach is that apart from repli­ manufacturing environments, such as logistic centers, hospitals or cating results that could have been achieved by applying existing met­ warehouses but also to transport parts within factories.4 rics (e.g. rising importance of medical robots, shift in popularity from The sketched technological transformation is being mirrored in the hardware to software technologies), we are able to look inside the work of the IFR’s statistical department. It bridges robot classification content of each particular topic (e.g. study technologies in surgery ro­ from the ISO standard and robot markets by collecting data on world­ botics) and their position in the complex space of knowledge comprised wide robot investments, sales and stocks differentiated by IR and SR; the by robotic patents (e.g. how central they are and what topics they are latter ones split up into robots for domestic and for professional use. For connected to). the heterogeneous domain of SR, the IFR provides a detailed technology The remainder of this paper is organized as follows. Section 2 pro­ breakdown covering 16 areas and 49 sub-areas (Table 1).5 In 2017 the vides some background information on service robotics and topic most important markets of SR have been in the fields logistic systems, modeling as a methodology to deal with patent data. Section 3 describes defense applications, public relation robots, field robots (especially our data and methods. Section 4 presents the results. Section 5 discusses milking), powered human exoskeletons and medical robots (IFR, policy implications and Section 6 contains concluding remarks. 2018b). According to IFR (2019), sales volumes of logistic robots have been the key driver of the SR markets also in 2018 followed by SR 2. Background 2.1. The rise of service robotics 3 See ISO 8371:2012, 2.11 (private use; synonyms are personal or domestic use) and 2.13. (professional use). Research on automation took off with the introduction of mass- 4 In 2018,7′ 700 units of logistic robots have been used in manufacturing production manufacturing methods. Since then, automated machines while103′ 000 units have been utilized outside of factories, e.g. in warehouses, have been continuously transformed to today’s multipurpose industrial logistic centers and hospitals (IFR, 2019). robots, i.e. robots that are able to adapt to a different application 5 The IFR provides information differentiating between IR and SR and cate­ without alteration of the mechanical system. Until today, key drivers of gorizing by application area, industrial branches, robot types or geographical robot diffusion have been automation deepening, i.e. an intensified use of region, and across time. 2 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 Fig. 1. Investment dynamics for IR and for selected SR application areas. Data is normalized so that the investment level in 2016 equals one. In both panels the red line refers to the level of one. Data for IR dynamics (until 2018) based on IFR data base; data on SR dynamics based on IFR (2018b) and IFR (2019). Starting from 2018 the investment data is predicted by IFR. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) applied for inspection and demolition tasks or medical service, most of To better understand the future development of robotic technologies, them being surgery robots. Considerable markets exist in field robotics we take a precise look at patenting as an early indicator of future in­ (e.g. milking), and big potential is seen in the application of robots for vestment in the respective area. In robotics, there is a strong correlation plowing (agriculture). Another fast growing market is public relations between patenting and investments (Fig. 2). This correlation pattern robots which are used to provide information in public spaces or shops also holds if one zooms into major geographical regions or robot ap­ thereby also increasingly utilizing humanoid robots. SR for private use plications (like handling or soldering). It is also worthwhile to stress that are dominated by sales of lawn mowers or floor- and window-cleaning robot patenting has been less susceptible than investment during and in robots, together with robotic toys and games. Technological advance­ the aftermath of the global financial crisis. These observations together ments in robot mobility also drive the adoption of robots in the (still with the aforementioned description of the transformation of robotic small) market for elderly and handicap assistance. Finally, SR for do­ technologies is the starting point of the subsequent empirical analysis of mestic use are reported separately. Their unit value is only a fraction of patents with a special - though not exclusive – focus on emergent SR. that of the many types used for professional use. Since they are produced for mass markets, they follow different pricing and marketing channels. The above summary illustrates the strong heterogeneity not only 2.2. Patent data analysis and use of topic modeling between IR and SR, but especially also within SR applications. They also differ with regard to unit price,6 life span7 and investment dynamics. Patent data is widely considered as the most complete description of While IFR reports reliable data on IR since 1993 onward, comparable innovations. Patents cover a multitude of technical fields over a long time series are not available for SR. A central reason for this is incom­ period of time, and thus reflect structural changes in those technologies. plete data due to the high fluctuation of providers in this dynamic There are specific criteria for an invention to be patented, which creates market segment. Another problem is that the reported data is difficult to an objective standard as to what counts as an invention. Information on compare over time. Since the focus of our analysis is on the trans­ patent citations has been long used in economics as an indicator for formation of robot technology, data on absolute investments play only a patent quality (Trajtenberg, 1990), knowledge diffusion (Jaffe et al., minor role. Instead, the dynamics of the specific SR areas, their evolu­ 1993) and its obsolescence (Jaffe and Trajtenberg, 1996). tion over time and their embedding within the robot technology space The use of patent citation data, however, has been always prob­ are important. In the following, we present the data on SR from IFR (IFR, lematic due to different practices of patent citation across patent offices 2018b; 2019). Recent investment dynamics as well as forecasts for with citation records being often incomplete (Michel and Bettels, 2001) 2019–2022 is approximated based on the IFR’s annual information on and, most important for the reasoning of the paper at hand, the fact that sales. visualizes this by setting the unit sales of the respective robot area patent examiners may have added extra citations not actually relevant in the year 2016 to one and showing the dynamics of factual and pre­ for the inventions (see Alcácer and Gittelman (2006) for USPTO evi­ dicted data. Fig. 1(a) reports IR investment dynamics beginning in 1993. dence). This bears the risk of distortions (expert bias) if one wants to Since it has a longer history and higher levels of installed units, their analyze the development of technology over time. Furthermore, cita­ development is less dynamic and even stagnates in 2018. Investment tions are usually made on the basis of legal considerations rather than dynamics of selected SR areas are displayed in Fig. 1(b) demonstrating from a technological perspective. Besides, patent classifications change SR growing many times faster than IR with the leading areas being lo­ over time (Lafond and Kim, 2019) and are also hard to compare between gistics and medicine. different patent offices. All this is challenging the use of structured in­ formation from the patent data, which is considered as the most comprehensive and accurate description on knowledge flows. What is 6 Medical and underwater robots sometimes cost several hundred thousand most striking is that a large body of information unstructured textual USD (IFR, 2018b), while toy robots often only a few hundred USD description contained in patents) has been long disregarded in economic 7 For industrial robots, the average duration of use is about 11 years, un­ research. derwater SR are utilized up to 10 years, while defense robots may have a life With the rising power of modern computers and availability of the cycle of one single operation. great amount of data, however, our choice of instruments to extract 3 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 Fig. 2. Robotic patents granted at USPTO (1077–2016) and robot investment worldwide (1993–2016). robotic industry has never been well studied with respect to the topics prevailing in the related patents.8 Apart from patents, TM has been widely applied to other type of textual information. Many studies focus on scientific literature pub­ lished either in a specific peer-reviewed journal across many themes (De Battisti et al., 2015; Griffith and Steyvers, 2004; Lüdering and Winker, 2016; Savin and van den Bergh, 2021) or all economic articles stored in a given database (such as JSTOR, Ambrosino et al. (2018)). Furthermore, some studies focus specifically on literature published on the theme of information security (Chang, 2016) or bioinformatics (Liu et al., 2016). While some of these studies look on abstract only (like De Battisti et al. (2015)), others take full texts of the academic papers into analysis (Ambrosino et al., 2018). Another popular field of application for TM is news articles. Those can be either specialized financial news (taken, e.g., from Dow Jones Newswires Archive (Larsen and Thorsrud, 2019) or financial analyst reports (Huang et al., 2017), policy statements and website articles related to climate change (Farrell, 2016) or publications Fig. 3. Length of patent non-technical summaries. from social media like Twitter (see, e.g., Chae and Park (2018)). Finally, topic modeling has been recently applied to survey open-ended ques­ tions (Roberts et al., 2014; Savin et al., 2021; 2020; Tvinnereim et al., 2017). This illustrates the generality of the approach that can be applied information from the textual data is rapidly growing. In this study we to very different type of data in terms of size and content. apply the so-called topic modeling (TM) approach to gain new insights about knowledge structure and dynamics in the robotic industry. TM is a 3. Data and methodology clustering approach for textual data aimed to identify meaningful topics in text data, analyze trends in topics, (re)classify and annotate docu­ 3.1. The robotics patent data set ments (Blei, 2012; De Battisti et al., 2015). TM posits the idea that patents represent a combination of topics, where each topic is a prob­ We focus on robotics utility patents granted at the United States ability distribution over a fixed vocabulary. While patent texts are Patent and Trademark Office (USPTO), where we take the first filed observed, the topics and their distributions are not and treated as latent patent of a family as representative.9 Our search strategy is based on variables. TM, thus, uncovers topics latent in a collection of patents and having the truncated keyword *robot* in the title or abstract of the identifies which composition of them best accounts for each patent. The patent10 or being classified in one of the CPC classes concordant to the advantage of TM over, e.g., keyword analysis is that i) it allows words to have different meanings depending on their contexts; and ii) it is data-driven: one does not need to specify topics a priori but generates 8 The only exception perhaps is by Kim et al. (2016) focusing on a small them from the data. More details on topic modeling related to our data fraction of robotic patents devoted to humanoid robots using less than 1000 set are presented and discussed in Section 3.3. patents from USPTO and analyzing their titles and abstracts only. TM has been recently applied to patent data in a number of studies. It 9 A family of patents refers to a group of patents that are issued in different has been used for patent (re)classification into product and technology countries for the same invention to obtain patent protection. We refrain from sub-classes to later explore technological convergence and geography of multiple counts of the same invention within the patent families since the size innovation in the photovoltaic technology between the US states (Ven­ of the family does not affect the technical facets of the invention. If we were to ugopalan and Rai, 2015); for identification of emerging topics among include every family patent in the analysis, this would distort the text corpus in triadic patent families (patented in the US, EU and Japan, Lee et al. favor of the larger families. 10 (2015)); for detection of pioneering patents introducing new topics We initially applied the same search strategy on the full texts of patents at (Kaplan and Vakili, 2015); and for prediction trends in patent topics the USPTO database. However, when checking the result it became obvious that this strategy yielded too many false positives. (Chen et al., 2017; Suominen et al., 2017). To the best of our knowledge, 4 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 Fig. 4. Data cleaning procedure . USPC class 901 (robots).11 This identification strategy resulted in22′ 927 3.2. Patent data pre-processing patents for the period 1977–2017 (see Fig. 2; black line for the evolution of patents over time). The first step of our pre-processing pipeline is lemmatization, where Patent texts consist of several parts which can be considered for inflected word forms are transformed to their dictionary form. Contrary textual analysis: title, abstract (typically less than 150 words), patent to stemming the words (i.e. cutting words to their word stem, or root description and patent claim. While title and abstract are too short to get form), lemmatization is done by identifying the intended part of speech any comprehensive understanding of technologies incorporated in the and meaning of a word in a sentence. We tried both approaches and patent, the claims concentrate on the differences of the technical novelty while both of them have limitations, topic keywords generated after compared to the prior art. Claim texts are thus also with limits suitable to stemming are harder to interpret as they usually contain only parts of reflect the content of new technological knowledge contained in the the words. Since multiple words with different meaning may be grouped patents. The patent description includes a non-technical summary as into the same stem, using lemmatization allows us to preserve inter­ well as a technical description. For our analysis, we skip the latter and pretability of our results. In particular, we use WordNetLemmatizer concentrate on the non-technical summary. This has the advantage that from the Natural Language Toolkit (NLTK) in the Python programming we avoid parts of text with too specific language and formula description language. that is less suitable for textual classification. On average, the length of Second, after lemmatization, we replace acronyms by their original the non-technical patent summary descriptions is 667 words (see Fig. 3 notation. This step is very important, as patents, similar to academic for descriptive statistics on the data). In total, 15.3 million words are literature, contain a lot of acronyms introduced once at the beginning of contained in non-technical summaries of our patent sample.12 patent description and used consistently in the document. Overall we Besides, each patent includes some structured data including classi­ find 1502 distinct acronyms, resulting in a total of18′ 628 replacements. fication to one or several patent classes. Existing concordances make it By identifying and replacing those acronyms with their actual meaning possible based on the evaluation of the CPC classes of patent specifica­ we solve several problems:13 (i) we reduce the chance that the same tions to assign the patents to industrial classification, e.g. by means of acronym used in different documents and actually standing for different NACE or ISIC concordance classifications Eurostat (2014). Similarly, word combinations will be recognized by our approach as the same patents can be assigned to technology fields (by means of WIPO tech­ word. Thus, although we find 1502 acronyms, there are only 1028 nology area concordance matrices WIPO (2008) with a strong bias to­ unique acronyms, meaning 474 acronyms have the same sequence of wards manufacturing industries. Based on this information it is possible, characters but a different meaning behind them. This implies that for example, to visualize the development of medical instruments or the without replacing acronyms we would have biased our results by not increasing importance of ICT components by exploiting structured in­ distinguishing the terms with non-unique acronyms. (ii) we consider­ formation of our robotic patent dataset (see Fig. A12 in A.2). Applying ably increase our ability to understand the formed topics later on as both concordances illustrates a shift over time from mechanical com­ instead of an acronym we can see the full expression. ponents to computers, an increase in medical instruments, a decline in Third, this is followed by transforming all characters to lowercase machine tools while measuring, testing and navigation remains rather and then removing stopwords (i.e. and, or, the) as well as all non- stable. We will later illustrate that our approach is capable to replicate characters.14 This step is standard for NLP (see, e.g., Grün and Hornik this pattern. (2011)). Note that we do that after replacing acronyms and lemmatizing words to minimize the amount of information that may be lost (e.g., 11 See https://www.uspto.gov/web/patents/classification/cpc/pdf/us901toc pc.pdf. A full list of concordant CPC classes can be found at: https://www. 13 The replacement is done by an n-gram with the words being connected in uspto.gov/web/patents/classification/cpc/html/us901tocpc.html##statTable. 12 one with ”_” symbol. We also conducted our analysis on the full descriptions (both non-technical 14 In particular, the NLTK’s list of English stopwords has been used extended and technical description) as well as the technical descriptions only. With the with a set of custom stopwords we created while testing different LDA models same number of topics chosen the results are fairly similar concerning the topic (see Section B.1 for more details. content and are available upon request. However, as technical summaries contain more formulas and other scientific notation, by concentrating solely on non-technical summaries we avoid supplying our NLP analysis with non-textual data that is hard to clean automatically. 5 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 short acronym of two letters only).15 the extent to which frequent words from the same topic tend to appear in Fourth, we create bi-grams out of words commonly occurring the same patents. Exclusivity analyzes whether popular words from one together. Bi-grams are created using gensim phrase module in Python. topic have low likelihood in all other topics. Thus, we follow Roberts For each bi-gram, the normalized pointwise mutual information (NMPI) et al. (2014) in arguing that semantically interpretable topics should score has been applied (see Bouma (2009)). In simple words, the NMPI consist of words that tend to co-occur within documents, and that their score measures how often any two words appear together versus how top keywords are unlikely to overlap with keywords from other topics. often they appear in text separately and forms out of those appearing More details on the different metrics used to determine the optimal predominantly together a bi-gram with a ”_” symbol (see Appendix B for number of topics are provided in Appendix B. the exact formula of the score and its explanation.). Specifically, a 0.5 Applying the topic modeling approach on the patent data results in threshold value to form bi-grams is used. The flow of the pre-processing two matrices. The first is the matrix of probabilities observing a word steps on our textual data is summarized in Fig. 4. given the topic, while the second is the matrix of topic prevalences in each of the documents. The information in these two matrices is the basis for labeling topics, visualizing them and conducting further anal­ 3.3. Topic modeling ysis such as matching the formed topics to existing SR classification and analyzing mutual interdependence between topics by means of network While there have been many algorithms developed for topic analysis. These two steps of analysis we address in the following sub­ modeling (for a recent overview see Liu et al. (2016)), in the following sections while subsections 4.2 and 4.3 provide the corresponding results. we concentrate on the most commonly used approach known as latent To illustrate the content of each topic and evolution of their impor­ Dirichlet allocation (LDA, Blei et al. (2003)). LDA as a probabilistic tance over time, we will henceforth use word clouds and diffusion curves. modeling approach assumes that each word in a document is generated Word clouds represent the 30 most frequent words given the topic with in two step process. First, assuming that each document has its own topic font size capturing how likely these words are given the topic. The distribution, a topic is randomly drawn from that distribution (the lightness of the color in turn reflects how exclusive this word is Dirichlet distribution). Second, assuming that each topic has its own compared to all other topics: the lighter the color, the more exclusive is word distribution, a word is randomly drawn from this word distribution the word. Diffusion curves represent the stage of technology adoption of the topic selected in the first step. A document is essentially a result of captured by our topics in the corpus of robotic patents over the period multiple repetitions of the statistical model consisting of those two steps 1977–2017. To produce such diffusion curves, we follow Lenz and the generative process defined by a joint probability distribution over the observed (documents and words) and hidden (topics) variables. This Winker (2020) in quantifying the probability that a given topic appears in the corpus of patent texts for each year ensuring that that these statistical model reflects the intuition that documents exhibit multiple topics, where each document exhibits the topics in different proportion. probabilities for any period sum up to one. In addition, we smooth the curves by estimating such probabilities not just for a single year but for a The LDA algorithm discovers the topic distribution for each document and the word distribution of each topic iteratively, by fitting this five-year time interval around a given period (see Appendix B for more details). Finally, to ease reading these plots, we classify the curves into two-step procedure to the observed documents until it finds the best set of variables that describe the topic and word distributions. Effectively, ’rising’, ’falling’ or ’in-between’ with green, red and blue color, respectively, depending whether the topic between 1977 and 2017 was LDA as a Bayesian model computes the conditional distribution of the hidden topics given observed data identifying the posterior distribution predominantly rising or falling in its prevalence, or none of the two (see Appendix B for details). of the latent topics in a collection of documents. Similar to cluster and principal component analysis, LDA reduces the dimensionality of linguistic data from words to topics based on the co- 3.4. Topic matching location of words in a collection of documents to infer the underlying topics in those texts and the weight of each topic in each individual While methods to form topic models have seen a rapid development document. According to Blei (2012, p. 79): “This can be thought of as over the last two decades, labeling and interpreting topics remains ‘reversing’ the generative process - what is the hidden structure that largely an ad hoc procedure. Most studies try to provide a concise label likely generated the observed collection”. For a more formal description consisting of frequent and exclusive words summarizing the essence of of LDA, see Appendix B. the topic. Apart from looking on single terms associated with the topics, The number of topics in the model affects the interpretability of its some papers in addition take illustrative documents where the preva­ results. Setting the number too low can result in topics that are too broad lence16 of respective topics is highest to demonstrate how topic labels fit and ambiguous. Conversely, setting the number too high may introduce in the context (see Roberts et al. (2014); Savin et al. (2020); Tvinnereim uninterpretable topics that pick out idiosyncratic word combinations and Fløttum (2015)). (Griffith and Steyvers, 2004). While one can follow computational lin­ In a recent survey, Boyd-Graber et al. (2017) distinguish between guistic literature and calculate perplexity score (fitness) of the LDA labeling methods that only use internal information from the topic model based on different numbers of topics (Blei et al., 2003), most of model against those that also use external knowledge sources. Labeling the time this number is chosen ad hoc and relatively small to maintain with exclusively internal information looks for phrases with high topic interpretability of results. It is by all means easier both to label and prevalence that well summarize the documents making them good overview 10 and not 100 topics. candidates for labels. Labeling with external knowledge sources either In our study we will choose the number of topics that aims to opti­ aims to weight words in a topic as prospective labels putting more mize not just perplexity score of the model, but also semantic coherence weight on words that are hypernyms17 (which is assessed through and exclusivity of the topics. In simple words, the perplexity score is external word library) and that co-occur often with other words from goodness of the LDA model trained on one part of the data to predict the that topic, or try to retrieve labels of the documents underlying the word distribution on the other part of it. Semantic coherence quantifies topics (if available), and form topics in line with those labels. Applica­ tions of such automatic topic labeling are very scarce (see e.g. Newman 15 Note that we decided against applying term-frequency inverse document 16 frequency filter, as we could delete too many relevant words necessary later for Related to our patent data set, topic prevalence refers to the degree (be­ our topic matching. Since we used for our simulations the high-performance tween 0 and 1) to which a patent document belongs to the respective topic. 17 computing cluster BwUniCluster https://wiki.bwhpc.de/e/Category:BwUniCl Hypernyms are words which meaning includes a group of other words, e.g. uster_2.0, this was feasible in terms of computational time. related to the word dog a hypernym would be pet. 6 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 Fig. 5. Perplexity, coherence and exclusivity for LDA models with different number of topics. Fig. 6. Deriving a threshold value for topic matching. Note: The red line shows how many topics are matched in total (i.e. including the same topic being matched to two or more SR sub-areas) given the threshold value, while the black one shows how many topics are matched uniquely to one SR sub-area as listed in Table 1. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) et al. (2010)) demonstrating a large room for improvement in methods matching, one being the lists of words associated to each topic with that can objectively label topics or match them to some external different probabilities19 and the other one being the descriptions of SR classification. areas from IFR, which serve as reference text. Then we look for words One of the major contributions of the paper at hand is the develop­ that appear both, in the topics and in the SR description, labeling them ment of a matching algorithm that allows for automatic identification of as ’candidate words’. Based on these words we calculate for each topic a SR topics, i.e. independent of subjective judgment (i.e. domain knowl­ ’topic matching score’ for each of the existing 49 SR sub-areas capturing edge of experts), specifically at the level of SR sub-areas in Table 1. To the extent to which they overlap. These topic matching scores tend to be the best of our knowledge, we are the first to propose such a method for higher if words from a given topic have a higher frequency (i.e. proba­ topic assignment to a specific technology field based on patent data. bility belonging to this topic) and exclusivity (i.e. low probability of For this, identification of an appropriate reference text distinguish­ appearing in any other generated topics) and, symmetrically, appear ing between SR areas based on textual information from the topics more often in the description of the given SR sub-area and less often in resulting from LDA is essential. In our case we rely on textual description other SR sub-areas. Furthermore, since we use only SR descriptions in of SR technologies as provided by the IFR SR report 2018 (IFR, 2018b, the reference texts (and not for example, descriptions of other techno­ chap 3, pp. 48–270) containing a detailed textual description of SR logical fields such as IR and beyond, we additionally reduce the topic technologies at the sub-area level).18 matching scores for those words, which are more common in the More precisely, we have two different textual sources relevant for 19 As was mentioned earlier, the main output of LDA is a matrix of size the 18 We also tried to apply the matching procedure on the level of 16 SR areas. number of topics times the number of words in the corpus, where each word is We prefer, however, to look on the less aggregated classification level to attributed to every topic with different probabilities. attribute topics more precisely. Knowing the SR sub-area, it is straightforward to identify the corresponding area. 7 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 Table 1 scores that allows us to uniquely classify the largest number of topics to SR applications: 16 areas (bold) and their 49 sub-areas as defined in IFR (2018b). SR sub-areas. This last step is data-driven, i.e. instead of choosing the Service Robotics threshold externally we let the data decide what should be the value of the matching score to assign a topic to any SR sub-area. In Appendix B for private use for professional use we explain all steps of the matching algorithm in detail, while Section 4 Robots for domestic task Field robotics presents the results of this approach. Robot companions, assistants, Agriculture humanoids Vacuuming, floor cleaning Milking robots and livestock robotics 3.5. Network space of topics Window cleaning Mining systems Lawn-mowing Space robots Having classified the patent descriptions into topics, we can assess to Pool cleaning what extent those topics are related to each other, i.e. how often these Professional cleaning Entertainment robots Floor cleaning topics appear in the same patent descriptions. The underlying assump­ Toys and hobby robots Window and wall cleaning (incl. wall-climbing tion is that the more the knowledge on technologies captured by the two robots) topics are used jointly in the same patents, the more they are interde­ Multimedia robots Tank, tube and pipe cleaning pendent. An example can be navigation of autonomous guided vehicles Education and research Hull cleaning (aircraft, vehicles, ships etc.) Other cleaning tasks based on the method of scanning the external environment. Capturing Elderly and handicap these relations between topics is important as it helps to look on the assistance topics not merely in isolation which is not adequate to understand Robotized wheelchairs Construction and demolition distinct technologies and to arrive at an overall picture of the knowledge Personal aids and assistive Nuclear demolition and dismantling space underlying the robotics industry. Therefore, the approach pre­ devices Building construction sented below is the second methodological contribution of the present Home security and Heavy/civil construction paper to the literature applying topic modeling to patent data. surveillance For this purpose, we define two topics as connected based on their Home security and surveillance Other construction systems (road construction) cosine similarity, i.e. co-occurrence of those topics in the same patents Logistic systems Automated Guided Vehicles (AGVs) in (see Appendix B for a formal description of the measure). Measuring this manufacturing environments similarity between any pair of topics we obtain a symmetric matrix with AGVs in non-manufacturing environments ones on the main diagonal as cosine similarity of two identical vectors is (indoor) one, and all other values bounded in [0,1). These values capture the Cargo handling, outdoor logistics strength of relation between any pair of topics and can be interpreted as Personal transportation Inspection and maintenance systems corresponding edge weights in an undirected graph between these Facilities and plants topics. In particular, a high weight implies that the two topics appear in Tank, tubes, pipes and sewers many patents with a high prevalence, and they are strongly linked. Since Other inspection systems (inspection robots for one patent can contain small parts of text belonging to many topics, we nuclear plants) Medical robotics get a virtually fully connected graph, where many edges have relatively Diagnostic systems low weight. To simplify its analysis and visualization, we assess signif­ Robot-assisted surgery and therapy icance of each particular edge in this weighted graph. To do so, we Rehabilitation systems follow Saracco et al. (2015) in constructing an appropriate null model – Other medical robots its randomly generated counterpart – which displays on average the Rescue and security applications Fire- and disaster-fighting robots same degree distribution (diversification of patents) and the same Surveillance/security robots ubiquity (weight distribution of each topic) 1000 times.22 Comparing Other surveillance and security robots the empirically observed weighted graph with their randomly generated Defense applications counterparts we preserve only those links which weight surpasses the Demining 95% threshold, i.e. they fall in the 5% of most outstanding edge weights Unmanned aerial vehicles (defense applications) Unmanned ground-based vehicles which could have been observed given the underlying data. The 5% Unmanned Underwater Systems threshold is taken as most conventional significance level observed in Underwater systems (civil / general use) the empirical literature. Note at this point that by deleting 95% of links Underwater systems (civil / general use) and transforming our network from weighted to unweighted one, we Powered human exoskeletons Powered human exoskeletons reduce the density of our topic networks to 0.05, i.e. only 5% of all Mobile platforms in general use possible links are present in our network. This, however, does not pre­ Mobile platforms in general use clude us from analyzing other important characteristics of the topic Public relations and joy rides network such as number of components capturing how many topics are Hotel and restaurant isolated from other topics or network centrality accounting for the con­ Guidance Marketing centration of edges on few topics (degree centralization) and the Robot joy rides dependence on topics that connect many other topics (betweenness centralization) proposed by Freeman (1979). It is important to remember, however, that the absence of an edge between any pair of scientific literature of contemporary American English.20 As a result, topics in our henceforth analysis does not imply that they never appear each topic belongs to each SR sub-area with a different matching together in any patent, but that the extent to which they co-occur does score.21 In the final step, we derive a threshold value for these matching not meet the bar to be considered as significant. 20 To this end, the database composed by Mark Davies (https://www.wordfre quency.info/) has been used. 21 Logically, if there was no candidate word appearing in the topic with a 22 positive probability and in the specific SR sub-area, the corresponding matching The same approach for testing link significance has been applied, among score equals zero. others, by Napolitano et al. (2018) and Pugliese et al. (2019). 8 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 Table 2 Topics from LDA model with K=380 matched to SR sub-areas. Topic Topic terms (30 most relevant) Matched SR sub-area Field Robotics 377 platen, flat, abrasive, crop, agricultural, harvest, harvester, tractor, abrade, lapping, tree, harvesting, label, lap, fertilizer, Agriculture farm, abrading, granite, fruit, agriculture, barcode_label, edging, farmer, queue_manager, floating_platen, hieroglyph, flat_surfaced, flatness, edger, contractor, 86 teat_cup, milk, animal, milking, teat, cow, milk_animal, harness, milking_animal, teat_animal, wire_harness, Milking robots and livestock robotics milking_parlor, udder, animal_milk, luggage, fodder, milk_parlor, udder_animal, disinfect, slotted, hind_leg, animal_udder, jumper, foremilking, occupation, beveling, facet, formboard, compartment, milker, 330 turret, conversion, last_teat, stepper_motor, light_curtain, deactivation_threshold, dairy_animal, strain_relief, Milking robots and livestock robotics umbilical_cord, milking, averaging_period, duration, garage_door, exploration, mending, position_sensor_print_circuit_board, stairstep, intensifier, nonlinear, recirculating, nonlinear_exploration, pause, unprogrammed, bipedalism, historical, dependence, gunshot, sclera, touchup, obstruction, 113 shift, traction, bucket, orbit, momentum, recite_claim, zmp, spacecraft, shifting, leg_hopping, upstanding, hst, midpoint, Space robots circumscribed, ferrule, stable, foamed_seal, insole, forwardly, shiftable, outwardly_shifts, mule, virus_strain, infectious_bacterial, wearable_assistive, backwardly_shift, inscribed_cylinder, outwardly_flexed, backwardly_flex, time_of_flight_diffraction, 37 seed, sucker, hood, mat, baseplate, nutrient, tissue_explants, soil, seedling, grit, firefighting, curved_backplane, handheld, Agriculture greenhouse, plant, loaded_filament, podium, seeding, tissue_explant, mature_plant, debridement, chimney, planting, leaf, sucker_revolving, harvest_crop, public_speaking, germination, presentment, suspendably_translatable, 260 course, unmanned_vehicle, truck, unmanned_dump, dump, dump_truck, topographical, foremilk, earth_sand, wherefore, Mining systems partial, contraband, haul, travel, survey, anonymous, wherewith, twice_diffracted, loader, presence_contraband, conflict, wait, manned, loading, idea_underlie, quarry, dumping, site, topological_map, lashing, Medical robotics 83 fiducial, jog_feed, canister, stereotactic, cockpit, transformable_toy, toy_doll, craft, skull, shin, neurosurgery, preoperative, Robot-assisted surgery and therapy astronaut, fiducials, fixable, sanding, sand, neurosurgical, anatomical, tiptoe, clinch_nut, aquatic, predrilled_hole, platter, satellite_platter, mri, intraoperative, raster, microsurgery, intracranial, 6↑ catheter, elongated, distal_end, medical, sled, elongate, handle, proximal_end, sheath, disposable, catheter_sheath, Robot-assisted surgery and therapy sterile_barrier, nose_cone, sterile, steerable_catheter, patient, ablation, knob, ablation_catheter, introducer, glass_crucible, sterility, catheter_introducer, electrophysiology_catheter, remote, introducer_sheath, mapping_ablation, electrophysiology, remotely, effective, 191↑ surgical, surgeon, surgical_instrument, surgery, minimally_invasive, surgical_procedure, surgical_site, incision, patient, Robot-assisted surgery and therapy medical, instrument, medical_procedure, endoscope, entry, intuitive_surgical, robotic, console, system, tissue, procedure, vinci_surgical, trauma, endoscopic, hospital_stay, perform, inc_sunnyvale, robotically, stereoscopic, orthopedic, recovery, 280 ↑ tissue, organ, accessory, forceps, endoscope, laparoscopic, capsule, endoscopic, trocar, laparoscope, blood_vessel, snake, Robot-assisted surgery and therapy laparoscopic_surgery, endoscopy, laparoscopy, minimally_invasive_surgical, abdomen, abdominal_cavity, continuum, vivo, abdominal, cavity, terminus, minimally_invasive, intracorporeal, cannula, connected, discomfort, clip_applier, invasive_surgical, 167 radiation, tumor, dose, patient, diagnostic, couch, ray, therapeutic, lesion, imaging, radiation_therapy, cancer, Robot-assisted surgery and therapy radiotherapy, therapeutic_radiation, treatment, therapy, ambulatory, sighting, radiosurgery, physician, biopsy, compute_tomography, magnetic_resonance, shooter, rays, computed_tomography, collimator, medical, proton, linac, 311 ↑ simulation, user, training, simulate, simulated, feedback, simulator, physical, haptic, virtual_reality, graphical, computer, Rehabilitation systems mouse, interaction, joystick, trainee, intuitive, real, feel, sensation, train, interface, realistic, interact, amusement, haptic_feedback, experience, environment, graphic, assistive, 308 ↑ prosthetic, prosthesis, ankle, artificial, wearer, amputee, knee, hip, foot, gait, flexion, limb, residual_limb, powered, Rehabilitation systems muscle, prosthetic_knee, toe, automatic_guided_vehicle, extremity, ankle_joint, amputation, flexion_extension, prosthetic_orthotic, damper, myoelectric, orthosis, heel, knee_joint, thigh, heel_strike, Robots for domestic task 130 ↑ rack, boundary, mower, lawnmower, lawn, grass, mow, mowing, saw_blade, helical_spring, mow_lawn, Lawn-mowing robotic_welding_assembly_device, root, geographic, traversal_launch, urls, mowable, traversal, substage_await, nursery_school, steeple, cubic, substage, substage_empty, photo_interrupter, rotor_steeple, young_child, unburned, angioplastic, playmate, 284 clean, dry, wash, cleaning, rinse, bath, wet, washing, drying, chemical, foreign, residue, cleaned, solution, jet, liquid, dirty, Vacuuming, floor cleaning pellicle_paste, impurity, immerse, remove, particle, towel, ipa, spray, diced, mop, dip, contaminate, water, Professional cleaning 7 pipe, piping, fuel_tank, pip, trough, circumferential, wiring_piping, diameter, worm_gear, minitube, tobacco_roll, Tank, tubes, pipes and sewers circumferentially, internal_finned, crest, inner, tooth_stump, gauge, axial_direction, butt, lapse_predicted, pipes, levelling_bench, circumference, minitubes, society_for_biomolecular_screening, swedish_specification, fueling_pistol, supportive, readable, pitting, 218 water, tank, rock, bay, submerged, sediment, submerge, reverse_osmosis, purified_water, scavenge, ingredient, oht, pump, Tank, tubes, pipes and sewers petroleum, chemical_mechanical_polish, underground, sink, borehole, purification, kitchen, buoyancy, rice, intrabay, liter, interbay, mpi_medium, cooking, immerse, purify, ballast, 223 housing, seal, sealing, bellow, inlet, outlet, opening, tight, leak, lip, sealed, interior, inner, evacuation, gland, venting, Tank, tubes, pipes and sewers bellows, sealingly, diaphragm, evacuate, outer, enclose, annular, house, manifold, reactive_ion_etch_mode, rings, vent, inlet_outlet, closed, Logistic systems 144 ↑ shelf, automated, warehouse, inventory, automate, sort, mail, retail, shipping, fulfillment, facility, forklift, sorting, order, Automated Guided Vehicles (AGVs) in pallet, good, aisle, deadlock, retailer, logistics, management, tote, shipment, acceleration_slowdown, shelve, distribution, manufacturing environments sale, depot, stock_keeping_unit, shelving, Defense applications 327 ↑ group, mission, game, weapon, surveillance, unmanned, lawn_mower, military, player, terrain, team, swarm, waist, Unmanned ground-based vehicles soldier, helicopter, enemy, threat, reconnaissance, tactical, mobility, ugvs, squad, unmanned_air_vehicle, fly, sport, combat, ground, launch, drone, opponent, Construction and demolition 353 lay, fire, bake, explosive, containment, chill, supercritical, branch, laying, underground_pipeline, branch_pipe, detonation, Heavy/civil construction detonate, baking, vernier, firing, rain_maker, wildfire, fire_fighter, buckle_arrestor, blasting_cap, photoresist, electric_discharge_machining, sewer, baked, disruption, sige, fire_fight, overlap_vernier, bake_chill, Note: For each topic, its top 30 most relevant (i.e. with high likelihood and exclusivity, see equation (8) for formal definition) words are displayed. Up-arrows indicate that the diffusion of the respective topic follows a positive trend, vice versa for down-arrow. 9 Topics within an area are sorted by their topic matching scores. I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 4. Results Table 3 Metrics for networks of topics . 4.1. Number of topics 1977- 1981- 1996- 2011- 2017 1985 2000 2015 As described in Section 3.3, we proceed with pre-processed texts by LDA model with 380 topics computing LDA models and plotting the corresponding scores on per­ Number of components 51 2 4 15 plexity, coherence and exclusivity for different numbers of topics Clustering 0.41 0.26 0.32 0.36 ranging from 10 to 500 with an interval of 10 (Fig. 5). The best model Mean path length in LCC 95.46 4.45 8.55 29.92 Modularity 0.28 0.27 0.28 0.29 should have the lowest perplexity score (i.e. lowest prediction error), Degree centralization 0.31 0.35 0.28 0.35 and the highest coherence and exclusivity. As one can see, the perplexity Betweenness 0.06 0.11 0.08 0.11 score improves with the number of topics, but the improvement is centralization marginally decreasing. The same pattern is observed for the exclusivity, LDA model with 190 topics Number of components 22 3 7 16 while coherence tends to aggravate for the growing number of topics. Clustering 0.37 0.29 0.31 0.33 Choosing an optimal number of topics K that would maximize all three Mean path length in LCC 42.09 6.72 14.55 31.26 dimensions in such circumstances is not possible. An additional fourth Modularity 0.33 0.34 0.35 0.36 criterion for choosing K is model complexity implying that with more Degree centralization 0.27 0.18 0.18 0.24 topics it becomes increasingly difficult to overview and interpret them. Betweenness 0.13 0.10 0.08 0.14 centralization For that reason, observing that after approx. 300 topics the three se­ lection criteria start to change much slower, we chose K to be equal to 380 topics as for this number we observe a (small) local optimium both generated word clouds and diffusion curves for the LDA models with 190 for perplexity and exclusivity. In addition, to demonstrate robustness of and 380 topics, two observations arise. First, the generated topics our results we henceforth report in parallel the topic model with 190 matched to SR in the two LDA models tend to strongly overlap. One topics, i.e. exactly half of K. We chose the alternative value to be smaller example is the pair of topics with largest topic matching scores on to increase readability of the results, as charts for networks, diffusion milking robots (topics 377/380 and 105/19024). Other good examples curves and word clouds tend to be simpler for smaller K. are topics on vacuuming and floor cleaning robots (284/380 and 6/ 190), robot assisted surgery (191/380 and 110/190) or tanks, tubes and 4.2. Matching topics to SR pipes (7/380 and 134/190). Second, the topics matched to SR tend to be rising topics. Among 380 topics, their chances to be classified as rising Applying our matching procedure described in Section 3.4 and are twice higher (38%, or 8 out of 21 topics) than for all topics con­ Appendix B on the constructed topics, we generate a distribution of topic structed on average (≈17%, or 63 out of 380 topics). For example, matching scores ranging for the case of 380 topics between 0 (i.e. no almost all of the topics matched to medicine (6, 191, 280, 308 and 311) candidate word overlapping between IFR reference texts on SR and our have experienced a fast growth in the last ten years, while in the 1980s topics was identified) and 5.62. The largest topic matching score has they were very small. There is also no falling topic among SR topics been obtained for topic 377 matched with the agriculture IFR sub-area. compared to 11% (41 out of 380) among all topics constructed. For 190 We define the threshold value maximizing the total number of uniquely topics and twenty of them being matched to SR, the situation is similar matched topics as 1.366 (see right plot in Fig. 6a). This is illustrated with with 40% (8 out of 20) vs. 20% (38 out of 190) for rising topics, and 5% the intersection of the black line showing the number of topics classified (1 out of 20 matched SR) vs. 18% (34 out of all 190 topics) for falling to only one SR sub-area and the red line demonstrating total number of topics. topic matches to SR sub-areas including those cases where the same topic has been matched to two, three or more SR sub-areas. Taking a 4.3. Network perspective threshold value below 1.366 will result in topics simultaneously matched to more than one SR sub-area, while a threshold value above The network generated for the whole period of 41 years (1977–2017) 1.366 would result in fewer uniquely matched topics having higher as well as three equally distant time windows of five years length are overlap with IFR SR descriptions. This threshold value resulted in 21 out presented in Figs. 9 and 10.25 As it was explained in Section 3.5, the of 380 topics matched to SR (see Table 3 for details on which topics have density of our networks at any period by construction is 5%, while the been matched to which SR sub-areas). Note that for 190 topics the topic absence of an edge between any pair of topics in our graph does not matching scores vary between 0 and 5.59 (Fig. 6b). The largest matching imply that those topics never appear together in our patent texts but that score is again from the sub-area of agriculture for topic 105. The the frequency and prevalence with which they co-occur is not high threshold value maximizing the number of uniquely matched topics is enough to be considered as significant. The topics belonging simulta­ 1.280. This results in 20 matched topics (see Table C5 in Appendix C). neously to 10% most central topics in terms of degree centrality (i.e. Thus, the results of applying the matching procedure for 190 and 380 number of edges) and betweenness centrality (i.e. how often these topics topics look very similar supporting the robustness of our results for are located on the shortest paths connecting any random pair of nodes in different K. the graph) are marked with squared node borders. One can conclude The left panels of Figs. 7 and 8 show word clouds we matched to that these topics represent some general technologies in the field of service robotics among the 380 and 190 topics, respectivelyFigure C18 robotics appearing in many different patents and connecting different and C24.23 For example, for model with 380 topics, topic 191 has words parts of the knowledge space in robotics together. In this sense, these can ‘surgical’ and ‘surgeon’ both very frequent and exclusive, while the be compared to so called ‘general purpose technologies’ (see Korzinov words ‘medical’ and ‘invasive’ are neither as frequent nor as exclusive. and Savin (2018) for a recent discussion) that just like electricity or The resulting diffusion curves can be found in Figure C19 and steam engine ’enable’ development of many new inventions. Word C25Appendix C for the whole set of topics generated, while in right panels of Figs. 7 and 8 we show diffusion curves for SR topics among the 380 and 190 topics, respectively. Inspecting and comparing the 24 Henceforth, for brevity reasons, we will indicate with ’/K’ the LDA model the topic index belongs to. 25 Since in the latter we take into account only particular sub-periods, pres­ 23 Word clouds for the whole set of topics generated in both LDA models can ence of edges between topics can vary from period to period under be found in Appendix C. consideration. 10 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 Fig. 7. 21 SR topics in the LDA model with 380 topics. Fig. 8. 20 SR topics in the LDA model with 190 topics. clouds of these topics are presented in Figs. C16 and C22 in Appendix C. network constructed for the whole period of 1977–2017. These topics Their visual inspection corroborates our intuition: these topics often are about technologies around seeding and harvesting (crops) using deal with processing information, positioning and orientating in space, robots. These areas of application as such are quite specific and do not while words in the word clouds having dark font imply that they have a come often neither with general topics we mentioned earlier nor with relatively low exclusivity. other topics like those matched to SR or identified as isolates. In contrast, the isolated topics (or so-called ’isolates’) represent parts The majority of topics in our graphs are naturally in between these of knowledge infrequently appearing together with other topics, having two extremes (most central topics and isolates). Identifying few edges predominantly exclusive terms in their word clouds26 and capturing for them we can see what other areas of robotic knowledge they are what we call specialized application areas of robotic technologies. Take closest to, which can help in their interpretation. E.g. topic 191 as an example topics 37 and 377 from the LDA model with 380 topics. (matched to SR sub-area robot-assisted surgery and therapy) in the LDA Both are matched to SR (sub-area agriculture) and are isolates in the model for 380 topics is connected to topics 338 (containing many terms on telesurgery), 9 (addressing endoscopy and filtering) and 96 (manip­ ulation and teleoperated machines). None of the latter three has been 26 Word clouds of these topics are presented in Figs. C17 and C23 in the matched to SR via our matching approach. Thus, the network perspec­ Appendix C. tive gives a broader picture on what kind of technologies tend to be 11 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 Fig. 9. Networks of topics for the period 1977–2017. Note: The size of the nodes represents the percent of the corpus of patents texts explained by the topic, while the presence of the edge captures the fact that the two topics tend to co-occur with large prevalence in many patent documents so that the chance to observe them together in a randomized null model is below 5% (see Section 3.5). Topics belonging to 10% most central in terms of degree and betweenness centrality are highlighted with squared frames, while topics matched to SR - with red color. Finally, cyan and magenta color are used to highlight the two clusters of topics on software and hardware, respectively. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) 12 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 Fig. 10. Networks of topics for three time windows: 1981–1985, 1996–2000 and 2011–2015 Note: The size of the nodes represents the percent of the corpus of patents texts explained by the topic, while the presence of the edge captures the fact that the two topics tend to co-occur with large prevalence in many patent documents so that the chance to observe them together in a randomized null model is below 5% (see Section 3.5). Topics belonging to 10% most central in terms of degree and betweenness centrality are highlighted with squared frames, while topics matched to SR - with red color. Finally, cyan and magenta color are used to highlight the two clusters of topics on software and hardware, respectively. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) 13 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 Fig. 11. Prevalence of topic clusters and SR over time. LDA model with 190 topics (left plot) and 380 topics (right plot). recombined with each other in different patents to further extend the that as formation of hierarchy in the knowledge network, where topics knowledge frontier. from one area of knowledge tend to become isolated from the topics in Finally, to identify communities of nodes forming strong inter­ the other areas. The intuition that our constructed networks are hier­ connected clusters in our graphs of topics, we apply the algorithm archical is further supported by themodularity measure being around proposed by Newman (2006) based on leading eigenvector of the 0.3.30 Concerning the importance of different topics in the network, we community matrix. This allows us to identify two large communities of use the concept of network centralization that is concerned not with the nodes dominating the networks.27 Looking closer on the word clouds of overall connectedness but with the particular structure of relations and topics forming those two communities, one can quickly realize that these relative positions of the nodes. We estimate two commonly used can be interpreted as clusters of technologies concentrated around either centralization metrics. The degree centralization measures the concen­ hardware technologies (e.g. body, movement, energy), or software tration of linkages on few nodes (1 would be a star network and 0 – a technologies (e.g. receiving, processing and storing information). See complete graph). The betweenness centralization (again between 0 and 1) word clouds with most central topics from each of the two cluster for indicates the dependence on nodes that connect many other nodes. Both both LDA models inFigure C14Figure C15Figure C20 and C21 in measures remain rather stable over time, at least if we look on the LDA Appendix C. We highlight these two clusters in Figs. 9 and 10 with model with 380 topics. Stability of these measures means that the different colors (magenta for hardware and cyan for software). asymmetry between topics in their centrality does not change and most Interestingly, if we sum up prevalences of topics belonging to these central topics do not become more important in more recent years.31 two clusters and to SR over time, we clearly see on Fig. 11 that i) the software cluster overtook the hardware one in the share of patent doc­ 5. Review, perspectives, and policy implications uments, which confirms the rising importance of ICT; ii) service robotics as expected is a rapidly growing research field in the last few decades.28 The present paper extends the existing economic literature on ro­ We calculate several graph indicators to quantify the generated botic research in several directions. Using unstructured patent infor­ networks and their development over time (Table 4). First, we see that mation on robotics that covers a time span of more than four decades the number of components, that captures how many nodes are isolated, and applying a topic modeling approach we discover latent topics in increases over time. This we interpret as a sign that topics in robotics those texts, which we later match to SR sub-areas using a novel matching became more ‘independent’ and can form patentable knowledge with procedure. The latter is done by identifying words that appear both, in other topics. Second, The clustering coefficient of the network increases the lists of words associated to each topic with different probabilities steadily. Apparently, the generated topics increasingly form densely and the descriptions of SR areas from IFR. Then we account for the connected cliques inside the largest connected component (LCC).29 frequency and exclusivity of these words (also compared to scientific Third, the mean path length – measuring the average number of edges to literature outside IFR descriptions), and calculate a threshold level connect any pair of randomly chosen topics in LCC (i.e. disregarding necessary to be reached to be identified as SR topic, which maximizes isolated topics) – clearly increased over time indicating a lack of short the number of unique SR topic matches. This way we contribute to the paths connecting the densely interconnected cliques. One can interpret literature on topic identification and interpretation which is rather nascent when it comes to an automatic topic matching and labeling. Another key novelty of our analysis is the embedding of topics into a network perspective. This allows to capture interdependencies between 27 In fact, the algorithm identifies more than two communities, but we focus robotic technologies by demonstrating which topics appear a lot only on the two largest ones in terms of number of nodes disregarding the together over many patent documents, and which do not. This allows to smaller communities. This choice is primarily motivated by their size, while all determine the relative position of SR topics within this network, observe other communities are much smaller and hard to classify to any area. 28 One can see that in the LDA model with 190 topics the share of texts belonging to SR is considerably bigger than in the LDA model with 380 topics. 30 This has to do with the fact that the model with less topics is less detailed in This estimate is close to the one by Valverde et al. (2007) obtained for disaggregating patent texts into topics. While both measures are imperfect since computed tomography of 0.41. 31 they miss topics which overlap with IFR SR classification to a smaller extent For 190 topics we observe an increase in both metrics of centralization in than our matching threshold, we believe that the results for 380 topics shall be the last five-year window (2011–2015). Since the model with 380 topics is considered as a more accurate one. considered by us as more accurate one, we abstain from interpreting this result, 29 LCC comprises all the elements of the network, where any two nodes are as it may be an outcome of more rough classification of documents into smaller connected to each other by a sequence of edges (path). number of topics. 14 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 an increasing hierarchy of the network with high clustering and in e.g. elderly care, logistics and cleaning, which is particularly impor­ distinguish between topics that connect to many areas of knowledge tant in the light of recent demographic trends in industrialized countries. enabling their recombination and topics that are rather isolated and Related to robot markets, we can link the dynamics of SR patenting focus on distinct application areas. Both these novelties are imple­ to SR sales. Comparing sales data and diffusion curves results in a robust mented in a data-driven way, independent of both experts’ idiosyncratic finding that SR in medicine and logistics are the leading sectors of ser­ knowledge and patents’ structured information (such as concordances). vice robotics. We thus go beyond existing literature (e.g. Kreuchauff and Korzinov The paper at hand has several policy implications. Looking in the (2017)) that bases their SR patent identification strategy on expert past, we see that robot patenting activity has been less susceptible to the judgment. Notice that patent classifications (such as International Pat­ financial crisis (2007–2009) than robot investment. However, in the ent Classification) also rely on expert knowledge and introduce new light of the current COVID-19 pandemic the potential to reduce conta­ classes with a considerable time delay compared to first patents filed on gion risks might not only spur robot research but also become a driver of that topic, thus being inherently weak to capture emerging trends in SR. robot demand.34 It is natural to assume that this will foster sales of While examining our results, one should remember that they just as service robots also in the long term and spur private investment in ro­ any other topic model are a result of stochastic optimization implying botic technologies. Governments could support this process by com­ that running the LDA algorithm several times one can obtain every time plementary basic research funding and subsidizing applied research in (slightly) different outcome. Therefore, it is important to focus on the ares where market incentives are not strong enough to fulfill actual qualitative results of our analysis. To ensure their robustness, we needs. compare two topic models: with 190 topics and 380 topics. Even though For example, related to care, lack of qualified labor is evident. the two models have very different number of topics, the results they Increasing the automation intensity here would help to solve this convey are very similar. First, they exhibit almost the same number of problem. An intensified use of robots would furthermore alter the topics matched to SR (approx. 20). Second, predominant number of associated work environment as well as the skill requirements of care­ topics matched to SR are in field and medical robotics. Third, we find takers making the job attractive for technology-oriented workers. As a disproportionally large number of rising patterns among the diffusion consequence, the educational system has to be adjusted to the new needs curves of topics matched to SR. Fourth, very similar structural properties and opportunities. To cope with this process in a timely and efficient of the networks of topics (two large communities reflecting hardware manner, a sound understanding of SR is important. And this is where our and software, position of topics matched to SR in the periphery of the study comes into play as we show that information out of patents can be network, similar level and dynamics for network metrics such as clus­ used as early signal prior to the actual transformation. Furthermore, in tering, modularity and centralization). Fifth, rising trend in the share of light of the recent progress in NLP and ML methods and growing amount service topics and software cluster among topics on robotics. These re­ of data being produced in sensitive application areas like healthcare, sults convince us that the patterns we observe for service robotics are adequate conditions related to data privacy and data security have to be robust to alternative specification of LDA parameters. established. The governments should especially think about how high Our approach enables us to replicate some stylized facts of robot quality data may be made available for independent scientists to avoid technology transformations as identified by the use of structured patent monopolization of this resource and thus to create a level playing field at data and may thus be compared to analysis based on well established the global level. concordances (WIPO (2008) for patent-technology mapping, Eurostat (2014) for patent-industry mapping). Among these facts are the relative 6. Conclusions and outlook shift of focus in robotics technology from hardware to software, the increasing importance of robotics in medicine (see Table A12 in the There is a vast amount of studies on robotics in the economic liter­ Appendix), agriculture and logistics (see discussion in Section 2.1). With ature. However, most of them are related to IR with a focus on labor the proposed method we can look closer into specific SR sub-areas. In market implications. Services, in contrast, are still rarely addressed case of medical robotics we see the increasing importance of surgery and although here the automation potential is huge. This gap in the litera­ therapy but also - and this is not yet common knowledge - a strong trend ture is mostly due to a lack of reliable data and the difficulties to of robot technologies related to rehabilitation.32 Besides, even specific quantify emerging technological fields. We shed new light here by areas of application can be identified and spotted. For example, we find applying topic modeling and graph theory to textual patent data. the term ’laparoscopy’,33 which is an operation performed in the Furthermore, given that topic labeling and interpretation is a chal­ abdomen or pelvis using small incisions (usually 0.5-1.5 cm) with the lenging task lacking objective data-driven methods, we proposed and aid of a camera. In the technology description of the SR report, the word applied a novel approach that uses external reference texts on SR ro­ ’laparoscopic’ shows up while describing various types of operations botics, namely technology descriptions of SR from the IFR. carried out by robots (IFR, 2018b, 136). Associated topic 280/380 has Our results reaffirm that at present various disciplines are involved in been matched to medical robots (see Table 3). technological development of robotics. Among the most prominent are We are able to interpret results of our analysis in the light of eco­ computer science, automation control, electrical and mechanical engi­ nomic theories and recent technological trends. Most importantly, we neering and bio-medicine (Analytics (2018)). Today’s robots are demonstrate automation potential materializing within services, an increasingly characterized by more intelligent components (e.g. smart economically important sector which until today has not yet seen much grippers), greater connectivity (e.g. plug & play interfaces, cloud automation. While robot adoption in manufacturing industries mostly computing) and are easier to use (e.g. programming by demonstration). has the advantage to relieve humans from dirty, dull and dangerous Although these technologies are also applied in IR, the future of modern tasks, in services non-standardized human-to-human interaction pro­ robots happens increasingly beyond predefined environments. As a cesses are essential. The use of SR may help to overcome labor shortages consequence, the areas of application also change constantly over time: logistics, medicine, agriculture, construction, cleaning. The rapid tech­ nology advances will likely make robots applicable also in other sectors 32 Within medical applications we are even able to further differentiate the that until now have not yet seen much automation. These technologies usage of robots, e.g. the increasing role of rehabilitation systems (topic 308/ might help to overcome a phenomenon which has become known as 380) and surgery (topic 311/380 focuses more on the hardware perspective while topic 191/380 relates to the complementary software). 33 34 From ancient Greek λαπαρα (lapara) ’flank, side’, and σ κωπϵω (skopeo) ’to Caselli et al. (2020) have recently shown that robotization indeed facilitates see’. social distancing and lowers the risk of contagion. 15 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 ’Baumol’s cost disease’35 or the phenomenon of secular stagnation, i.e. position of countries in mastering those technological topics. It is higher productivity in service sectors which is due to automation worthwhile to assess revealed technological advantages of each country contributing to aggregate productivity growth. in different SR sub-areas and how these were changing over time. This This study has a few limitations, which can also be seen as directions would be of particular interest for policy makers to design their sup­ for further research. First, we could not find comparable sources of de­ porting measures for local companies competing within the interna­ scriptions to automatically label the remaining (non-SR) topics. tional technology markets. Furthermore, since topic diffusion curves Applying our topic matching on a broader classification of robotics is may be linked to sales dynamics, one could extend the analysis by therefore welcome. Second, patent data covers mainly technical side of building market forecasts based on the patenting activities. technologies in robotics. One could go for a broader perspective and use LDA for other related data sources such as scientific publications, CRediT authorship contribution statement funding programs, newspapers and other media coverage. Such data would address societal discourses and political debates about robotics to Ivan Savin: Conceptualization, Formal analysis, Methodology, elicit not technologies but cultural differences, such as arguments in Writing – original draft. Ingrid Ott: Conceptualization, Formal analysis, favor and against robots as drivers or doctors. This way one could draw a Writing – original draft. Chris Konop: Conceptualization, Formal comprehensive picture of SR acceptance including the perspectives of all analysis, Methodology. stakeholders within the technology system. Third, we did not look on the Appendix A. Classification of robots: ISO standard and IFR operationalization A1. Robot categories according to standard ISO 8373:2012(en) (ISO 8373:2012, 2.6) A robot is an actuated mechanism programmable in two or more axes with a degree of autonomy, moving within its environment, to perform intended tasks. Autonomy in this context means the ability to perform intended tasks based on current state and sensing, without human intervention. (ISO 8373:2012, 2.10) A service robot is a robot that performs useful tasks for humans or equipment excluding industrial automation application. Note: The classification of a robot into industrial robot or service robot is done according to its intended application. (ISO 8373:2012, 2.11) A personal service robot or a service robot for personal use is a service robot used for a non-commercial task, usually by lay persons. Examples are domestic servant robot, automated wheelchair, and personal mobility assist robot. (ISO 8373:2012, 2.12) A professional service robot or a service robot for professional use is a service robot used for a commercial task, usually operated by a properly trained operator. Examples are cleaning robot for public places, delivery robot in offices or hospitals, fire-fighting robot, rehabilitation robot and surgery robot in hospitals. In this context, an operator is a person designated to start, monitor and stop the intended operation of a robot or a robot system. (ISO 8373:2012, 2.15) A robot system is a system comprising robot(s), end-effector(s) and any machinery, equipment, devices, or sensors sup­ porting the robot performing its task. Fig. A12. Robot patent–industry as­ signments and their relative importance across three time windows Note: Results are provided in weighted patent counts mapped according to the PATSTAT database. Color highlighting only refers to those industries that we address throughout our study. Green (orange) color highlighting indicates an increase (decrease) in relative importance over time, which can be seen e.g. in com­ puters and peripheral equipment or medical and dental instruments (e.g. machine tools). 35 Baumol and Bowen (1966) observed that rapid productivity growth in some sectors (e.g. in manufacturing) relative to other sectors (e.g. service industries) could result in a ’cost disease’ at the aggregate level if the slowly growing sectors constitute a large part of the economy. Baumol (2012) provides a recent discussion with a key focus on healthcare, while Aghion et al. (2017) adapt this reasoning to AI and automation. 16 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 A2. Robot patent-industry assignments based on PATSTAT mapping Appendix B. Further details on the methods used B1. Data pre-processing The following stopwords have been added to the NLTK’s list of English stopwords: ’would’, ’could’, ’figure’, ’whether’, ’wherein’, ’respecti­ ve’,’primary’, ’certain’, ’etc’, ’described’, ’herein’, ’first’, ’second’, ’third’, ’fourth’, ’fifth’, ’moreover’, ’furthermore’, ’according’, ’yes’, ’present’, ’disclosure’, ’preferably’, ’feature’, ’like’, ’embodiment’, ’comprising’, ’preferred’, ’example’, ’exhibit’, ’develop’, ’demonstrate’, ’comprise’, ’consist’, ’prefer’, ’prefering’, ’preferred’, ’otherwise’, ’alternative’, ’alternatively’, ’discuss’, ’generally’, ’exemplary’, ’substantially’, ’typically’, ’additional’, ’upon’, ’suitable’, ’feature’, ’aspect’, ’design’, ’employ’, ’create’, ’multiple’, ’understood’, ’non’, ’depict’, ’previously’, ’partially’, ’say’, ’take place’, ’great’, ’suitable’, ’design’, ’generally’, ’exemplary’, ’aspect’, ’combination’, ’accordance’, ’appreciate’, ’arrangement’, ’disclose’, ’typically’, ’understood’, ’multiple’, ’skilled’, ’configuration’, ’create’, ’employ’, ’upon’, ’direct’, ’mechanic’, ’alter’. The normalized pointwise mutual information (NMPI) score is calculated as follows: log(Pr(worda , wordb )/(Pr(worda ) × Pr(wordb )) score(worda , wordb ) = (1) − log(Pr(worda , wordb )) This score sets the probability of two words occurring together in relation to the probability of them occurring together in case of independence. A score of -1 (in the limit) means two words are never occurring together, 0 in case of independence (they occur together as often as expected, based on their independent probability), and +1 in complete co-occurrence. In our case, two words will be merged into a bi-gram with a ”_” symbol, if the score is above 0.5. B2. Formal description of LDA and the parameters used More formally LDA can be described as on Fig. B13. The shaded circle represents the observed data (w(d,n) , nth word in each document d). The unshaded circles denote latent (unobservable) variables: z(d,n) - the topic assignment for the nth word in document d (or, alternatively, the assignment of words to topics), θd - the topic proportions for the dth document, and ψ (1:K) - topic distributions over the vocabulary. Arrows indicate the conditional dependencies between variables, while frames (the boxes in the figure) refer to repetitions of sampling steps, with the variable in the lower right corner referring to the number of samplings: number of documents D and the number of topics K. Thus, the inner frame over z(d,n) and w(d,n) represent the repeated sampling of topics and words until N words have been generated for each document d; the frame surrounding θd illustrates the sampling of a distribution over topics for each document d for a total of D documents; the frame surrounding ψ k illustrates the repeated sampling of word distributions for each topic assignment until the word probabilities of K topics have been generated. We have used the parallelized Latent Dirichlet Allocation from gensim, an open-source library for unsupervised topic modeling and natural language processing, using modern statistical machine learning implemented in Python. To estimate the LDA model, several simplifying assumptions have to be made. First, θ is assumed to be a random draw from a uniform Dirichlet distribution with scaling parameter α. Second, ψ is assumed to be a random draw from a Dirichlet distribution with parameter β. Following Griffith and Steyvers (2004) and the default settings of the LDA gensim package (https://radimrehurek.com/gensim/models/ldamodel.html), we set α = 1/K, while β as a prior is learned by LDA from the data. For T>100 β << 0.01, which shall result in a fine-grained decomposition of the texts into topics that address specific areas of knowledge.36 Then, fixing the number of topics K ex ante and using Gibbs sampling one can iteratively approximate the hidden LDA model using the Markov Chain Monte Carlo approach converging to the target distribution by repeated sampling from it (Griffith and Steyvers, 2004). Note also that the number of documents to be used in each training chunk of the LDA algorithm (so-called ’chunksize’) was set to 1800, the number of passes through the corpus during training (’passes’) equal to 20 and the maximum number of iterations through the corpus when inferring the topic distribution of a corpus (’iterations’) equal to 1000. All other parameters were set to default values. Fig. B13. Graphical model representation of LDA (based on Blei et al. (2003)). 36 In our sensitivity analysis we experimented with setting either both these parameters or just one of them to higher values that are popular in the literature, such as α = 0.1 or β=0.1 or β=0.01 (Chen et al., 2017; Huang et al., 2017; Kaplan and Vakili, 2015). However, those alternatives resulted in worse performance of topics in terms of perplexity and exclusivity (discussed later in this paper), and so we stick to the default setting. 17 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 B3. Determining the number of topics The perplexity score measures the ability of an LDA model estimated on a subset of documents (training data) to predict the word distribution in the remaining subset of documents (testing data). It is defined as the exponential of the negative normalized predictive likelihood under the model. A lower perplexity score indicates that the model has better generalization performance. Formally, for a testing data (Dtest ) with T1 documents, the perplexity score is equal to ∑T1 logp(wd ) − ∑ d=1 T1 (2) N perplexity score(Dtest ) = e d=1 d , where Nd is the number of words in document d, wd is a vector of all the words in document d, and p(wd ) is the probability of observing the word vector wd in document d given the LDA model estimated from the training data. Semantic coherence is a criterion developed by (Mimno et al., 2011) and is maximized when the most probable words in a given topic frequently co-occur together. The authors show that the metric correlates with human judgment of topic quality. Formally, let D(v, v ) be the number of times that ′ words v and v appear together in the same document. Then for a list of the M most probable words in topic k, the semantic coherence for topic k is ′ given as: ∑ M ∑i− 1 ( ) D(νi , ν) + 1 Ck = log ( ) (3) i=2 j=1 D νj A smoothing count of 1 is included to avoid taking the logarithm of zero. Finally, drawing on previous work on exclusivity and diversity in topic models (Bischof and Airoldi, 2012), the exclusivity of a topic measures if words with high probability under topic i have low probabilities under other topics. If so, we can conclude that topic i is exclusive. A topic that is both cohesive and exclusive is more likely to be semantically useful. More concretely, for topics Tn ∈ T and words wn ∈ Ti , we measure the exclusivity by: ∑∑ ∑ ∑ 1 exclusivity = |T| × (4) T1 ∈T w1 ∈T1 T ∈T\T w2 ∈T2 ∩T1 weight(w 1 ) × weight(w2 ) 2 1 B4. Plotting diffusion curves for LDA topics To produce diffusion curves, we follow Lenz and Winker (2020) in quantifying the probability that a topic Ti appears in the corpus of patent texts C for a certain year t as: ∑ P(Ti , d) P(Ti , Ct ) = d∈Ct , (5) Dt where d stands for patent documents in which we sum the prevalences of that topic in that time period, normalized by the overall number of doc­ uments in period t, Dt . Hence, at any period of time, the probabilities of all topics sum up to one. In addition, we smooth the diffusion curve by applying a five-year moving average window so that at any period the equation (5) is estimated not just for the year t, but also for two years preceding t and two years following it. Furthermore, to ease visual inspection of the diffusion curves, we classify the lines into three categories. To be classified as a ’rising’ topic (with green color), the following inequality should hold: ∑ ∑ ∑ ΔP(Ti , Ct ) = (P(Ti , Ct ) − P(Ti , Ct− 1 ))〉υ × |ΔP(Ti , Ct )| (6) where υ ∈ [0,1]. In other words, over the period of consideration the topic should predominantly rise and not fall. Reversely, to be classified as ’falling’ topic (with red color), the inequality in (6) must hold with the opposite sign. If none of the two conditions is fulfilled, the topic is classified as ’in- between’ (with blue color).37 B5. Topic matching Technically, the matching procedure consists of the following steps: 1. Define the reference texts: take IFR report on SR (IFR, 2018b). Lemmatize these texts and replace capital letters with small letters. This results in reference texts covering 49 sub-areas (compare Table 1). 2. Identify ’candidate words’: search for words which appear simultaneously among 100038 frequent words for each topic and in the SR reference texts. 3. For each candidate word wj of topic t and a reference text q, determine a word matching score according to (7): 37 The value of υ is chosen to be equal to 0.3 to separate the diffusion curves into the three groups to be as distinct as possible. As long as υ is positive and not very close to zero (to avoid classifying as rising those topics where the sum of ups and downs is approximately the same) it does not affect our later conclusions. 38 We look on the 1000 words having highest probability to be assigned to each individual topic. Looking beyond 1000 most likely words does not improve the results as the probabilities of the remaining words are virtually equal to zero. 18 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 ( ⃒) ( ) relevance wj ⃒t ∑n 1 word matching score wj , t, q = θ− 1 ( ) × (7) c × ρ wj i=1 i where ⃒ • relevance(wj ⃒t) is the relevance of the word to the topic according to (8) Sievert and Shirley (2014), which captures how exclusive this word is in this topic. ⃒ Let p(wj ⃒t) be the probability of word j in topic t as resulted by LDA and p(wj ) the marginal probability of word j in the empirical distribution. p(wj |t) The lift is defined as the ratio of a words’ probability in a topic to its marginal probability across the corpus, p(wj ) (Taddy, 2012). Then the relevance of word j for topic t is defined as ( ( ⃒⃒ )) ( ⃒) ( ( ⃒ )) p wj t relevance wj ⃒t = λ × log p wj ⃒t + (1 − λ) × log ( ) , λ ∈ [0, 1] (8) p wj We chose λ equal 0.5, meaning equal weighting of the word’s probability for topic and its lift. The idea is to reduce the weight of words appearing in other topics. • θ ∈ [1, 49] is the total number of SR sub-areas the word occurs in. This way we discriminate words appearing in multiple SR sub-areas applying a nonlinear penalty cθ− 1 . We experimented with several scalars for c above one. Larger values of c apply a stronger penalty on words that appear in multiple SR sub-areas reference texts, thus radically reducing the weight of these words and thus shifting the focus to words which are more unique for particular SR sub-areas. We decided for c = 4 as increasing it further was not changing the results. as • ρ is a generality score of the word based on its frequency in the English language and derived from the word frequency database composed by Mark Davies, which describes how often a word appears in the English language per million words. We normalize this frequency by our corpus’ length and take its log to derive ρ for each word we matched.39 This way we take into account that the reference texts represent only a fraction of all technologies (i.e. not covering IR and technologies outside the field robotics) and do not sufficiently penalize common words being matched by our algorithm (e.g. ’cup’). • n is the number of times the word appears in the same SR sub-area. Thus, every time a word appears in the same reference text, its marginal score will be increased with a declining margin based on the frequency it appeared.40 Thus, the word matching scores in equation (7) are unique for each word Wj and each topic – SR sub-area pair. 4. Determine the topic matching score of a topic t belonging to a reference text q by summing the word matching scores of all relevant candidate words appearing in topic t and a reference text q: ∑ ( ) topic matching score(t, q) = word matching score wj , t, q (9) wj ∈t∩q 5. Derive a threshold for the resulting topic matching scores. To this end, consider the distribution of values for each topic and each SR sub-area and choose one that maximizes the number of topics classified to at most one SR sub-area. The threshold value is calculated on the interval between zero and max(topic matching score(t,q)), where smaller values are associated with many topics matched to several SR sub-areas simultaneously (i.e. not uniquely), while increasing the threshold value reduces the amount of multiple matches until eventually each matched topic may be un­ equivocally be assigned to a single SR sub-area. This way, the threshold value is dependent on the distribution of topic matching scores: the higher the values of these matches indicating the overlap with multiple SR sub-areas, the higher the threshold value. We prefer to use this rule to define the threshold value as it best suits our purpose to identify as many topics capturing technologies from SR as possible, while remaining data-driven. Fig. 6 illustrates the derivation of the threshold value. B6. Network space of topics Remember that one of the outputs from applying LDA is a matrix of topic prevalences across the documents. Thus, each topic is a vector of weights, bounded between 0 and 1, of length D (number of our patents being22′ 927), so that cosine similarity of topics a and b is: ∑D ab i=1 ai bi Cosine similarityab = = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅ ∑D 2̅√̅̅̅̅̅̅̅̅̅̅̅̅̅̅ ∑D 2̅. (10) ‖a‖‖b‖ i=1 ai i=1 bi Appendix C. Further results C1. LDA model with 380 topics 39 https://www.wordfrequency.info/. The list of around100′ 000 words and their forms together with the frequency of use in the academic literature of the corpus of contemporary American English. 40 Thus, the first time the word appears, its score is one, whereas the n-th time the word appears, its score is 1n. 19 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 Fig. C14. Word clouds of 16 most central topics belonging to hardware cluster among 380 topics. Note: Topics are listed in the order of centrality. Topic 223 is SR topic. Fig. C15. Word clouds of 16 most central topics belonging to software cluster among 380 topics Note:Topics are listed in the order of centrality. Topic 191 is SR topic. 20 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 Fig. C16. Word clouds of 21 most central topics among 380 topics Note: Topics are listed in the order of centrality. 21 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 Fig. C17. Word clouds of isolated topics among 380 topics Note: Topics 37, 83, 353 and 377 belong to SR. 22 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 Fig. C18. Word clouds of 380 topics. 23 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 Fig. C19. Diffusion curves for 380 topics . 24 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 C2. LDA model with 190 topics Table 4 Topics from LDA model with K = 190 matched to SR sub-areas. For each topic, its top 30 most relevant (i.e. with high likelihood and exclusivity, see Eq. 8 for formal definition) words are displayed. Up-arrows indicate that the diffusion of the respective topic follows a positive trend, vice versa for down-arrow (which only apply for matches within the 190 topic model). Topics within an area are sorted by their topic matching scores. Topic Topic terms (30 most relevant) Matched SR sub-area Field Robotics 105 test, testing, tester, crop, agricultural, plant, harvest, tested, farm, harvesting, contactor, fertilizer, prober, actor, Agriculture nondestructive_testing, pest, retinal, nutrient, soil, retinal_ganglion, destructive_testing, stimulus, pesticide, subsurface, farmer, corn, flower, probers, tractor, aerial_farm, 109 box, rack, milking, milk, auxiliary, teat_cup, animal, teat, milk_animal, teat_animal, animal_milk, pinion, scraper, Milking robots and livestock robotics dairy_animal, automatically, talus, gouging, instructor, microbial_agent, esf, mtp, relevant, compartment, console, subtalar, implement, head_mount_display, rear, watch, farmer, 56 steer, steering, guidance, vehicle, lane, car, run, running, road, self_propelled, parking, driver, travelling, steerable, park, Agriculture cage, pedal, weighted_constant, race, wheel, driverless, reconfigurable_toy, droplet_ejection, passenger, amusement, propulsion, lateral, ahead, appraisal, brake_pedal, 40 apparatus, candidate, spatial, fruit, accord, seventh, eighth, element, ninth, discriminant_plane, trajectory, region, predicted, Agriculture sign_language, sixth, tenth, eleventh, condition_satisfied, twelfth, move, recognize, satisfy, thirteenth, condition, possibility, shade, action, predict, toner, tree, Inspection and maintenance systems 134 pipe, pipeline, inner, filter, outer, piping, main, branch_pipe, pip, diameter, yoke, lining, pig, liner_bag, smoking_article, Tank, tubes, pipes and sewers circumferential, lay, leak, oil, underground, tubular, inside, rig, wall, interferogram, liner, dig, laying, interferometry, flux_leakage, 14 ↓ seam, bead, weld, sealant, junction, pool, penetration, puddle, external_recharging, wax, seam_tracker, fusion, Tank, tubes, pipes and sewers pseudo_random, quality, weldment, gas_metal_arc_welding, butt, profiler, irregularity, corrugate, join, geometry, lateral, molten, slag, orbital_welder, gtaw, width, toolspeed, conformation, 159 air, powder, tank, flow, water, duct, pump, booth, enclosure, hopper, jet, discharge, exhaust, compressed_air, blast, Tank, tubes, pipes and sewers pressurized, intake, stream, decontamination, orifice, spray_booth, airflow, flow_rate, fan, blower, curtain, blasting, interior, cyclone, mist, Construction and demolition 129 panel, jig, side, roof, automotive, automobile, hanger, window_glass, assemble, trunk_lid, wheelchair, recovery_powder, Heavy/civil construction fresh_powder, framing, station, retinal_vein, fender, underbody, car, preassembly, cowl, line, enhancement_understanding, toilet, exclusive, forcel, mule, inner, bonnet, ophthalmol, Logistic systems 158 container, pallet, load, station, loading, unload, unloading, transport, lift, warehouse, pack, palletizing, stock, handle, carton, Automated Guided Vehicles (AGVs) in handling, empty, truck, workstation, good, delivery, palletizer, palletized, load_unload, loaded, hoist, loading_unloading, manufacturing environments shipping, shelf, beverage, 180 ↑ location, navigation, identify, identification, beacon, radio, cod, map, identity, rfid_tag, global_positioning_system, Automated Guided Vehicles (AGVs) in radio_frequency, antenna, indoor, determine, rfid, wireless, destination, transceiver, navigate, transponder, transmitter, manufacturing environments processor, identifier, information, system, dead_reckoning, message, include, identified, Medical robotics 47 ↑ catheter, medical, distal, patient, proximal, distal_end, sled, proximal_end, sheath, lumen, heart, elongate, bed, sterile, Robot-assisted surgery and therapy flexible, catheter_sheath, diagnostic_therapeutic, sterile_barrier, cardiac, elongated, handle, nose_cone, manipulation, introducer, ablation, stabilizer, sterility, steerable_catheter, therapy, ventricle, 43 panoramic, imager, laparoscope, anatomical, bypass_graft, fade, zoom, artery, coronary_artery, colon, Robot-assisted surgery and therapy secondary_containment, tract, tremor, anastomosis, swarm, insufflate_gas, heart, scissors_stapler, clamp_graspers, balloon, blood, endoscopy_laparoscopy, pavement, practitioner, prosthetist, bagger, surgery_endoscopy, file, pain, vital_sign, 131 ↑ brain, tissue, tumor, neural, therapy, stereotactic, nerve, cancer, organ, mri, stimulation, biopsy, dose, clinical, Robot-assisted surgery and therapy radiation_therapy, lesion, compute_tomography, patient, nervous, prostate, vol, spinal_cord, sobot, emg, needle, skull, computed_tomography, multiblock, ultrasound, magnetic_resonance, 110 ↑ instrument, surgical, surgeon, surgery, surgical_instrument, minimally_invasive, surgical_procedure, endoscope, surgical_site, Robot-assisted surgery and therapy patient, incision, tissue, medical, endoscopic, laparoscopic, medical_procedure, telesurgical, forceps, telesurgery, trocar, cannula, laparoscopic_surgery, procedure, abdominal_cavity, robotic, intuitive_surgical, abdominal_wall, console, entry, organ, 126 ↑ virtual, simulation, interface, user, simulated, feedback, simulate, haptic, physical, simulator, virtual_reality, interaction, Other medical robots sensation, stylus, tactile, mouse, haptic_feedback, graphical, training, feel, cursor, haptic_interface, articulatable, environment, system, real, realistic, computer, experience, visual, Powered human exoskeletons 169 ↑ joint, actuator, tendon, shoulder, actuation, muscle, wearer, rehabilitation, exoskeleton, passive, linkage, revolute_joint, Powered human exoskeletons degree_freedom, actuate, universal_joint, limb, ball_socket, jointed, artificial, human, elbow, mechanical, autonomous_and_remote_control_all_purpose_machine, skeletal, actuated, muscular, dofs, exercise, expansible, leg_hopping, Robots for domestic task 72 path, movement, along, traverse, automatic_guided_vehicle, swath, follow, spatial, curvature, mow, implement, contour, Lawn-mowing refinement, optoelectronic, traversal, transformation_affine, supplementary, chess, automatic_guided_vehicles, move, mwo, centerpoint, prepositioning, desired, jolt, system, pas, matrix_mow, anticipated, ceramic_optoceramic, 97 lens, emitter, coverage, receiver, perimeter, boundary, hydration, cliff, debris, ophthalmic_lens, lenses, eyeglass, sonar, Lawn-mowing polymerization, confinement, gateway, deterministic, monomer, piperazine, ester, airlaid_layer, pat_larsen, emission, mow, implementation, maskant, carbon_atom, spectacle_lens, acid, lawn, 6↑ cleaning, clean, brush, floor, cleaner, collecting, waste, collect, carpet, scrub, collection, chassis, fan, dirty, bristle, Vacuuming, floor cleaning loose_particulate, intake_port, squeegee, scrubbing, scrubber, fluid, liquid, contaminant, smear, sponge, surface, epsilon_sup, aft, dust, wet, 39 ↑ wheel, main_body, cleaner, chassis, dust, dock, deck, docking, docking_station, flipper, climb, caster, dust_collector, Vacuuming, floor cleaning foreign_substance, payload_deck, omnidirectional_wheel, wheeled, axle, spoke, dirt, agitator, front, rotate, bristle_row, include, spacer_grid, dust_dirt, turnover, climbing, suspension, 25 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 Fig. C20. Word clouds of nine most central topics belonging to hardware cluster among 190 topics Note: Topics are listed in the order of centrality. Fig. C21. Word clouds of nine most central topics belonging to software cluster among 190 topics Note: Topics are listed in the order of centrality. 26 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 Fig. C22. Word clouds of nine most central topics among 190 topics Note: Topics are listed in the order of centrality. Fig. C23. Word clouds of 21 isolated topics among 190 topics. 27 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 Fig. C24. Word clouds of 190 topics. 28 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 Fig. C25. Diffusion curves for 190 topics. 29 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 References Graetz, G., Michaels, G., 2018. Robots at work. The review of economics and statistics 100 (5), 753–768. https://doi.org/10.1162/rest_a_00754. Griffith, T., Steyvers, M., 2004. Finding scientific topics. Proceedings of the National Acemoglu, D., Lelarge, C., Restrepo, P., 2020. Competing with robots: firm-level Academy of Sciences 101, 5228–5235. evidence from France. AEA Papers and Proceedings 110, 383–388. https://doi.org/ Grün, B., Hornik, K., 2011. Topicmodels: an R package for fitting topic models. Journal 10.1257/pandp.20201003. of statistical software 40 (13), 1–30. Acemoglu, D., Restrepo, P., 2018. Artificial intelligence, automation and work. National Huang, A.H., Lehavy, R., Zang, A.Y., Zheng, R., 2017. Analyst information discovery and Bureau of Economic Research Working Paper 24196. https://doi.org/10.3386/ interpretation roles: a topic modeling approach. Management science 64 (6), w24196. 2833–2855. Acemoglu, D., Restrepo, P., 2019. Automation and new tasks: how technology displaces IFR, 2018. World Robotics – Industrial Robots. Technical Report. International and reinstates labor. Journal of Economic Perspectives 33 (2), 3–30. https://doi.org/ Federation of Robotics. 10.1257/jep.33.2.3. IFR, 2018. World Robotics – Service Robots. Technical Report. International Federation Aghion, P., Jones, B.F., Jones, C.I., 2017. Artificial Intelligence and Economic Growth. of Robotics. Technical Report. NBER. IFR, 2019. IFR press conference 2019. Shanghai. https://www.ifr.org/downloads/press Agrawal, A., Gans, J.S., Goldfarb, A., 2019. Artificial intelligence: the ambiguous labor 2018/IFR-World-RoboticsPresentation - 18 Sept 2019.pdf. market impact of automating prediction. Journal of Economic Perspectives 33 (2), Jaffe, A., Trajtenberg, M., 1996. Flows of knowledge from universities and federal 31–50. https://doi.org/10.1257/jep.33.2.31. laboratories: modeling the flow of patent citations over time and across institutional Alcácer, J., Gittelman, M., 2006. Patent citations as a measure of knowledge flows: the and geographic boundaries. Proceedings of the National Academy of Sciences 93, influence of examiner citations. Review of Economics and Statistics 88 (4), 774–779. 12671–12677. Ambrosino, A., Cedrini, M., Davis, J.B., Fioria, S., Guerzoni, M., Nuccio, M., 2018. What Jaffe, A., Trajtenberg, M., Henderson, R., 1993. Geographic localization of knowledge topic modeling could reveal about the evolution of economics. Journal of Economic spillovers as evidenced by patent citations. Quarterly Journal of Economics 108, Methodology 25 (4), 367–377. 577–598. Analytics, C., 2018. Scientific Literature and Patent Analysis of Robotics. Technical Kaplan, S., Vakili, K., 2015. The double-edged sword of recombination in breakthrough Report. Clarivate Analytics. innovation. Strategic Management Journal 36, 1435–1457. Atack, J., Margo, R.A., Rhode, P.W., 2019. “Automation” of manufacturing in the late Kim, J., Lee, J., Kim, G., Park, S., Jang, D., 2016. A hybrid method of analyzing patents nineteenth century: the hand and machine labor study. Journal of Economic for sustainable technology management in humanoid robot industry. Sustainability 8 Perspectives 33 (2), 51–70. https://doi.org/10.1257/jep.33.2.51. (5), 1–14. Autor, D.H., Dorn, D., 2013. The growth of low-skill service jobs and the polarization of Korzinov, V., Savin, I., 2018. General purpose technologies as an emergent property. the US labor market. American Economic Review 103 (5), 1553–1597. https://doi. Technological forecasting and social change 129, 88–104. org/10.1257/aer.103.5.1553. Kreuchauff, F., Korzinov, V., 2017. A patent search strategy based on machine learning Baldwin, R., Forlsid, R., 2020. Globotics and Development: When Manufacturing is for the emerging field of service robotics. Scientometrics 111 (2), 743–772. https:// Jobless and Services are Tradeable. Technical Report. NBER. doi.org/10.1007/s11192-017-2268-3. Baumol, W., 2012. The Cost Disease: Why Computers Get Cheaper and Health Care Lafond, F., Kim, D., 2019. Long-run dynamics of the U.s. patent classification system. Doesn’t. Yale University Press, New Haven. Journal of Evolutionary Economics 29 (2), 631–664. Baumol, W.J., Bowen, W.G., 1966. Performing Arts – The Economic Dilemma. Twentieth Larsen, V.H., Thorsrud, L.A., 2019. Business Cycle Narratives. Technical Report. CESifo Century Fund. Working Paper Series. Bessen, J., Goos, M., Salomons, A., van den Berge, W., 2020. Firm-level automation: Lee, W.S., Han, E.J., Sohn, S.Y., 2015. Predicting the pattern of technology convergence evidence from the Netherlands. AEA Papers and Proceedings 110, 389–393. https:// using big-data technology on large-scale triadic patents. Technological forecasting doi.org/10.1257/pandp.20201004. and social change 100, 317–329. Bischof, J., Airoldi, E., 2012. Summarizing topical content with word frequency and Lenz, D., Winker, P., 2020. Measuring the diffusion of innovations with paragraph vector exclusivity. In: Langford, J., Pineau, J. (Eds.), Proceedings of the 29th International topic models. PloS one 15 (1), e0226685. https://doi.org/10.1371/journal. Conference on Machine Learning (ICML-12). New York, NY: Omnipress, pone.0226685. pp. 201–208. Liu, L., Tang, L., Dong, W., Yao, S., Zhou, W., 2016. An overview of topic modeling and Blei, D.M., 2012. Probabilistic topic models. Communications of the ACM 55 (4), 77–84. its current applications in bioinformatics. SpringerPlus 5 (1), 1608. Blei, D.M., Ng, A.Y., Jordan, M.I., 2003. Latent dirichlet allocation. The Journal of Lüdering, J., Winker, P., 2016. Forward or backward looking? the economic discourse Machine Learning Research 3, 993–1022. and the observed reality. Journal of Economics and Statistics 236 (4), 483–515. Bouma, G., 2009. Normalized (pointwise) mutual information in collocation extraction. Mansfield, E., 1989. The diffusion of industrial robots in Japan and the United States. Proceedings of the International Conference of the German Society for Research policy 18, 183–192. Computational Linguistics and Language Technology, GCSL 09.. Michel, J., Bettels, B., 2001. Patent citation analysis: a closer look at the basic input data Boyd-Graber, J., Hu, Y., Mimno, D., 2017. Applications of topic models. Foundations and from patent search reports. Scientometrics 51, 185–201. Trends in Information Retrieval 11 (2–3), 143–296. Mimno, D., Wallach, H.M., Talley, E., Leenders, M., McCallum, A., 2011. Optimizing Caselli, M., Fracasso, A., Traverso, S., 2020. Mitigation of risks of Covid-19 contagion semantic coherence in topic models. Proceedings of the Conference on Empirical and robotisation: Evidence from Italy. In: Press, C. (Ed.), Covid Economics: Vetted Methods in Natural Language Processing. PA: Association for Computational and Real-time Papers, pp. 174–188. Linguistics, pp. 262–272. Chae, B.K., Park, E.O., 2018. Corporate social responsibility (CSR): a survey of topics and Napolitano, L., Evangelou, E., Pugliese, E., Zeppini, P., Room, G., 2018. Technology trends using twitter data and topic modeling. Sustainability 10 (7), 1–20. networks: the autocatalytic origins of innovation. Royal Society open science 5, Chang, H.C., 2016. The synergy of scientometric analysis and knowledge mapping with 172445. topic models: modelling the development trajectories of information security and Newman, D., Noh, Y., Talley, E., Karimi, S., Baldwin, T., 2010. Evaluating topic models cyber-security research. Journal of Information & Knowledge Management 15 (4), for digital libraries. Proceedings of the 2010 Joint Conference on Digital Libraries 1650044. (JCDL), pp. 215–224. Chen, B., Marvin, S., While, A., 2020. Containing COVID-19 in China: AI and the robotic Newman, M.E.J., 2006. Finding community structure using the eigenvectors of matrices. restructuring of future cities. Dialogues in Human Geography forthcoming. https:// Physical Review E 74, 036104. doi.org/10.1177/2043820620934267. Pugliese, E., Cimini, G., Patelli, A., Zaccaria, A., Pietronero, L., Gabrielli, A., 2019. Chen, H., Zhang, G., Zhu, D., Lu, J., 2017. Topic-based technological forecasting based Unfolding the innovation system for the development of countries: co-evolution of on patent data: a case study of Australian patents from 2000 to 2014. Technological science, technology and production. Scientific reports 9 (16440). forecasting and social change 119 (C), 39–52. Roberts, M.E., Stewart, B.M., Tingley, D., Lucas, C., Leder-Luis, J., Gadarian, S.K., Cheng, H., Jia, R., Li, D., Li, H., 2019. The rise of robots in China. Journal of Economic Albertson, B., Rand, D.G., 2014. Structural topic models for open-ended survey Perspectives 33 (2), 71–88. https://doi.org/10.1257/jep.33.2.71. responses. American journal of political science 58 (4), 1064–1082. Dauth, W., Findeisen, S., Suedekum, J., Woessner, N., 2018. Adjusting to robots: Worker- Saracco, F., DiClemente, R., Gabrielli, A., Squartini, T., 2015. Randomizing bipartite level evidence. 10.21034/iwp.13. networks: the case of the world trade web. Scientific reports 5, 10595. Dauth, W., Findeisen, S., Suedekum, J., Woessner, N., 2019. The Adjustment of Labor Savin, I., van den Bergh, J., 2021. Main topics in EIST during its first decade: Markets to Robots. Techreport. University of Würzburg.https://sfndsn.github.io/do acomputational-linguistic analysis. Environmental Innovation and Societal wnloads/AdjustmentLaborRobots.pdf Transitions 41, 10–17. De Battisti, F., Ferrara, A., Salini, S., 2015. A decade of research in statistics: a topic Savin, I., Drews, S., van den Bergh, J., 2021. Free associations of citizens and scientists model approach. Scientometrics 103 (2), 413–433. with economic and green growth: a computational linguistics analysis. Ecological EPO, 2017. Patents and the Fourth Industrial Revolution – The Inventions behind digital Economics 180, 106878. https://doi.org/10.1007/s10584-020-02842-y. transformation. Technical Report. European Patent Office. Savin, I., Drews, S., Maestre-Andres, S., van den Bergh, J., 2020. Public views on carbon Eurostat, 2014. Patent Statistics: Concordance IPC V8 - NACE rev.2. Technical Report. taxation and its fairness: a computational-linguistics analysis. Climatic change 162, Eurostat.https://circabc.europa.eu/sd/a/d1475596-1568-408a-9191-426629047e 2107–2138. 31/2014-10-16-Final&20IPC_NACE2_2014.pdf Sievert, C., Shirley, K., 2014. LDAvis: A method for visualizing and interpreting topics. Farrell, J., 2016. Corporate funding and ideological polarization about climate change. Proceedings of the Workshop on Interactive Language Learning, Visualization, and Proceedings of the National Academy of Sciences 113 (1), 92–97. Interfaces. Association for Computational Linguistics, Baltimore, Maryland, USA, Freeman, L.C., 1979. Centrality in social networks conceptual clarification. Social pp. 63–70. https://doi.org/10.3115/v1/W14-3110.https://www.aclweb.org/ant Networks 1 (3), 215–239. hology/W14-3110 Goeldner, M., Herstatt, C., Tietze, F., 2015. The emergence of care robotics - a patent and publication analysis. Technological forecasting and social change 92, 115–131. https://doi.org/10.1016/j.techfore.2014.09.005. 30 I. Savin et al. Technological Forecasting & Social Change 174 (2022) 121280 Suominen, A., Toivanen, H., Seppänen, M., 2017. Firms’ knowledge profiles: mapping Zeng, Z., Chen, P.-J., Lew, A.A., 2020. From high-touch to high-tech: COVID-19 drives patent data with unsupervised learning. Technological forecasting and social change robotics adoption. Tourism Geographies. https://doi.org/10.1080/ 115 (9), 131–142. 14616688.2020.1762118. Taddy, M., 2012. On estimation and selection for topic models. Artificial Intelligence and Statistics 1184–1193. Ivan Savin is a senior researcher at the Institute of Environmental Science and Technol­ Trajtenberg, M., 1990. A penny for your quotes: patent citations and the value of ogy, Universitat Autșnoma de Barcelona, Spain. Also Dr. Savin is a senior researcher at the innovations. RAND Journal of Economics 21, 172–187. Ural Federal University in Yekaterinburg, Russia. Ivan Savin completed his Ph.D. on Tvinnereim, E., Fløttum, K., 2015. Explaining topic prevalence in answers to open-ended computational econometrics at the Justus Liebig University in Giessen in 2011. In 2017 he survey questions about climate change. Nature climate change 5, 744–747. completed his habilitation at the Karlsruhe Institute of Technology. Ivan works on topics Tvinnereim, E., Fløttum, K., Gjerstad, O., Johannesson, M.P., Nordø, A.D., 2017. Citizens’ related to economics of innovation and structural change, economics of complexity and preferences for tackling climate change. quantitative and qualitative analyses of economics of networks. More recently Ivan Savin started also working on climate policy their freely formulated solutions. Global Environmental Change 46, 34–41. analysis. Valverde, S., Solé, R.V., Bedau, M.A., Packard, N., 2007. Topology and evolution of technology innovation networks. Physical Review E 76 (5), 056118. Venugopalan, S., Rai, V., 2015. Topic based classification and pattern identification in Ingrid Ott is a full professor at the Karlsruhe Institute of Technology (KIT) since 2010. She patents. Technological forecasting and social change 94, 236–250. is leading the chair in Economic Policy. Since 2008 she is also a research associate at the WIPO, 2008. Concept of a Technology Classification for Country Comparisons. Technical Kiel Institute for the World conomy (IfW). Between 2007 and 2009 she was a head of Report. World Intellectual Property Organization.https://www.wipo.int/export/sit research division New Technologies and Regional Innovation Systems at the Hamburg es/www/ipstats/en/statistics/patents/pdf/wipo_ipc_technology.pdf Institute of International Economics (HWWI). And between 2002 and 2007 assistant Yang, G.-Z., Nelson, B.J., Murphy, R.R., Choset, H., Christensen, H., Collins, S.H., professor at Leuphana University Lȭneburg, Chair in Innovation and Growth. Her research Dario, P., Goldberg, K., Ikuta, K., Jacobstein, N., Kragic, D., Taylor, R.H., McNutt, M., interests are in economics of innovation and economic policy. She also has a special focus 2020. Combating COVID-19 -the role of robotics in managing public health and on robotics and general purpose technologies. infectious diseases. Robots and Society 5 (40), eabb5589. https://doi.org/10.1126/ scirobotics.abb5589. Chris Konop is M.Sc. from the Karlsruhe Institute of Technology (KIT). He is specializing Youn, H., Strumsky, D., Bettencourt, L.M.A., Lobo, J., 2015. Invention as a combinatorial in natural language processing. process: evidence from US patents. Journal of The Royal Society Interface 12 (106), 20150272. 31