Patents and knowledge diffusion

.


Introduction
"English's emergence as the global language, along with the rapid progress in machine translation [. . . ] make it less clear that the substantial investment necessary to speak a foreign tongue is universally worthwhile. While there is no gainsaying the insights that come from mastering a language, it will over time become less essential in doing business [. . . ]." (Lawrence H. Summers -NY Times, 2012) One rationale for the existence of the patent system is that it fosters innovation by facilitating knowledge diffusion. Patents "promote the progress of science" (US Constitution art. I, § 8, cl. 8) by granting inventors a temporary monopoly on their inventions while ensuring the disclosure of the technical information on which the invention is based. For this reason, the patent system is often characterized as quid pro quo system in which the inventor trades the public release of new technical knowledge to society in exchange for the rights to its exclusive use. The scholarly debate, however, is far from consensus on the positive effect of patent disclosure on the pace of technical progress. Several scholars have questioned the legitimacy of disclosure theory as a justification for the existence of the patent system and argue that the disclosure requirement is not really working as planned for society (Lemley, 2012;Risch, 2007;Devlin, 2010). This paper aims to contribute to this discussion and shed light on the mechanism of knowledge diffusion through the patent system, focusing on cross-border and cross-language knowledge flows.
Historically, language differences represented one of the main barriers to the diffusion of scientific and technical knowledge. Over the last three centuries, scientists, researchers, and inventors have resorted to different solutions to facilitate communications between diverse language communities. The use of Latin as a lingua franca, the development of international auxiliary languages, such as Ido and Esperanto, and the publication of journal abstracts that systematically translated the abstracts of the contributions appearing in the most important scientific journals are only a few well-known examples of attempts to overcome language barriers to the progress of science (Gordin, 2015).
Even though it was not its primary objective, the progressive harmonization of patent law across countries was probably one of the main elements that fostered the diffusion of novel technical knowledge across language borders. In the current system, the legal protection granted by a patented invention could be extended to other jurisdictions within twelve months from the original application. 1 Most jurisdictions require the foreign application to be translated into one of the official languages of the country of the receiving patent authority. Therefore, a Japanese inventor who would like to obtain the right to exclude others from using her invention in the United States must translate her patent application into English and file it with the United States Trademark and Patent Office (USPTO). In such a case, the existence of the patent system not only pushes the inventor to publicly disclose the technical knowledge that should allow her invention to be reproduced, but the prospect of protection beyond the national context gives her the incentive to translate said knowledge, allowing for a potentially much larger interested audience.
Despite its potential relevance, the importance of mandated translation in fulfilling the objectives of the disclosure requirement has received very limited attention in the scholarly debate (Ouellette, 2017). This is especially surprising in an era in which increasingly sophisticated machine translation tools are lowering the costs of access to knowledge behind language barriers in an unprecedented way in human history. This paper aims to fill this gap in the literature and, more specifically, to assess whether improved access to the patent literature through automated translations has fostered the diffusion of technical knowledge.
To do so, we exploit a recent natural experiment. In 2013, Google launched a major improvement to Google Patents, adding documents from the Chinese Patent Officethe China National Intellectual Property Administration (CNIPA) -and using its own neural machine translation service to translate patents previously available to the public only on in Chinese via the Chinese patent office's website. Thanks to Google Patent's automatic translations, Chinese patents were made available and searchable in both their original Chinese version and in English. 2 This event provides an ideal research setting for evaluating whether the disclosure of technical knowledge through patents and their translation facilitates knowledge diffusion. We exploit this sudden change in the availability of translation of Chinese patents to implement a difference-in-differences analysis in which we evaluate whether US inventors make more references to knowledge embodied in Chinese patents after the automated translation provided by Google, relative to a set of suitable control patents. The control group is populated with patents issued by the Korean patent office.
We measure knowledge diffusion through patent citations. Although several studies have highlighted the limitations of using citations to capture knowledge spillovers due to their heterogeneous and strategic nature (Sampat, 2010;Lampe, 2012;Cotropia et al., 2013;Kuhn et al., 2020), we follow the recommendations provided in recent contributions and focus exclusively on the types of citations that are more likely to capture actual knowledge flows (Jaffe and de Rassenfosse, 2019;Corsino et al., 2019;Kuhn et al., 2020).
Our results show a positive effect of the machine translation of Chinese patents on knowledge flows. We observe an increase of about 7.2 % in citations received by Chinese patents after the translation. The effect is more pronounced in technological areas of which China is at the technological frontier, such as computing-and data processing-related technologies, with up to a 14.5 % increase in citations received by translated patents. We also show that machine translation particularly increases the likelihood of Chinese patents' receiving a citation from a small entity, a non-profit organization, or an independent inventor in the US.
Excluding citations from patents filed by US residents of Asian origins strengthens the aforementioned results. Furthermore, we find that inventors are more likely to cite translated patents that are considered more readable.
All in all, our results suggest that improving access to the technical knowledge disclosed in the patent literature through machine translations increases the likelihood of cross-border knowledge flows. These findings are in line with empirical literature that studies how institutions reducing the costs of access to existing knowledge facilitate knowledge dissemination (Furman et al., 2021;Furman and Stern, 2011;Agrawal and Goldfarb, 2008).
The mere production of knowledge does not itself ensure that others will be able to use it; effective diffusion requires both awareness of the existence of and the willingness to pay the cost of access to that knowledge (Furman and Stern, 2011;Mokyr, 2002). Language barriers increase the costs of both discovery and access to potentially relevant knowledge. We show that the machine translation of patent literature reduces these costs and is especially beneficial for resource-constrained inventors.
The rest of the paper is organized as follows. Section 2 provides the theoretical framework of the disclosure function of the patent system. Section 3 illustrates our identification strategy, and Section 4 provides an overview of the data and method used. Section 5 discusses the econometric results and a few robustness checks. Section 6 introduces two additional tests in support of the idea that machine translation facilitates actual knowledge diffusion. Section 7 offers a conclusion.

Patent disclosure and cumulative innovation
The promotion of innovation through the disclosure of technical knowledge is one of the pillars on which the patent system rests (Ouellette, 2012;Lemley, 2012;Furman et al., 2018). Patentees obtain exclusive rights to the commercial exploitation of their inventions and can earn quasi-rents on them for a limited time (Sampat and Williams, 2019). In return, the inventors must disclose their inventions to society so that science and technology can progress by building on the knowledge divulged (Ouellette, 2012;Fromer, 2009). To fulfill this function, in most jurisdictions, a patent application must include a precise written description of the invention that would allow a person trained in the relevant field to reproduce it. Indeed, disclosure is one of the requirements for which the TRIPS Agreement imposes a minimum standard for protection. Article 29.1 of the TRIPS states that "an applicant for a patent shall disclose the invention in a manner sufficiently clear and complete for the invention to be carried out by a person skilled in the art and may require the applicant to indicate the best mode for carrying out the invention." Several legal scholars and economists, however, have argued that the information disclosed in patent applications is of little use to inventors, as they do not learn their science from patents (Lemley, 2012;Boldrin and Levine, 2013). In particular, the works in this strand of literature suggest that patents are generally drafted using rather obscure and vague jargon aimed at hiding strategic information or at broadening the scope of the patent's claims (Risch, 2007;Devlin, 2010). If these claims were correct, at least a part of the patent bargain would not be satisfied. Society would be paying the costs that exclusivity entailsi.e., higher prices for innovated products and services and the consequential underconsumptionwithout receiving the benefits that patent disclosure should bring about in terms of knowledge diffusion. At present, the empirical evidence on the positive effect of patent disclosure on knowledge spillovers remains limited and does not fully rule out this possibility.
On the one hand, the empirical studies based on surveys administered to inventors provide mixed results. Jaffe et al. (2000) asked US patent inventors to list the most important inputs for the conception of 1 Patent application filed via the international route provided by the Patent Cooperation Treaty can be filed to a national office up to thirty months from the original application.
2 One of the declared objectives of Google's effort was to increase the discoverability of foreign inventions to improve the quality of patents in the US and worldwide.
their patented inventions. Only 5 % of the respondents identified the patent literature as having a significant influence on the invention process. A similar survey administered to R&D managers by Cohen et al. (2000) asked them to rank the relevance of different information channels to the completion of a recent R&D project. Results showed that patents were considered less important than other information sources like scientific publications and informal exchanges. More recent surveybased studies paint a more optimistic picture for disclosure theory. Based on a survey administered to researchers active in the nanotechnology field, Ouellette (2012) shows that 64 % of the respondents have read at least one patent for research purposes. In a follow-up study based on a survey administered to a more diverse group of researchers, Ouellette (2017) finds that only a minority of respondents never read a patent and that 60 % of the patent readers found useful scientific information in the most recent patent they read. On the other hand, a recent strand of research based on quasiexperimental approaches has produced growing empirical evidence of a positive effect of patent disclosure on innovation (Furman et al., 2018;Baruffaldi and Simeth, 2020;de Rassenfosse et al., 2020). Furman et al. (2018) investigate the effect of information disclosure through patents on subsequent innovation and exploit the opening of new patent libraries in the pre-internet era for identification. They find that the improved access to the patent literature made possible by the opening of a library in a given region lead to a 17 % increase in local patenting relative to suitable control regions. Furman et al. (2018) firmly suggests that the increase in patenting is driven by the disclosure of technical information in the patent documents. Baruffaldi and Simeth (2020), on the other hand, focus on the effect of the early disclosure of patent applications and investigate the impact of a policy change that affected their publication time in the US. In 2000, the American Inventors Protection Act reduced the default publication time for patent applications filed at the United States Patent and Trademark Office (USPTO) to eighteen months. Their findings confirm the importance of early disclosure in facilitating information diffusion, highlighting that technological knowledge may become obsolete quickly. In addition, they find the effect largely unaffected by geographical distance. The results of Baruffaldi and Simeth (2020) imply that knowledge diffusion increases as the time frame between the invention and its disclosure decreases, highlighting the importance of the timing of disclosure to the usefulness of the disclosed information for future inventions.
de Rassenfosse et al. (2020) evaluate the extent to which knowledge flows are disrupted when a patent application to the USPTO is temporarily kept secret because of national security concerns. 3 Their analysis shows that patented inventions cited by a patent that is subject to a secrecy order received, on average, between 30 and 50 % fewer forward citations than a group of suitable control patents during the period in which the secrecy order was in place. These findings suggest that secrecy orders hinder the cumulative effects in knowledge production, highlighting the importance of patent disclosure to enable follow-on invention.
All in all, these recent studies appear to confirm that the disclosure function of the patent system is working as planned. Inventors seem to be using the knowledge disclosed in patent documents as input for their inventive efforts. In addition, patent disclosure might also benefit inventors and society in ways that go beyond the direct provision of inputs in the invention process. Besides promoting incremental inventions, disclosure also facilitates efficient bargaining by clarifying property rights and helping identify the bounds of a patent application's scope (Devlin, 2010). Hence, good knowledge of the technical and legal information disclosed in the prior art helps applicants draft more sustainable patents in terms of validity and enforceability (Sampat, 2010).
Still, most of the works that have analyzed the benefits (or lack thereof) of disclosure have focused on the flows of knowledge that patent disclosure allows within a single jurisdiction. Ouellette (2017) suggests instead that patent disclosure might be particularly beneficial for improving access to knowledge that would otherwise be inaccessible to local inventors. In particular, patents might be especially useful in facilitating information flows across language borders through mandatory translations of knowledge that would be otherwise trapped behind a language barrier. Anecdotal evidence supports this hypothesis. For instance, the adoption of one of the most fundamental innovations in coronary angioplasty treatment, the balloon catheter developed by Andreas Grüntzig, was delayed because the authors described the new method in an article published exclusively in German. The publication could be read only by individuals proficient in the German language and was restricted to German-speaking countries (Husmann and Barton, 2014). The global adoption of this technique followed the filing of several patent applications for the balloon angioplasty device in Switzerland, Germany, France, the United Kingdom, the United States, and Japan. In a recent paper, Choudhury and Kim (2019) exploit a natural experiment to find that ethnic migrant inventors are instrumental in transferring contextual knowledge across borders. They constructed a dataset of herbal patents to evaluate whether knowledge of Chinese and Indian herbal medicine is transferred to the West by firstgeneration migrant Chinese and Indian scientists. Choudhury and Kim (2019) find that an increase in the number of first-generation ethnic migrant inventors increases the rate of codification of herbal knowledge by US patent assignees by 4.5 %. Their results indicate that knowledge locked within specific cultural regions becomes accessible only through migrants capable of speaking the language, confirming the importance of language as a barrier to knowledge flows.
The benefits of patent disclosure may not exclusively derive from the requirements imposed within a single jurisdiction but also from the existence of an international patent system that builds on a broadly harmonized patent law. The Paris Convention for the Protection of Industrial Property, one of the first international treaties on intellectual property matters, establishes the right to priority. This right means that, on the basis of a regular first application filed with the patent authority of one of the contracting states, the applicant may apply for protection in any of the other Contracting States within twelve months from the first filing and that these subsequent applications will be examined as if they were filed on the same date as the original application. 4 The Patent Cooperation Treaty extends the time window for filing subsequent applications in different jurisdictions to thirty months. 5 Each jurisdiction requires the extended application to be filed in accordance to the local filing rules and, where applicable, this entails the translation of the original application in (one of) the official language(s) of the receiving patent authority. The additional time that the treaties grant to applicants for extending their applications abroad is specifically intended to provide enough time for the applicants to translate and adjust their patent applications to the requirements of a foreign patent office. In addition, all the main international IP treaties in force todaythe Paris Convention, the PCT, and the TRIPSimpose the "national treatment" principle, which states that, within each jurisdiction, foreign applicants must receive treatment equal to that accorded to domestic applicants. 6 3 Secrecy orders are imposed by the Commissioner of Patents in accordance with the Invention Secrecy Act of 1951. 4 The Convention was originally signed in 1883, and it is still in force today with 177 signatory members. See https://www.wipo.int/treaties/en/ip/paris/. 5 The PCT is an international treaty making it possible to seek patent protection for an invention simultaneously in a large number of countries. On the one hand, the progressive harmonization of patent law ensures that, in most countries, applicants should comply with the disclosure requirements. On the other, the existence of an international patent system provides a framework in which inventors can safely extend the protection granted to their inventions beyond their domestic jurisdiction. All in all, the role of the patent system in fostering knowledge diffusion across countries might have been underappreciated so far, and much of this effect might be driven by the translation of knowledge that would be otherwise trapped behind a language barrier. This lack of attention toward the role of patent translation is even more surprising in an era in which advances in computer science have made machine translation tools highly sophisticated and lowered the costs of access to knowledge codified in a foreign language. While research exists on the precision and improvements of machine translation in general, its reliability for translations of patent documents or highly technological information has been only recently assessed. 7 Research by Zulfiqar et al. (2018) evaluates the accuracy of Google Translate, among others, with an emphasis on German scientific literature. 8 Their study shows that the service is reliable and an instantaneous tool for translating "not only short phrases but even large passages" into English. Focusing on the disclosure of patent information, research done by Larroyed (2018), shows the effectiveness of machine translation. Larroyed (2018) estimated the level of disclosure at almost 80 % of the patent content between Western languages and almost 70 % for Chinese to English translations. The estimate is based on a selection of 100 patents related to clean technologies applied for between 2013 and 2015. Based on the LISA (Localization Industry Standards Association) Quality Assessment, the machine translation of these patents was evaluated against the manual translation done by native speakers. In addition, the translated documents were blind-reviewed by persons skilled in the art. The translation was evaluated based on multiple factors, such as accuracy, terminology, syntax, and style, followed by the classification of errors as minor, major or critical. The average number of errors from both evaluations was 22.75 for Western languages, resulting in a disclosure of 80 % of the information. Although the author states that machine translation still needs improvement, it "clearly discloses patent information" and represents one of the main tools of communication within the patent system. Indeed, machine translation is increasingly used by patent specialists, companies, and even patent authorities. The examination guidelines of the European Patent Office currently state that "in order to overcome the language barrier constituted by a document [. . . ], it might be appropriate for the [patent] examiner to rely on a machine translation of said document" (EPO, 2021a). 9 In short, one of the channels through which patent disclosure positively affects knowledge diffusion might be the translation of otherwise unavailable technical knowledge. If that is the case, we would expect that the automatic translation of a large body of patent literature previously unreachable for a substantial share of the global population of inventors should have a relevant and positive effect on the extent to which inventors engage with the newly available knowledge. In particular, machine translation reduces the cost of searching for and accessing potentially useful knowledge, leveling the playing field across and within countries in terms of access to prior art.
The precise aim of this paper is to contribute to the discussion on the role of patent disclosure in fostering knowledge diffusion and to provide evidence about the importance of automated translation for crossborder and cross-language knowledge flows. The next section discusses the strategy we adopt to identify the effect of machine translations on knowledge spillovers.

Identification strategy
To investigate the impact of machine translation of patent documents on knowledge diffusion, it is necessary to identify a well-defined study population and a set of treatment conditions under which it is easy to distinguish between a treatment and control group and between pretreatment and post-treatment time periods. Ideally, we would run an experiment on a sample of patents searchable on a publicly available patent search website that are written in a specific language, for which no translation, automated or otherwise, is available. We would then randomly split the group into a treated and a control set and make available a machine translation for the patents in the treatment group. We would then assess whether the patents in the treatment group lead to an increase in knowledge transfer compared to patents in the control group after the treatment. Even though it would be hard to implement such an experiment, in this paper, we exploit a natural experiment that could be considered a close substitute of the ideal setting described above. In 2013, Google launched a major improvement to Google Patents, Google's search engine that allows users to search and read the full text of patents issued by and filed with several different patent authorities. 10 In September 2013, Google patents added the text of the patent documents from the China National Intellectual Property Administration (CNIPA), the Chinese patent office, using Google's neural machine translation service Google Translate to automatically translate patents publicly available only in Chinese through an online platform previously made available only by the Chinese patent office (Orwant, 2013). 11 In the same update, Google Patents also added the full text of patent documents from four other patent authorities: the German Patent and Trademark Office (DPMA), the Canadian Intellectual Property Office (CIPO), and the World Intellectual Property Organization (WIPO). Since September 17, 2013, the text of all Chinese patents has been available on Google Patents in both its original language and in English. The observational data from this natural experiment allows us to construct an ideal treatment group: patents published in China before September 17, 2013.
Choosing patented inventions realized in China as a treatment group is highly interesting for multiple reasons that go beyond their publication on the Google Patent platform. In global comparison, China's patent activity experienced unprecedented growth both in terms of quantity -CNIPA received the highest number of patent applications in recent years (WIPO, 2019; WIPO, 2020a, 2020b) -and quality. Dominguez Lacasa et al. (2019) show China's growing relevance in the innovation and intellectual property landscape in the context of technological catch-up by the BRICS economies. Their findings show that China is unique among BRICS in terms of its rapid improvements of technological intensity, fast structural change in the direction of dynamic frontier activities, and technology diversification, which is also expressed in the 7 Research shows, that conventional phrase-based machine translation is able to translate even complex languages into grammatically correct English, yet Groves and Mundt (2015) demonstrate it is still error-prone. However, recent improvements, the implementation of Google's neural machine translation, further bridges the gap between human and computer-aided translation (Junczys-Dowmunt et al., 2016). The modern approach reduces translation errors by an average of 60 % compared to a side-by-side human evaluation on a set of isolated simple sentences (Wu et al., 2016). 8 While Google Patents is the database storing the patent information, Google Translate is the backbone, providing the translation of the content. 9 The USPTO is also highlighting the importance of machine translation in the examination guideline by stating that examiners are "encouraged to use […] machine translations where possible in the early phases of examination". diversification of technological knowledge. China's rapid scientific progress is also evident in the production of scientific papers. Zhou and Leydesdorff (2006) provide evidence that its contribution to world science shows exponential growth not only in quantity but also in quality. Along with the exponential increase in scientific publications, the citation rates of Chinese publications are also increasing exponentially. Still, knowledge produced in China seems to be more relevant in certain technological areas, such as renewable energies (Trancik, 2014) and physical sciences and electronics (Zhou and Leydesdorff, 2006). Especially in the fast-growing field of data processing, China has evolved into a technological leader. In 2017, China's global share of research papers in the field of artificial intelligence vaulted to 27.68 %. In addition, the number of Chinese companies working in this domain grew to 1189, second only to the US (Li et al., 2021). The increasing centrality of China in the global knowledge economy makes Chinese patents the ideal treatment group with which to assess the impact of improved access to knowledge through automated translations. To identify a suitable control group, we searched for jurisdictions that have patenting activities comparable to China but that were not affected by the Google Patents improvement of 2013. South Korea proved to be a good candidate. Like their Chinese counterparts, Korean patents are locked behind a language barrier, but unlike those at the CNIPA, patents filed at the Korean Intellectual Property Office (KIPO) were added to the Google Patents platform only on August 30th, 2016, about three years after the inclusion of the Chinese documents. 12 Therefore, while English-speaking inventors could access the knowledge disclosed in Chinese documents through machine translations available on Google Patents, the content of Korean patents was not machinetranslated by Google for a longer period (Wetherbee, 2016). In addition, South Korea is an important global competitor in the IP domain, responsible for the most patent filings after China, the U.S., and Japan in 2019 (WIPO, 2020a, 2020b) and the most filings in the field of big data technologies and AI after China and the US.
Comparing knowledge flows generated by patents filed at the Korean and Chinese patent offices before and after the machine translation of the Chinese documents operated by Google is the corner stone of our identification strategy. However, the identification of the effect of machine translation confront us with a few additional challenges.
First, we have to ensure the comparability of the treated and control group. In particular, patents filed at CNIPA and KIPO, may have a different propensity to be extended to a foreign jurisdiction and, therefore, to be officially translated into another language in order to seek patent protection abroad. To mitigate the risk of such a relevant confounding factor, we decided to focus exclusively on granted patents filed at CNIPA (KIPO) and never extended (or applied for) to another jurisdiction, i.e., we constrain the sample to single-child patent families. This choice substantially reduces the chance of the invention being translated and publicly disclosed to the non-Chinese (Korean) speaking world through a mechanism other than Google's machine translation. Limiting the data to single-child families, we isolate a body of knowledge that is only available at the national level and that was likely to remain trapped behind the language barrier for Western inventors. To further reduce the possibility of access to this knowledge before the machine translation implemented by Google in 2013, we remove from our working sample any patent that lists an inventor with a residence address outside of China (South Korea) and any patent whose applicant is a non-Chinese (non-Korean) entity, based on the residence address of the applicant. The rationale for this choice is that non-Chinese-speaking (non-Korean-speaking) or mixed teams could already share knowledge in a different language using communication channels beyond the one offered by patent disclosure.
Second, we have to determine a way to measure knowledge flows to our treated and control patents and, more specifically, a measure of knowledge diffusion able to capture the impact of the reduced cost of access to information codified in a foreign language. Here we follow an extensive literature that uses patent citations as quantifiable trails of knowledge flows (e.g., Belenzon, 2012;Galasso and Schankerman, 2014;Moser et al., 2018). As explained by Hall et al. (2005), if Patent B cites Patent A, it implies that Patent A represents a piece of previously existing knowledge upon which Patent B builds and over which Patent B cannot have a claim. Following the works in this strand, we use the count of forward citations as a measure of knowledge diffusion. To better capture the effect of machine translation of documents written in Chinese, we count the number of citations to the patents in our treated and control group arriving from patents granted at the USPTO and filed by US-based inventors.
A few important works, however, discuss fundamental limitations entailed by the use of citations as a proxy for knowledge flows due to their heterogeneous and strategic nature (Alcácer et al., 2008;Sampat, 2010;Lampe, 2012;Cotropia et al., 2013;Kuhn et al., 2020). In particular, Kuhn et al. (2020) find that self-citations, examiner citations, citations made by patents citing more than 20 references, and citations added after the original applications have been filed are the ones less likely to be associated with any transfer of knowledge between applicants. To limit the noise these limitations introduce in our measure of knowledge diffusion, we follow the recommendations provided in recent contributions and focus exclusively on the types of citations that are more likely to capture actual knowledge transfer (Jaffe and de Rassenfosse, 2019; Corsino et al., 2019;Kuhn et al., 2020). First, we exclusively consider applicant citations. Subsection 5.3.1 provides a brief account of what happens when we use examiner-added citations as the outcome variable and show that our findings are valid exclusively for applicantadded citations, which are more likely to signal actual knowledge transfer. Second, by design, our citation counts omit self-citations. We exclusively count citations from US patents filed by US applicants to Chinese or Korean patents filed by Chinese or Korean applicants. Third, we remove citations arriving from patents with more than 250 backward citations. 13 The rationale behind this choice is precisely to mitigate the concern raised by Kuhn et al. (2019) that a small minority of patents citing hundreds of patents for strategic reasons might generate a large share of the observed patent citations. Fourth, although our data does not allow us to disentangle between citations added by the applicant in the original applications and citations added at a later stage, Cotropia et al. (2013) show that the overwhelming majority (about 83 %) of the citations made by US patents to foreign patents are submitted with the original application.
All in all, we believe that the strategy we adopt for counting citations allows us to measure knowledge flows. Therefore, we adopt a differencein-differences approach and compare the number of forward citations made by US inventors to our treated patents-patents filed at CNIPA by Chinese applicants that were never extended to another jurisdiction-before and after the introduction of machine translations by Google Patents, with the number of US citations received by the patents in our control group-patents filed by Korean applicants, which were never extended to a foreign patent authority and were translated by Google Patents only in late 2016.
We would interpret a relative increase in the treated groups after the 12 In an effort to reach global coverage, in this update Google Patents also included documents from the following authorities: the Japan Patent Office  13 We re-estimated the regression based on different thresholds, excluding patents with more than 20, 100, 400 and 500 backward citations and not removing any patents at all. The findings for all these thresholds are robust.
translation as an indicator of the positive influence of machine translation on knowledge diffusion. This would strongly suggest that US inventors used the information made available by the translated documents. The next section presents the data and the methodology we use to estimate the effect of machine translation.

Empirical model
As discussed in Section 3, we use a difference-in-differences approach to assess the effect of machine translations on knowledge diffusion by contrasting a group of Chinese patents, our treated group, and a group of Korean patents, our control group. To do so we estimate an econometric model that follows from our identification strategy: where the dependent variable y it is the number of citations received from US patents to Chinese or Korean patent i, divided into the time t before and after the introduction of the automatic translation by Google. Chinese reports whether the focal patent i is Chinese or not (South-Korean) and Translated is our treatment indicator, which takes the value 1 in the  period after the automated translation and value 0 before the translation. Our main variable of interest is Chinese × Translated: a positive and significant coefficient for this interaction term would confirm a positive effect of the machine translation of Chinese patents on knowledge diffusion in the US. δ ipc accounts for International Patent Classification (IPC) fixed effects and δ year includes publication year fixed effects of the patents in our sample.
The vector X i captures the features of the focal patent, including the number of applicants and the number of inventors (as listed on the patent document), the number of different IPC sections assigned to the patent, the number of citations received from patents published within the same country, the number of backward citations, the number of independent claims included in the patent and the number of patents by Chinese (Korean) applicants granted in the US before and after the treatment. In Subsection 4.2 we further discuss the relevance of these features. ϵ is the error term.
We estimate our baseline model using an OLS and, to account for the count nature of the citation data, a Poisson regression. It is important to note that the validity of the difference-in-difference analysis is dependent on two factors: the absence of any pre-event trends within the data and the assumption that both groups follow a similar trend before the treatment. We discuss the potential dissimilarities between the treated and the control group and the presence of possible pre-trends in Subsection 5.4.

Data preparation and descriptive statistics
Our main data source is the EPO's Worldwide Patent Database (PATSTAT, April 2020 release). PATSTAT is one of the world's most comprehensive patent databases and contains data on more than 100 million patents and patent applications from 90 patent authorities. To populate our treatment and control group, we exploit PATSTAT to identify patents granted by the Chinese and Korean patent authorities and first published between 1995 and September 17, 2013. To ensure that they were never extended, and hence potentially translated, outside their countries of origin, we limit our samples to patents filed by Chinese (Korean) residents that belong to patent families of size one: i.e., they are not claimed as priority filing by any other patent application within or outside the country of first filing. Since the family size of a patent is commonly used as an indicator of its value (Moser et al., 2018;Higham et al., 2021), in a robustness check described in Section 5.5.1 we use alternative sampling strategies to make sure our results are not driven by the lower quality of the patents in our sample. 14 Due to the different standards for protection required by the Chinese and Korean patent laws, we excluded Utility models from the sample. Based on the person's country code provided in PATSTAT, we also excluded all patents that listed among its inventors or applicants an individual or entity that reported a foreign residence address. Removing these patents further decreases the probability of the content being shared outside of the local language environment prior to the machine translation.
The final set contains 49,004 Chinese and 19,758 Korean patents. Fig. 1 displays the distribution of patents by country of filing and year of filing. Both countries follow a similar trend, with a peak between 2008 and 2010, though the number of Chinese patents grew drastically since 2005. Fig. 2 shows the distribution by IPC sections. 15 Previous contributions agree on the importance of controlling for technology classification and the focus on specific industrial sectors (Baruffaldi and Simeth, 2020;Berkes and Nencka, 2020;Furman et al., 2018). The preliminary descriptive statistics shown in Fig. 2 already indicate a significant difference. The technological sections Physics (G) and Electricity (H) account for about 60 % of all patents. China's catching up to the global frontier in computing and data processing can also be observed in our data. If we split Physics into classes, the majority of Chinese patents are classified as 'G06', which refers to 'computing' including the 'processing of information and the structure of the database.' Fig. 3 confirms that most the patents in our sample belong to a technological domain in which China supposedly leads.
We exploit the PATSTAT database to construct our dependent variable by counting forward citations arriving from US patents and divide the citation count into pre-treatment and post-treatment time periods, our treatment being the machine translation of the Chinese patents in 14 In particular, in Section 5.5.1 we consider patents that were first applied for in China (Korea) and later extended to additional jurisdictions in other countries. 15 The International Patent Classification (IPC) provides a hierarchical system of letters for the classification of patents and utility models according to the different areas of technology to which they pertain. The IPC divides technology into eight sections, classified into approx. 70,000 classes (WIPO, 2020a(WIPO, , 2020b. September 17, 2013. The pre-treatment time period includes data from January 1, 2008 to September 17, 2013, and the post-treatment time period a similar timespan from September 18, 2013 to December 31, 2017. Forward citations to the USPTO were chosen due to the stricter rules of the US patent office concerning citations. Because of the incompleteness of non-granted patent information in the original data set, only forward citations from granted patents were considered. Following the critique by Sampat (2010), we focus exclusively on forward citations made by the applicants. Arguing that citations added by others, e.g., the patent office examiner, were not relevant in tracking the diffusion of technical knowledge from the original inventor of the Chinese (Korean) invention to the US. In Section 5.3.1, however, we use examiner citations to control for the robustness of our approach. 16 By design all Chinese and Korean patents used in the analysis were cited at least once by a US patent. The main rationale for focusing on this group of patents is to avoid focusing on a sample of low-quality inventions. The literature on patenting trends in China has often raised concerns about the quality of the patents filed by local applicants, mainly due to the strong patent subsidy policy implemented by the central and local governments since 2011 (Boeing and Mueller, 2019) and, to a lesser extent, due to the potential home bias of the Chinese patent authority (de Rassenfosse and Raiteri, 2022). To avoid populating our working sample with patents of dubious interest for inventors in the US, we focus on the universe of inventions that received at least one citation from a US patent at any point in time in the period taken into account. We interpret those patents as covering inventions with at least some potential to become relevant to the US market, and, hence, with the potential to be more easily discovered through the automatic machine translation introduced by Google in 2013. 17 To limit the probability of a native speaker being involved with the US patent application process, all citations from US patents filed by Chinese (Korean) inventors and companies were excluded from the count. 18 The distinction was made based on the residence address reported in PATSTAT, leaving 128,292 citations to the patents in our full sample. However, this approach does not fully rule out the possibility that some of these citations may come from Chinese (Korean) inventors able to read Chinese (Korean) patent applications independently of the machine translation offered by Google Patents. Patent documents, and hence the PATSTAT database, do not report the nationality of the inventors but merely their registered address of residence. For example, a Chinese inventor living in the US would be labeled as US inventor following the method discussed above, yet she would still be able to fully understand the Chinese language and would not need translation to access the information disclosed in Chinese patents. Following the approach of Breschi and Lissoni (2009) and Breschi et al. (2015), to make a more fine-grained distinction, in Subsection 5.2 we made use of two different natural language processing libraries to construct two additional dependent variables that exclude citations from two types of patents. First, we exclude citations coming from patents filed by inventors that have a name that the algorithm identifies as Chinese or Korean. Second, to mitigate issues that may arise from names that are used for both Chinese and Korean individuals, we exclude from the citation count all the citations coming from patents with inventors with names the algorithm identifies as Asian.
After constructing the outcome variable, we also construct several control variables capturing characteristics of the focal patent that may potentially influence the relationship of interest. From PATSTAT we extract the number of applicants and inventors, as listed in the patent document, the number of IPC classes a patent is assigned to, the number of backward citations received, and the number of citations received from other Chinese or Korean patents.
The number of IPC classes reports the total number of four-digit IPC classes assigned to a patent. Patent documents covering many IPC classes are commonly used in literature as proxy for the technological scope of the invention (Harhoff et al., 2003).
Both the number of backward citations received and the number of citations received from other Chinese or Korean patents are used as a proxy for the patent's quality. The latter in particular provides information about the patent's significance within the Chinese or Korean language environment and therefore relates to the importance of US inventors' ability to access its content.
Additionally, we used Google Patents to create the variable number 16 In addition to using citations added by the examiner separately, we reestimate our baseline model including citations added by both the examiner and the applicant. In Subsubsection 5.3.1 we discuss the results. 17 To make sure that our sampling strategy is not driving our main results, we estimate our model on four alternative samples. Appendix A provides the details of this analysis. First, we consider also patents that never received a citation by a US patent and we find a positive and statistically significant effect of machine translation, although smaller in magnitude compared to the baseline results. Given the high share of patents that never received at least one citation from a US patent (about 90 % of the 763.292 patents) we then also reestimated the effects when we remove patents from IPC subclasses that never received a citation by a US patent, patents that never received an international citation, or by patents that never received a citation by patents filed at the EPO, JPO or USPTO. All estimates confirms the validity of the results of the focal analysis.
18 Re-estimating our baseline model without removing individuals or entities reporting a foreign residence address still shows a positive and statistically significant effect. Compared to our baseline result it yields a much smaller marginal effect, implying an increase of 5.2 % in forward citations after the translation.
of claims, which reports the number of independent claims listed in the patent application. Independent claims describe the essential features of the invention and this variable could then be interpreted as a proxy for the scope of the invention.
To account for a general rise in Chinese (Korean) innovation because of increasing international interest, we include a control variable, reporting the number of granted US patents filed by Chinese (Korean) applicants both before and after the treatment (henceforth referred to as No. of local applicants). Table 1 shows the summary statistics of the key variables used in our analysis for the full sample, divided by the treated and control focal patents. On average, the treated patents appear to receive a smaller number of forward citations. Moreover, Chinese patents make on average fewer backward citations and have a significantly smaller number of claims. Table 2 presents the results of the baseline OLS estimates for the main dependent variable. As previously mentioned, to better identify the effect of knowledge diffusion, we restrict the analysis exclusively to the sample of treated focal patents and their corresponding control patents that were cited at least once by a US patent filed between 2008 and 2018. The results, reported in column 1, show a positive and highly statistically significant coefficient for the interaction term Chinese × Translated, confirming a positive relationship between machine translation and knowledge diffusion. The OLS results suggest that Chinese patents receive on average about 0.186 more citations than Korean patents for which no automated translation existed at the time on Google Patents. Compared to an average of 1.93 citations received by the patents in our sample, this implies a 9.7 % increase in forward citations after the treatment period.

Baseline result
To account for the fact that citations are count data, we also estimate the baseline model using a Poisson regression. Table 2 reports the estimates in column 2, which yields a slightly smaller marginal effect than OLS. The Poisson regression estimates an increase of 0.138, implying an increase of 7.2 %.
While the baseline results show a positive and statistically significant effect, the effect of translation on knowledge diffusion becomes more evident when we split the data by technological sections. The Poisson estimates reported in column 1 of Table 3 point to a strong effect, especially in the Chemistry, Physics, and Electricity classes. With an increase of 14.5 % in forward citations after the translation, patents from the Physics domain seem to drive the results. As previously discussed, Physics is the IPC section under which technologies related to computing are categorized. Information technologyand computer science in particularis one domain in which China has rapidly evolved to become a global leader in the past ten to fifteen years (O'Meara, 2019; Li et al., 2021). The fact that the benefits of machine translation are especially evident in fields in which China is closer to the technological frontier seems to suggest that improved access to knowledge allows US inventors to identify and use the relevant technological contributions made by Chinese inventors. We also find a substantial positive effect in the Chemistry domain but, quite surprisingly, a strong negative effect in the Human Necessity category. The latter effect appears to be driven by two technological subclasses that cover surgical and diagnostics instruments, and by citations originating from large entities but we cannot provide a conclusive explanation for this result.

Citation origins
As discussed in Subsection 4.2, one major limitation of the data that might potentially threaten our interpretation of our baseline estimates lies in the way patent documents report the country of origin of inventors and applicants: PATSTAT reports their registered residence address and not their actual nationality. For this reason, a Chinese Table 2 The effect of machine translation on forward citations, OLS and Poisson estimates.  Notes: Significant at ***1 %, **5 % and 10 %*. Robust Standard errors in parentheses. The dependent variable records the cumulative forward citations that the focal patent i has received from U.S. patents earliest filed between 2008 and 2019. Not-granted patent citations and foreign (not-U.S.) citations are excluded from the count of the dependent variable. The variable Chinese x Translated is a binary indicator that identifies those focal patents that was treatedadded to Google Patent database and automatically translated in 2013. To address the skewness of citation data, that a small minority of patent applications is generating a large majority of patent citations (Kuhn et al., 2019), citations arriving from patents with more than 250 backward citations were removed.

Table 3
The effect of machine translation on forward citations divided into technological areas -Poisson estimates. inventor who lives in the United States would simply be registered as US resident in PATSTAT. But a Chinese inventor living in the US obviously does not face the same language barrier as her non-Chinese-speaking colleagues when it comes to searching and reading the Chinese-only patent literature. In such a scenario, we cannot entirely rule out the hypothesis that the effect we recover in the baseline estimates could be driven by confounding factors, such as a sudden increase in high-skilled Chinese migration happening contemporaneously to our treatment. Indeed, the number of high-skilled workers from China is steadily increasing; since 2013, they have represented the second-largest group receiving H1-B visas in the US (USCIS, 2020).
To rule out the concern that this migratory phenomenon is driving our result, we use two natural-language-processing algorithms to infer the nationality of the inventors based on their given names and surnames. First, we use the library name2nat (Park, 2020), which was trained on names and nationalities extracted from Wikipedia and assign a specific nationality to a combination of first name and surname. Second, we use Ethnicolr, a machine-learning-based classifier trained on a specific dataset and implemented in Python to infer the ethnic background (Laohaprapanon and Sood, 2021). This algorithm assigns individuals based on their first and last names to categories that combine ethnic backgrounds. The algorithm was trained on Florida voter registration data and Wikipedia data from 2000 and 2010. It assigns to a given name a specific probability of belonging to one of four classes: 'White', 'Black', 'Asian' or 'Hispanic', and then assign the imputed ethnic background by selecting the one associated with the highest probability.
Using the name2nat library on the names of the US patent inventors citing our focal patents, we identified patents that listed inventors considered as Chinese or Korean according to the algorithm. We then constructed an alternative outcome variable for which we remove these patents from the citation count. Using this approach, we excluded 43,356 citations from the original count, leaving 84,936 citations from US patents without any Chinese or Korean inventors involved.
We instead use the Ethnicolr classifier to create a citation count that excluded any patents listing inventors with a combination of given name and surname that the algorithm identifies as Asian. We remove patents with inventors that have a 50 % or higher probability of being Asianwhich results in the removal of 79,235 citations. 19 We then re-estimate the baseline model using the newly created dependent variables. Table 4 reports the marginal effects recovered through the Poisson regression, together with the OLS coefficient as a benchmark.
All results show a positive and statistically significant effect, that is overall stronger than the baseline effect. Compared with the average of 1.28 citations received by the focal patents, the Poisson estimate implies (see column 2) a 9.5 % increase in forward citations from patents without any inventor with a Chinese or Korean name involved, after the translation. If we consider citations arriving from patents without any inventors with an Asian name, the estimator yields a stronger marginal effect, implying a 13.1 % increase of forward citations.
The overall difference with the baseline estimate appears to be rather small, but if we investigate the different technological areas, in some areas we can observe substantial changes. Table 5 reports the results of the Poisson regression of the dependent variable without Chinese and Korean inventors involved in column 1 and without any Asian inventors in column 4.
While the tendencies of the results are similar to the baseline estimates, the effect is once again mainly driven by fields in which China became is on the technological frontier; the effect in the Physics domain almost doubles after the removal of patent citations from Chinese and Korean inventors. Based on an average of 1.253 citations per patent, Chinese patents related to Physics show a 24 % increase in forward citations after the treatment period when compared with the control group. Column 4, the citation count without Asian inventors involved, confirms these results, indicating a 25 % increase in the number of citations after the implementation of machine translation by Google.
These estimates indicate that our initial assumption discussed in the identification strategy in Section 3 is highly relevant. The effect is not driven by citations coming from patents with inventors already capable of speaking the language but from inventors who likely got access through translation. 20 As in the baseline, this is especially evident in areas in which China has become a technological leader, such as computing and data processing.

Small entities
Although Google Translate is arguably the most popular machine translation service, it is fair to assume that specific kind of organizations such as large multinational enterprises may have in-house professional translation and machine translation services already in place and they might use these services also to parse and examine the patent literature relevant to them. Clearly, we should expect the benefits of a generalized machine translation service to be quite limited in such cases. Still, these professional translation services or ad hoc machine translation software programs are pricey, and their cost can be an obstacle especially for small entities or independent inventors. Being open access and free of charge, Google Patents sets itself apart from paid solutions, therefore lowering the barriers to its use and enabling small entities to make use of it. For this reason, we would expect a stronger effect of the treatment for Table 4 The effect of machine translation on citations arriving from patents without Chinese and Korean inventors and without Asian inventors involved. OLS and Poisson estimates.  citations arriving from patents that were filed by smaller entities and single inventors. Since there is a high chance of large companies' already having internal machine or professional translation services in place, we constructed an additional alternative outcome variable that exclusively counts citations arriving from US patents filed by small entities. To identify small entities, we follow the definition used by the USPTO (section 3 of the United States Small Business Act) to grant a 50 % reduction in application fees. This definition includes independent inventors, small businesses, and nonprofit organizations. Using the USP-TO's Patent Examination Research Dataset (PatEx) (Graham et al., 2015), we identify the citations arriving from small entities and create the new outcome variable. 21 In addition, by exploiting the information available through the Patstat database, we construct another version of our outcome variable that only considers citations arriving from patents filed by individual inventors. Table 6 reports the results for citations arriving from patents without any Asian inventors involved. 22 All OLS and Poisson estimates are positive and statistically significant. The non-linear estimator yields a slightly smaller marginal effect than OLS, yet the magnitude of the effect is much stronger than the baseline results if we consider only citations from small entities. Based on an average of 0.135, the model estimates an increase of 16.3 % (34.1 % for the OLS) in citations after the treatment. Though the effect for citations arriving from patents filed by individual inventors is smaller in magnitude than the citations arriving from small entities, implying an increase of 7.9 %, it is still slightly stronger than the baseline results. These results confirms that the benefits of improved access to knowledge behind a language barrier are especially important for those categoriesmainly small businesses and nonprofit organizations but also individual inventorsfor which the costs of access to this knowledge through alternative means are higher.

Examiner citations
As suggested in Section 3, using patent citations as an indicator for knowledge flows has limitations that must be addressed. One important aspect to consider is whether citations are added by the applicant or by the patent examiner who reviews the patent application. Previous Table 5 The effect of machine translation on citations arriving from patents without Chinese and Korean inventors and without Asian inventors involved divided into technological areas -Poisson estimates. CN Table 2, see text for details. The dependent variable record the cumulative forward citations that the focal patent i has received from U.S. patents earliest filed between 2008 and 2019 and with Chinese and Korean inventors removed (columns 1-3) and Asian inventors removed (columns 4-6). Not-granted patent citations and foreign (not-U.S.) citations are excluded from the count of the dependent variable. The variable Chinese x Translated is a binary indicator that identifies those focal patents that was treatedadded to Google Patent database and automatically translated in 2013. Notes: Significant at ***1 %, **5 % and 10 %*. Robust Standard errors in parentheses. This table shows only the regressor of interest. Regressors not listed are those of Table 2, see text for details. The dependent variable in columns 1 and 2 records the cumulative forward citations that the focal patent i has received from U.S. patents earliest filed between 2008 and 2019 and filed by small entities, columns 3 and 4 report the estimates for citations from patents filed by individual inventors and columns 5 and 6 reports estimates for citations from patents filed by small entities without individual inventors. Citations from patents with Asian inventors involved were removed from all samples. To identify small entities we follow the definition used by the USPTO (section 3 of the United States Small Business Act) to grant a 50 % reduction in application fees. To identify individual inventors, we used the invt seq nr within Patstat. In addition, not-granted patent citations and foreign (not-U.S.) citations are excluded from the count of the dependent variable. The variable Chinese x Translated is a binary indicator that identifies those focal patents that was treatedadded to Google Patent database and automatically translated in 2013. 21 To flag small entities, we used the USPTO Patent Examination Research Dataset (PatEx). The PatEx dataset provides access to the bulk data collected by the Public Patent Application Information Retrieval system (Public PAIR) and is available at https://www.uspto.gov/learning-and-resources/electronic-dataproducts/patent-examination-research-dataset-public-pair. 22 Re-estimating the results based on a less narrowly defined outcome variable not removing citations from patents filed by Asian inventors or removing only patents with Chinese/Korean involvedshows a similar effect, but weaker in magnitude.
literature acknowledges that examiner citations add measurement error and that not reporting them separately adds unknown noise to the data (Alcácer and Gittelman, 2006). In our identification strategy, we consider only citations added by the applicants, even though patent examinersgovernment agents who approve patent applicationsare also involved in drafting the content of patents. 23 On the other hand, their citations are unlikely to reflect knowledge flows since they are not involved in the innovation and innovation process itself (Jaffe and Trajtenberg, 1999). Still, Alcácer et al. (2008) show that examiner citations account for 63 % of all citations in the average USPTO patent, making them a relevant factor to consider when testing our model's robustness.
Besides not being involved with the initial knowledge flow, patent offices and examiners already had translation services in place before Google's patent translation. They could already search and access the Chinese patent literature before 2013. Hence, citations added by examiners should not be affected by the machine translation services offered by Google. To verify this assumption, we recompute our dependent variable by exclusively counting citations coming from patent examiners. 24 Table 7 reports the results of the OLS and the Poisson regression. No associated coefficients differ significantly from zero, showing no change in forward citations after the translation in 2013. This result appears to confirm that the treatment effect in the focal analysis is not driven by confounding factors that would also affect examiner behavior, but by the inventors actually benefiting from improved access to the Chinese patent literature.

Robustness checks
As discussed in Section 3, there is a general threat to the validity of our approach which mainly relates to the comparability of the treated and the control group. As the descriptive statistics in Section 4.2 show, Chinese and Korean patents do present some differences in their observable characteristics raising the concern that these differences could be correlated to unobservable time-variant factors that are not adequately addressed by our difference-in-differences approach.
To rule out these concerns, we perform two additional robustness checks that confirm the validity of the results presented in the previous sections.

A conditional Difference-in-Difference approach
To better account for potential differences between the treated and the control group, we show that the results are robust to the implementation of an exact matching approach combined with the differencein-differences method used in the focal analysis, i.e., to the adoption of a conditional Difference-in-Difference approach. We perform an exact matching on the IPC class 25 and the application year of our focal patents. Therefore, this procedure creates bins containing treated and control inventions that belong to the same technology field and were developed around the same time. Considering the imbalance of our core set, having three times as much Chinese patents as Korean patents, we accounted for the robustness of our baseline results by creating two different matching sets. In the first one, each treated patent is matched to one perfect twin in the control group, whereas in the second, we match using a two-toone ratio, keeping up to two Chinese patents in each bin. We removed patents without a suitable twin in both cases.
The one-to-one exact matching removes roughly half of the patents, resulting in a loss of 33,864 Chinese and 4610 Korean patents. The overall sample size of the matched set is 30,288. Exact matching with a ratio of two-to-one gives a sample size of 41,142 patents, removing 24,266 treated patents.
We re-estimate the baseline model from patents without Chinese and Korean inventors involved using the Poisson and OLS estimators based on the two matched samples. Table 8 reports the estimates together with the previously-estimated marginal effects for the full sample as a benchmark. The Poisson regression yields smaller marginal effects than the OLS, yet overall all effects are statistically significant and positive. If we consider citations to the one-to-one matched sample, the OLS estimate implies a 17.1 % and the Poisson a 5.2 % increase. The estimators for the two-to-one matched sample yield similar positive and significant effects: an increase of 18 % (OLS estimate) and 7.2 % (Poisson estimate) in the citations arriving from US patents after the machine translation.
Comparing the results for the matched samples to the one obtained in the focal analysis, the matched samples show a slightly weaker effect but, overall, the results are qualitatively very similar to the one obtained for the full sample, confirming the validity of the findings discussed above.

Pre-event trends
Another potential threat to our identification strategy comes from the possible presence of pre-trends driving our results. To mitigate this concern, we perform a placebo test. To do so, we construct the data in the same way as introduced in Subsection 4.2, but pretend that the machine translation took place in 2011, two years prior to the real treatment. In addition, we remove patents published after September 17, 2011, and adjusted the time period for the collection of citations coming from US patents accordingly. Table 9 presents the estimates of the fake shock for the core set in columns 1 and 2 and, additionally, only for Physics-related patents in columns 3 and 4. 26 The OLS estimator for the core set is not significantly different from zero (column 1, row 1). The Poisson estimate shows a weak statistically Table 7 The effect of machine translation on citations added by examiners: counting citations from U.S. patents without Chinese and Korean inventors based on core and matched samples. Notes: Significant at ***1 %, **5 % and 10 %*. Robust Standard errors in parentheses. This table shows only the regressor of interest. Regressors not listed are those of Table 2, see text for details. The dependent variable records the cumulative forward citations added by the examiner that the focal patent i has received from U.S. patents earliest filed between 2008 and 2019 (column 1), the 1:1 matched sample (column 2) and the 2:1 matched sample (column 3). Notgranted patent citations and foreign (not-U.S.) citations are excluded from the count of the dependent variable. The variable Chinese x Translated is a binary indicator that identifies those focal patents that was treatedadded to Google Patent database and automatically translated in 2013. 23 Re-estimating our baseline model without excluding citations added by the examiners, still shows a highly significant and positive effect, implying a 5.5 % increase in citations after the translation. Compared to our baseline results, the increase is smaller if examiners citations are included in the sample. 24 An interesting observation is that the examiner citations in our data only account for roughly 4.2 % of overall citations. 25 There are in total 573 unique IPC classes in the data set. 26 The results for the Physics class is reported because, as discussed in Section 5, citations to the patents belonging to this class appear to be driving the results.
significant negative effect. The treatment group was cited less after the fictitious treatment compared to the control patents.
Re-estimating the model using the previously introduced matched samples, the one-to-one (column 1, row 1 and 2) matched and the twoto-one matched sample (column 2, row 1 and 2), shows no effect of the placebo treatment, both the Poisson and OLS coefficient for the interaction term are not significantly different from zero.
The test confirms that there is a weak pre-event trend when we consider the full sample. Yet, it is a weak negative trend, therefore, if anything, it should have led to a reduction in the magnitude of the actual effect and does not seem to affect the validity of our results. In addition, the effect disappears when we increase the similarity between the treated and the control group by matching on the application year and the technology class.

Alternative sampling
As discussed in Section 4, to ensure that the knowledge disclosed by our focal patents was not already translated in international applications simultaneously filed at other patent authorities, we decided to focus on patents never extended beyond the Chinese (Korean) jurisdiction. The size of a patent family, however, is a well-known indicator for the value of a patent document (Moser et al., 2018;Higham et al., 2021), and we might be concerned that limiting our sample to patents filed at only one authority could introduce a potential bias towards lower value inventions.
To ensure that our results are not driven by low quality inventions we run an additional analysis in which we estimate our baseline model with two main modifications.
First, we use three alternative sampling strategies. More specifically, we identify patents filed in China (Korea) and extended to any national patent offices other than China or Korea no later than September 2013. In this way, we select our first alternative sample, composed of 23.754 patents. Since the majority of these 23.754 patents were first extended to either the USPTO or the Japanese Patent Office (JPO), we create two additional samples: patents first extended to the USPTO and patents first extended to the JPO.
The second difference concerns the dependent variable and illustrates a major limitation from including patents with a family size larger than one. In fact, a patent having additional family members in other jurisdictions, for instance a US equivalent to a Chinese or Korean inventions, would introduce some noise in the measurement of our outcome variable. Even if a US inventor learned about a new patented invention via the machine translation of the Chinese document done by Google Patents, she might then become aware of the existence of the US patent application and cite the US document instead of the original Chinese document. To make sure we capture the potential effect of translation in such cases, we need to count forward citations at the family level, and not just to the Chinese family member, introducing additional noise in our measure (Criscuolo, 2006;Bakker et al., 2014). Therefore, for this analysis, we count citations from US patents to any of the patents in the family of the focal patent, i.e., to any DOCDB family member. Table 10 presents the Poisson estimates based on the alternative sampling. All results show a positive and statistically significant effect, with differences in magnitude. Compared to the average, the Poisson estimate for patents first extended to the US (column 1) implies a 3.9 % increase in forward citations. When we consider citations to patents extended to Japan (column 2) or to any other patent authority (column 3), the effect is stronger in magnitude, implying a 7.7 % and 9.6 % increase in forward citations.
These estimates suggest that machine translation seems to have a positive impact also for patents considered to be of higher value or of higher international importance.
Interestingly, the results appear to be weaker in magnitude when we consider patents extended to jurisdictions requiring an official translation into the English language of the patent application, such as the USPTO. Clearly, in such cases, the need for a machine translation to access the information included in the Chinese document is reduced by the publication of an official translation when the USPTO patent application becomes public. The results are instead in line with our baseline results when we consider patents first extended to the Japanese patent office or any other jurisdiction.

Patent value based on local citations
In Section 5.5.1, we demonstrated that our results are robust to the inclusion of patents that belong to larger patent families and are, in principle, more valuable. As discussed above, such exercise involved, however, the use of alternative samples and a different method to count forward citations. We now turn our attention to the relationship Notes: Significant at ***1 %, **5 % and 10 %*. Robust Standard errors in parentheses. This table shows only the regressor of interest. Regressors not listed are those of Table 2, see text for details. The dependent variable records the cumulative forward citations that the matched focal patent i has received from U.S. patents earliest filed between 2008 and 2019 The matched samples were also re-estimated with citations arriving from non-Asian patents, which yielded statistically significant and positive effects, but not reported in the table. In addition, not-granted patent citations and foreign (not-U.S.) citations are excluded from the count of the dependent variable. The variable Chinese x Translated is a binary indicator that identifies those focal patents that was treatedadded to Google Patent database and automatically translated in 2013.

Table 9
Effect of a robustness test to investigate possible pre-event trends for citations without Chinese and Korean inventors on the core and matched sets -OLS and Poisson estimates. Notes: Significant at ***1 %, **5 % and 10 %*. Robust Standard errors in parentheses. This table shows only the regressor of interest. Regressors not listed are those of Table 2, see text for details. To investigate possible pre-event trends, we removed patents published after 2011. The dependent variable records the cumulative forward citations that the focal patent i has received from U.S. patents earliest filed between 2006 and 2017 (column 1), on the 1:1 matched sample (column 2) and 2:1 matched sample (column 3). Not-granted patent citations and foreign (not-U.S.) citations are excluded from the count of the dependent variable. The variable Chinese x Translated is a binary indicator that identifies those focal patents that received the fake treatment.
between machine translation and the value of our focal patents. In fact, we might have the concern that machine translation, by lowering the costs of prior art searches, simply increases the probability of finding marginal patents: knowledge of little relevance for US inventors, that would otherwise go unnoticed. To test for this possibility, we use forward citations received from local patents as a proxy for the value of our focal inventions and split the sample into four quartiles based on the distribution of local citations. We then split the sample by the four quartiles and estimate our baseline model for each group separately. Fig. 4 graphically depicts the marginal effect of machine translation for each quartile. As Fig. 4 shows, the effect of machine translation is substantially larger for patents in the top quartile of the local citation distribution. Machine translation leads to an increase of 32.3 % for the top quartile. Although positive and significant, the effect is an order of magnitude weaker for the central quartiles (3.1 % increase). Interestingly, the effect is quite strong for the patents in the bottom quartile of the value distribution. Machine translation increases citations from US patents to the least locally cited patents by 26.3 %. 27 All in all, this test confirms that our results are not entirely driven by marginal patents. In fact, machine translation appears especially valuable in fostering the diffusion of knowledge embedded in more relevant, valuable inventions. The result, however, also indicates that reducing costs of search substantially increases the chance of discovery of marginal patents that would not receive attention otherwise.

Do inventors read machine translations?
In Section 5.2, we discussed a few factors that may have confounded our results and rule them out by using alternative methodologies to construct our outcome variables. However, these additional analyses only provide indirect evidence that the main channel through which machine translations of Chinese patents by Google patents lead to an increase in citations by inventors located in the US is an improvement in the accessibility to knowledge that was previously available only in Chinese. This section aims to provide more direct evidence of a direct link between machine translations and knowledge flows. To do so, we run two additional analyses in which we adopt a triple-differences approach.

University-owned patents
In a recent article Kong et al. (2020) investigate the difference between university and corporate patents. They use linguistic measures to show that university patents need 1.1 to 1.6 years of education less to read. Considering corporate patents, the gap is 2.2 to 2.6 times larger between the top 100 applicants, which further supports the hypothesis that this difference may stem from a strategic motive whereby corporations intentionally obscure their inventions to deter competitors from adopting the innovation. The relevance for machine translation is evident. Considering the finding of Larroyed (2018)-patents with easier sentence structure result in higher quality translations-we would expect a stronger effect of machine translation on patents owned by universities that have a higher readability than corporate patents. Clearly, if patents with a higher readability received more citations as result of machine translation in comparison to patents with a lower readability, this would strongly suggests that US-based inventors actually read those patents and use the knowledge codified in the patent documents. A first descriptive analysis of our data seems to confirm this hypothesis. In total 22,843 patents are owned by universities. Five out of ten of the treated patents with the highest increase in citations after the treatment are University-owned patents, whereas in the control group no university-owned patents appear in the top ten.
To investigate the role of university patents, we expand our initial   27 We additionally split the sample into four quartiles based on the distribution of forward citations received from international patents (excluding citations from patents filed at the USPTO, CNIPA and KIPO) and reestimated our baseline model. The results follow a similar pattern, showing that the effect is mainly driven by the top and bottom quartile.
model and use a difference-in-difference-in-differences estimator, which distinguishes between patents filed by universities and corporations. We estimate the following model: The baseline model discussed in Subsection 4.1 is extended by the variable University, which reports whether the focal patent i is filed by a University. 28 In the triple-differences model, the main variable of interest is the variable Chinese × Translated × University. This is a binary variable taking value 1 if the patent is Chinese, a University patent, and translated after 2013. Therefore, a positive and significant coefficient will provide indications on the effect of machine translation of Chinese university patents on knowledge diffusion in the US. Table 11 reports the marginal effects of the Poisson estimators for citations arriving from patents without Chinese and Korean inventors involved (Columns 1), and citations without Asians (Column 2).
The analyses shows a highly significant and positive effect of the triple-interaction term for the dependent variables. These results imply that more readable patents get more frequently cited by US-based inventors as a consequence of the machine translation. This, in turn, appears to confirm that US inventors learn new technical knowledge by reading machine translated patents.

Illustrations in patents
In the previous section, we show that university-owned patents likely to be more readable than corporate patents are cited more often after the machine translation. We interpret this finding as evidence that US-based inventors read and learn from Chinese patents once the cost of accessing knowledge is lowered. Another characteristic that can make a patent more easily understandable and readable is the presence of illustrations in its claims. The availability of a graphical figure in the patent claims could straighten potential errors in the interpretation deriving from an inaccurate or flawed machine translation. In addition, illustrations increase the overall readability of patent claims. If this is the case, we would expect patents with illustrations in their claims receive more citations on average after the introduction of the treatment in 2013.
In order to test this hypothesis, we estimate the following new tripledifference model: The variable Figure reports whether the focal patent i contains a figure in its claims. 29 Our main variable of interest is Chinese × Translated × Figure, which takes value 1 for Chinese patents with at least one figure in their claims after the translation. In total 3714 patents in the data set have at least one illustration in their claims.
The coefficients of the variable of interest reported in Table 12 are always positive and highly significant in all models. The Poisson regression estimates that a Chinese patent with a figure in its claims, once translated, receives on average about 0.601 citations more. Compared to an average of 1.25 citations, this implies a 48.3 % increase in forward citation after the treatment period. The effect is even stronger if we re-estimate the model for the dependent variable counting citations from patents without Asian inventors involved. The result of the Poisson estimates indicates a 56.8 % increase.
To our knowledge, no research has yet investigated the relationship of translation, knowledge diffusion, and graphical figures in patent documents. Our results strongly suggest the existence of such a relationship. Patents with pictures in their claims receive more citations after being translated, implying an easier access to this knowledge once it is translated into English.

Discussion and conclusion
One of the main rationales for the existence of the patent system is the promotion of knowledge diffusion through the disclosure of technical knowledge. The scholarly debate on the actual impact of patent disclosure on knowledge diffusion, however, remains ongoing. Our paper seeks to contribute to the discussion by investigating the impact of machine translation.
Our identification strategy exploits an interesting natural experiment, the automatic translation of patent documents issued by the Chinese Patent Office implemented by the Google Patents service in 2013. This improvement in the patent search service allowed inventors to search and read in English patents previously available exclusively in Chinese and, hence, lowered the cost of access to the prior art for English speakers.
Our results suggest that translation facilitates knowledge diffusion. We use citations from US patents to 49,217 prior Chinese patents that were never filed, and hence translated, outside of China. We use a group of 19,860 Korean patents as a control group, exploiting the fact that patents issued by the Korean patent office were included and automatically translated in English by Google Patents three years after the 28 Patents classified as University patents include all patents filed by educational and academic entities. To identify such patents, we use the PATSTAT reported applicant name. Every patent containing the following keywords is classified as a University patent: 'university', 'school', 'academy', 'college', and 'institute'. The Chinese and Korean translation of these terms is used accordingly. 29 The data was obtained from Google Patents. We only considered figures in the claims, to ensure the relevance of the figure for the innovation itself.
patents issued by the Chinese patent office. We find a positive and statistically significant relationship between the machine translation and the number of forward citations. Our estimates indicate that treated patents receive on average 7.2 % more forward citations than the control group after translation. The effect is particularly strong in technical domains in which China is a technological leader, such as computingand data-processing-related technologies, and for resource-constrained entities, such as non-profit organization, small enterprises, and independent inventors. The main results are robust to a broad range of alternative specifications, like assuming a fake treatment period to account for potential pre-event trends, and to alternative sampling strategies. We find no effect for examiner-added citations, providing further evidence supporting the disclosure argument. The effect is robust to exclusively considering citations arriving from non-Chinese or non-Asian inventors, as identified by natural language processing algorithms, indicating that the effect appears to be driven by inventors incapable of understanding the content of the patent prior to the translation.
In addition, we investigate the heterogeneous effect of translation for university-owned and corporate-owned patents. University patents, due to their easier readability, have a higher likelihood of resulting in an accurate translation. Our findings support this claim and suggest that US-based inventors read and use the knowledge disclosed in Chinese documents. We reach a similar conclusion by focusing on patents that include illustrations in their claims.
These findings have several policy implications. One, immediate consideration is that the patent system produces the intended effects at least when it comes to the disclosure function. Lowering the costs of access to the knowledge disclosed in patent documents leads to an increase in the use of that knowledge.
Second, our findings show that one of the main channels through which the benefits of patent disclosure are realized is the translation of knowledge that would otherwise remain trapped behind a language barrier. The existence of an international and harmonized patent system makes relevant knowledge available to a wide and interested audience and facilitates knowledge flows between geographically and culturally distant areas of the world.
Third, our results also show that there are still substantial gains that could be obtained in terms of knowledge dissemination by improving the quality and availability of the knowledge disclosed in patents. As shown by the differential impact of translation between university and corporate patents, it is clear that the usual critique made by scholars skeptical about disclosure theory is not far-fetched: corporate patents do often hide, rather than disclose, relevant information. The adoption of policies to enhance the enforcement of the disclosure requirements during the patent prosecution process could be a step toward enjoying the full benefits of patent disclosure.

Declaration of competing interest
None of the authors have any conflicts of interest or financial ties to disclose.

Appendix A. Including focal patents without US backward citations
To confirm that the effect of machine translation is robust to alternative sampling methodologies, we include a robustness check in which we run our focal analysis on four different samples. In the first sample, we include single-child patents, applied by local applicants at the CNIPA or KIPO, without imposing further restrictions. In this way, the construct a sample of 763,292 patents. About 91 % of these patents never received at least one citation from a US patent document (neither before nor after the treatment). For our next sample, we start selecting inventions that build on knowledge potentially more relevant to the US market, i.e. at a higher risk of receiving a citation from US inventors. We do so by eliminating all the patents belonging to IPC classes (at the subgroup level), that never received at least one citation from a US patent in the period taken into account, therefore technological areas that we interpret as of little relevance to inventors in the US. In this way, we remain with a sample of 553,648 patents that belong to IPC subgroups cited at least once by US inventors. For our third sample, we select only patents that received at least one citation from an office other than the CNIPA or KIPO, in the period taken into account, which created a sample of 698,904 patents. Finally, we come closer to the sample used in the focal analysis and select only patents that received at least one citation either from a USPTO, JPO, or EPO patent. In this way, we identify a sample of 164,346 patents that received attention from documents issued by at least one of the three most important patent offices in the period taken into account and therefore could be considered as disclosing knowledge that has a stronger international appeal. Patents from the last sample received on average 1.027 citations from US inventors in the period taken into account.
As Table 13 shows, the impact of automatic translations is always positive and significant and appears to be stable, in relative terms, across all samples, ranging from a 6 % increase in the first and fourth sample to a 16 % increase in the second sample. All in all, these results confirm the Notes: Significant at ***1 %, **5 % and 10 %*. Robust Standard errors in parentheses. This table shows only the regressor of interest. Regressors not listed are those of Table 2, see text for details. The dependent variable records the cumulative forward citations that the focal patent i has received from U.S. patents earliest filed between 2008 and 2019 and without Chinese/Korean inventors involved (column 1) and without Asian inventors (column 2). In addition, not-granted patent citations and foreign (not-U.S.) citations are excluded from the count of the dependent variable. The dummy coded variable 'Figures' indicates, that a patent has at least one figure in the claimbased on the data retrieved from Google Patents. In total 3714 patents in the data set have at least one illustration in their claims. The variable Chinese × Translated × Figures is a binary indicator that identifies those focal patents that were treatedcontaining at least one figure in the claims, added to Google Patent database and automatically translated in 2013.
idea that our main findings are not driven by the specific sampling strategy we adopted.

Table 13
The effect of machine translation on forward citations, including Chinese (Korean) patents never cited. Notes: Significant at ***1 %, **5 % and 10 %*. Robust Standard errors in parentheses. This table shows only the regressor of interest. Regressors not listed are those of Table 2, see text for details. The table only reports Poisson estimates, the OLS results are similar in magnitude and significance. The sample was extracted following the criteria described in Subsection 4.2, but without removing Chinese (Korean) patents that never received a citation by a US patent, resulting in 763.292 Chinese and Korean patents. The table reports estimates on the whole sample (column 1), on patents filed in IPC subsections which at least received one citation (column 2), on patents that received at least one citation from an office other than CNIPA or KIPO (column 3) and only considering patents that were cited by patents filed at the EPO, JPO or USPTO (column 4). The dependent variable records the cumulative forward citations that the focal patent i has received from U.S. patents earliest filed between 2008 and 2019. Not-granted patent citations and foreign (not-U.S.) citations are excluded from the count of the dependent variable. The variable Chinese x Translated is a binary indicator that identifies those focal patents that was treatedadded to Google Patent database and automatically translated in 2013. To address the skewness of citation data, that a small minority of patent applications is generating a large majority of patent citations (Kuhn et al., 2019), citations arriving from patents with more than 250 backward citations were removed.