The impact of language translation quality on commerce: The example of patents

Abstract

Language matters, and it is an overwhelming stylized fact that language translation is an unavoidable part of global business. In this paper, we quantify the impact of translation difficulty reflected by the presence of multiple-meaning words in the original text. We focus on international patent applications because patent prosecution is nation-based. An inventor who seeks patent protection in a foreign jurisdiction with a different official language will need to file a translated version of the same document. Our estimates show that applications with more ambiguous original (English) text, are up to 25 percentage points less likely to receive a grant in the non-English jurisdictions (China, Japan, and South Korea). The results suggest that language translation difficulty can serve as a potential source of distortion in the global patent system. Ultimately, such translation difficulty may reduce the level of investment in global innovation activities, potentially leading to significant welfare loss. These findings serve to illustrate why international businesses should have adequate language translation strategy to address any translation difficulty arising from the presence of ambiguous words even when the deal involves the cross-border transfer of highly codified knowledge such as patents.

Résumé

La langue compte. Il est indéniable que la traduction linguistique est une partie inévitable du commerce mondial. Dans cet article, nous quantifions l'impact de la difficulté de traduction reflétée par la présence de mots à sens multiples dans le texte original. Nous nous concentrons sur les demandes de brevet internationales car les poursuites en matière de brevets sont basées sur les pays. Un inventeur qui cherche à obtenir une protection par brevet dans une juridiction étrangère dont la langue officielle est différente devra donc déposer une version traduite du même document. Nos estimations montrent que les demandes dont le texte original (en anglais) est plus ambigu ont jusqu'à 25 pourcents moins de chances d'être acceptées dans les juridictions non anglophones (en Chine, Japon et Corée du Sud). Les résultats suggèrent que les difficultés de traduction peuvent être une source potentielle de distorsion dans le système mondial des brevets. En fin de compte, ces difficultés de traduction peuvent réduire le niveau d'investissement dans les activités d'innovation mondiales, ce qui peut entraîner une perte importante de bien-être. Ces résultats illustrent la raison pour laquelle les entreprises internationales devraient disposer d'une stratégie de traduction linguistique adéquate pour faire face à toute difficulté de traduction découlant de la présence de mots ambigus, même lorsque la transaction implique le transfert transfrontalier de connaissances hautement codifiées telles que les brevets.

Resumen

El lenguaje importa y es un hecho ampliamente estilizado que la traducción del lenguaje es una parte inevitable de los negocios globales. En este artículo cuantificamos el impacto de la dificultad de la traducción reflejada en la presencia de palabras de múltiples significado en el texto original. Nos concentramos en las solicitudes de patentes internacionales ya que la búsqueda de patentes nacionales es de base nacional. Un inventor que busca la protección de una patente en una jurisdicción extranjera con un idioma oficial diferente va a necesitar la versión traducida del mismo documento. Nuestros estimados muestran que las solicitudes con mayor número de de texto ambiguo original en inglés son 25% puntos menos probables de recibir una beca en las jurisdicciones de habla no inglesa (China Japón y Corea del Sur). Los resultados sugieren que la dificultad de la traducción del lenguaje puede servir como una fuente potencial de distorsión en el sistema de patentes a nivel global. Finalmente esta dificultad de la traducción puede reducir el nivel de inversión en actividades de innovación global, lo que potencialmente llevaría a una pérdida significativa de bienestar social. Estos hallazgos sirven para ilustrar por qué los negocios internacionales deben tener una estrategia de traducción del lenguaje adecuada para atender alguna dificultad de traducción que surja por la presencia de palabras ambiguas, incluso cuando el asunto involucre la transferencia de conocimiento altamente cualificado como son las patentes.

Resumo

Idiomas são importantes, e é um fato estilizado esmagador que a tradução de idiomas é uma parte inevitável dos negócios globais. Neste artigo, quantificamos o impacto da dificuldade de tradução refletida pela presença de palavras com múltiplos significados no texto original. Nós nos concentramos em pedidos de patentes internacionais porque o processamento de patentes é nacional. Um inventor que busca proteção de patente em uma jurisdição estrangeira com um idioma oficial diferente precisará apresentar uma versão traduzida do mesmo documento. Nossas estimativas mostram que aplicações com texto original (inglês) mais ambíguo têm até 25 pontos percentuais menos probabilidade de receber uma concessão nas jurisdições cujo idioma oficial não é o inglês (China, Japão e Coréia do Sul). Os resultados sugerem que a dificuldade de tradução do idioma pode servir como uma fonte potencial de distorção no sistema global de patentes. Em última análise, essa dificuldade de tradução pode reduzir o nível de investimento em atividades globais de inovação; potencialmente levando a uma perda significativa de bem-estar. Essas descobertas servem para ilustrar por que as empresas internacionais devem ter uma estratégia de tradução de idiomas adequada para lidar com qualquer dificuldade de tradução decorrente da presença de palavras ambíguas, mesmo quando o negócio envolve a transferência internacional de conhecimento altamente codificado, como patentes.

摘要

语言很重要，语言翻译是全球商务不可避免的一部分，这是一个压倒性的典范事实。在本文中我们量化了原文中有多重意义的词语所反映的翻译困难的影响。我们专注于国际专利申请，因为 专利起诉是基于国家的。在使用不同官方语言的外国司法管辖区寻求专利保护的发明人需要提交相同文件的翻译版本。我们的估计表明，在非英语管辖区（中国、日本和韩国），原始文本（英语）更混淆的申请获得资助的可能性低了多达25个百分点。研究结果表明，语言翻译困难可能是全球专利系统中失真的一个潜在来源。最终，这种翻译困难可能会降低全球创新活动的投资水平；可能导致严重的福利损失。这些发现有助于阐明为什么国际商务应有适当的语言翻译策略，以解决因存在歧义词而产生的翻译困难，即使是在交易涉及例如专利等的高度编码知识的跨境转让的时候。

Zusammenfassung

Sprache ist wichtig, und es ist eine unumstößliche Tatsache, dass die Übersetzung von Sprachen ein unvermeidlicher Bestandteil des globalen Geschäftslebens ist. In dieser Arbeit quantifizieren wir die Auswirkungen der Übersetzungsschwierigkeiten, die sich im Vorhandensein von Wörtern mit mehreren Bedeutungen im Originaltext widerspiegeln. Wir konzentrieren uns auf internationale Patentanmeldungen, da die Patentverfolgung auf nationaler Ebene erfolgt. Ein Patentanmelder, der Patentschutz in einer ausländischen Gerichtsbarkeit mit einer anderen Amtssprache anstrebt, muss eine übersetzte Version desselben Dokuments einreichen. Unsere Schätzungen zeigen, dass Anmeldungen mit mehrdeutigem (englischem) Originaltext in den nicht-englischsprachigen Ländern (China, Japan und Südkorea) mit bis zu 25 Prozentpunkten geringerer Wahrscheinlichkeit eine Erteilung erhalten. Die Ergebnisse deuten darauf hin, dass Übersetzungsschwierigkeiten eine potenzielle Quelle für Verzerrungen im globalen Patentsystem sein können. Letztlich können solche Übersetzungsschwierigkeiten die Höhe der Investitionen in globale Innovationsvorhaben verringern, was möglicherweise zu erheblichen Wohlfahrtsverlusten führt. Diese Ergebnisse verdeutlichen, warum internationale Unternehmen über eine angemessene Übersetzungsstrategie verfügen sollten, um Übersetzungsschwierigkeiten, die sich aus dem Vorhandensein mehrdeutiger Wörter ergeben, zu bewältigen, selbst wenn es sich um den grenzüberschreitenden Transfer von hochgradig kodifiziertem Wissen wie Patenten handelt.

INTRODUCTION

It was the late 1980s and the delegation of Chinese academics visiting the University of Melbourne were discussing the use of regression analysis in economics. In the ensuing discussions, the interpreter from China, translated the term ‘unbiased estimators’ as ‘Marxist–Leninist estimators’. Luckily, there were good intentions on both sides and the meeting was a grand success.¹ Nevertheless, this anecdote highlights how language translation is not mechanical, even for highly technical terms and even when it involves highly codified knowledge.

Language matters, and it is an overwhelming stylized fact that trade, migration, and collaborations between countries with common national languages is significantly higher than those without, all other things considered (Cohen & Levinthal, 1990; Kogut & Zander, 1992; Palangkaraya, Jensen et al., 2017). As there are many benefits from interacting with businesses and organizations residing in nations of different language environments, language translation is an unavoidable part of global business (Piekkari, Welch, and Welch, 2014). However, not all organizations have the benefit of bilingual staff within their work teams when trading across linguistic divides. Many therefore resort to the use of machine translators and agents to negotiate deals.

In this paper, we quantify the impact of translation difficulty, as reflected by the degree of translation ambiguity measured by both the presence of multiple-meaning words in the original (English) text and the presence of bilingual team members, on business outcomes. As an exemplar, we use data on the successful examination of international patent applications – legal documents to recognize inventions. To the best of our knowledge, this study represents the first empirical estimate of the effects of translation difficulty on business outcomes using a large comprehensive database.

Our study extends the only existing study on the challenge in the global IP management arising from the translation of patent applications (Grünewald & Wurzer, 2013). Unlike Grünewald and Wurzer’s (2013) data, which is based on a subjective survey of a sample of fewer than 400 observations, our dataset provides an objective measure of the translation challenge and is drawn from a much larger sample of 115,423 patent applications. These data belong to 63,589 patent families which originated in English but were subsequently translated and filed in China, Korea, and Japan. A patent family refers to an (essentially similar) document that was lodged in several patent offices around the world. These filings were chosen as they required language translation from English to each respective official language of the patent office. The text from each application is then examined for sense and clarity at each office in their native language.

We use regression estimates to show how the probability of a patent grant is negatively affected by the degree of translation ambiguity,² and, that this probability improves with the involvement of a native speaker (as indicated by the presence of a local inventor in the patent applicant’s team of inventors). Our metric of translation ambiguity is defined as the percentage of multiple-meaning words in the original English text. Our measure of a successful business deal is whether the patent application passes examination and is therefore granted by said foreign patent office.

The remainder of this paper is structured as follows. In "Background" section, we provide a brief review of the studies on the importance of language translation for international businesses. In Sect. Empirical Framework we discuss the empirical framework including the construction of the language translation ambiguity measure, the specification of the estimating model, and the data for the estimating sample. In Sect. Results and Discussion we present the results and discuss their interpretation. Section Conclusion concludes.

BACKGROUND

With increasing globalization, the transmission of codified knowledge involving more than one language has been rising, at least since the 1980s (see Thomson, 2013; de Rassenfosse, Dernis, Guellec, Picci, & de la Potterie 2013; Buera & Oberfield, 2020). Language translation is big business. The European Commission estimated that, in 2004, it spent € 1–1.2 billion on translation services (Fidrmuc & Ginsburgh, 2007). On a grander scale, a marketing analysis report from Common Sense Advisory, a research firm, estimated that annual global language services cost US $34 billion and were growing at 12% a year (The Economist, 2012, as cited by Piekkari et al., 2014).

Language in the international business context is primarily about knowledge transfer and communication. Welch and Welch (2008) describe the basic communication model as: code a signal, transmit it, and then decode it. The effectiveness of transmission is affected by whether the sender encodes the signal fully and accurately, the medium or channel of transfer, and how the receiver decodes and understands the signal.

The focus of most knowledge transmission analysis to date has been on the transmission of tacit, not codified, knowledge. The successful transmission of tacit knowledge relies on fuzzy and imprecise channels such as face-to-face interaction, non-verbal forms of communication, social networks, staff movements, and relationships (Polanyi, 1958). Issues such as homophily, staff exchanges, social networks and communication modes between international groups dominate the academic discourse (Welch & Welch, 2008; Tenzer, Terjesen, & Harzing, 2017). To the extent language translation issues are raised, it is largely considered a coding problem to be solved by mechanical means. Fixman (1990), for example, found in her 32 interviews of nine US multinational companies that whereas cultural understanding was viewed as important, foreign language skills were regarded as a technical matter that could be solved by a hired translator.

An exception to this view is Janssens, Lambert, & Steyaert (2004) who argue that international business language strategies take three forms: mechanical, cultural, and political. Frequent misunderstandings arise if the translation process is only based on a mechanical language strategy, involving a simplistic process of replacing terminologies from one language to another (Holden & Michailova, 2014; Chidlow, Plakoyiannaki, & Welch, 2014; Brannen & Mughan, 2016).

Yet, we observe that in recent practice, because of the cost of employing human translation, a mechanical language strategy appears to be the chosen strategy of many international patent applicants. Language translation of patent documents increasingly rely on automated machine translations (Cavalier, 2001; Choi, 2009; List, 2012; Pouliquen, 2015; Wang, 2009). Unfortunately, there are no data to know for sure the extent on the use of machine or a mix of machine and human translation in patent prosecution. Smyth, Barker, and Belcher (2015) reported a case of the use of machine translation instead of certified human translation to support an EPO filing. The reason is to reduce translation cost. Interestingly, the author also highlighted the fact there was ambiguity in the machine translation, which resulted in costly appeal and further examination. However, a cursory Internet search would reveal various websites of patent attorneys, which seems to suggest that machine translation is usually recommended only for novelty search, but not for drafting patent application. For example, Nurminen (2019) claimed that whereas machine translations might have been used widely by patent professionals for decades, machine translations might be used increasingly only as the first step in the preparation of translation of patent application with additional human editing at a later stage. On the other hand, the USPTO Manual of Patent Examining Procedure contains some provisions for examiners to use machine translation to support, for example, a rejection decision.³ A similar allowance guideline is also implemented at the EPO.

The reliance on a mechanical language translation approach, involving either humans or artificial intelligence (AI), in international patent filing may be based on the belief that patent documents are highly formalized forms of codified knowledge and that the precision of codified knowledge promises more definite translation outcomes. But is this always the case? If the culturally specific context and elements of language use are important in identifying the specific meaning of any given terminology,⁴ we can expect that a translation of a highly technical document such as a patent application to be subject to critical errors if account is not make for context.

For multinational companies with workers, customers, government institutions, and business partners speaking and writing in many different native languages, language translation of written text is unavoidable (Brannen, Piekkari, & Tietze, 2014, p. 495; Piekkari et al., 2014). This is also true for companies seeking patent protection across different jurisdictions, each of which may require the submission of patent applications in a different language from the original language. Despite this need, existing research is limited to exhorting more investigation into the role of language translation on global innovation and operational success. There is a pervasive lack of hard evidence on the issue (Piekkari et al., 2014). Unlike the literature that investigates difficulties in cross-border transmission of tacit knowledge, studies that analyze problems with the transmission of codified knowledge via text translation are often done through a series of anecdotes (e.g., Welch, Welch, & Marschan-Piekkari, 2001).

In this paper, we investigate how the ambiguity in the original language of codifying knowledge affects business outcomes by analyzing a large sample of text data from international patent examination outcomes. We ask whether multinational companies need to move beyond a mechanical perspective on translation when seeking patent protection in foreign jurisdictions. We investigate how much these companies can gain by taking a cultural perspective to correctly interpret words with multiple meaning in the culturally dependent context given by the written text. For the case of patents, this means that the same technical terms may have different meanings in different languages or that the same concepts may need to be defined in written formats in different ways. This need for taking a cultural perspective is potentially significant for translation of patent documents because new terms or new way of using existing terms in the original language may often be use to describe the novel invention (Osenga, 2006).

The translation of highly technical text such as patent claims is a complex and difficult process. Lise (2021) argues that a patent translator needs to understand what the claims mean in the original language (that is, to correctly interpret the claims) and how to translate them to the target language to claim the same invention. Evidence suggests that understanding what patent claims mean in the original language is far from straightforward.⁵ Osenga (2006) argues that to understand what is being conveyed by the written words in patent claims the interpreter needs to consider both internal and external linguistic clues. The first set of clues consists of “syntactic hints” from the structure of the grammar and sentence of the language of the patent claims. The second set of clues consists of non-linguistic factors or situational pragmatics constituting the underlying knowledge of the interpreter (and translator), which are helpful for a correct interpretation of the claim text. In this sense, a mechanical perspective of translation is unlikely to produce an accurate translation due to its sole focus on the internal linguistic clues.

By contrast, a cultural perspective of translation would consider both linguistic and non-linguistic clues. Such a “culturally aware” approach is in practice the other side of the same coin referred to by investing in “bilingual” or “local” experts. Golden (2008) argues that there are both technology-centered and lawyerly aspects of claim meaning. Even in the lawyerly aspects, for a patent attorney to accurately interpret the meaning of the claims, he or she requires both legal expertise and access to the technical knowledge of the artisan. Therefore, the presence of a “bilingual” local inventor who is fluent in the original and target (local) language would enable the implementation of a cultural perspective approach to translation. In terms of Osenga’s (2006) argument above, we can expect the access through a “bilingual” local inventor to improve the accuracy of the translation because of his or her superior internal and external clues to correctly interpret and translate patent claims.⁶ Thus, for example, a bilingual inventor⁷ could have prevented the incorrect translation of the Italian word “semiliquido” as “half liquid”, which can mean half solid and half liquid instead of the intended meaning of a gel-like substance or “semi-liquid”.⁸ Our empirical analysis aims to confirm this hypothesis.

A recent study has shown how machine translation benefits international trade (Brynjolfsson, Hui, & Liu, 2019), however it is likely that even with state-of-the-art machine translators, disambiguity in word meaning can lead to errors in machine translation. There is evidence that sense disambiguity as reflected by the extent of polysemous words in the text is associated with one of the most pervasive types of machine translation errors (Costa, Ling, Luís, Correia, & Coheur, 2015) or lower machine-translation quality (Araghi & Palangkaraya, 2019). A classic polysemous word is ‘bank’, which can mean a type of financial institution or a sloping river side. In this study, we assume that text ambiguity can be measured by the extent of words which have multiple meanings (that is, polysemous words) in the source language. Mishra, Bhattacharyya, and Carl (2013) found a positive relationship between linguistic features of a sentence to be translated including "average of words polysemy" on the difficulty of a sentence to be translated based on human translator experimental data. Larroyed (2018) estimated that machine-translated patent information disclosed an average of 80% of the original content. Whether or not the 20% deficiency is critical for patent examination and systematically related to observed characteristics such as those of the invention and the inventor is an open question.

Patent applications have several advantages as the subject for quantifying the impact of translation difficulty. Patents are legal monopoly rights granted to those who can prove in a written document that the described invention meets certain conditions. Language translations of codified patent knowledge are common – the data are abundant – and there is a clear and reasonably objective indicator of success – a technical decision about whether to grant the patent. Multiple patent data are available for the same invention or the same examining office, which allows us to hold constant the quality of the invention and the stringency of the patent office, so we can identify language ambiguity. For our purposes, what is relevant is that the original text clearly communicates the description of the invention and why it meets these conditions. Inventors will almost always use a specialist patent attorney who is skilled in the relevant technical area to write the application, in our case in English. Applications must then be made separately to each national patent office (with the possibility of treating Europe as one jurisdiction) in the office’s national language. Each application will have a patent attorney who is local to each office. If a translation from the original language is required, the local patent attorney will subcontract this to a translator who commonly uses both AI and humans to translate the document. The costs of a translation is considerable, and has been estimated to vary from US $1600 to US $4600. The application is then judged by patent office examiners to see if it passes the patentability test. If the application is unclear or confusing, we expect that the probability of a grant will be reduced.

The existing literature suggests that potentially costly translation errors can become a problem at every step of the patent prosecution. For example, costly disputes due to translation errors may come in the form of invalidation and rejection of patents by judicial and quasi-judicial bodies. Scott and O’shea (2021) discussed various cases related to translation repercussions that may arise in the form of discrepancies in translated patent claim terms, ambiguity in machine translations, patent infringement and malicious translation exploitation, and unenforceability of patent. Chisum and Farmer (2009) provide an in-depth discussion on patent translation, translation errors, and the impact of translation errors. Lastly, Grünewald and Wurzer (2013) reported that a survey of European IP professionals conducted by Steinbeis Institute for Intellectual Property Management found that incorrect translations of patent applications commonly occurred and could have seriously negative consequences on the scope of protection of the granted patent. The survey also found that budget constraints may lead to filings in fewer jurisdictions if the translation costs are too high. These findings suggest that the negative impact of translation difficulty can potentially go beyond this paper’s measured impact in terms of the probability of patent grant.

EMPIRICAL FRAMEWORK

The Model

For any patent application i written in English, we estimate the impact of translation ambiguity and a bilingual team member on the examination outcome in a given patent office j using a reduced form binary dependent model. Specifically, denoting patent examination outcome of application i filed and examined at patent office j and ${y}_{ij}=1$ as granted and ${y}_{ij}=0$ as rejected, following the approach of earlier studies such as Guellec and van Pottelsberghe (2000) and Webster, Jensen, and Palangkaraya (2014), we estimate the conditional probability of grant as:

$$Pr\left[{y}_{ij}=1|{T}_{i},{{L}_{ij},{\varvec{X}}}_{{\varvec{i}}},\alpha ,\beta ,\delta ,\theta ,{\varvec{\gamma}}\right]=\text{f}\left(\alpha +\beta {T}_{i}+{\delta L}_{ij}+{\theta {T}_{i}L}_{ij}{+{\varvec{X}}}_{{\varvec{i}}}{\varvec{\gamma}}\right)+{\varepsilon }_{ij}$$

(1)

where f(.) can be either a linear or non-linear logistic function; ${T}_{i}$ is translation ambiguity level defined as the proportion of polysemous words in the first claim; ${L}_{ij}$ is whether the inventor is local to office j; and ${{\varvec{X}}}_{{\varvec{i}}}$ denotes a vector of patent application i’s characteristics such as technology area, length of the first claim, number of claims, which are known to affect the examination outcome.⁹

The use of a binary dependent variable in the estimating Eq. (1) implies at least two important simplifying assumptions. First, we assume that all grants are the same. When a patent application is granted, it might be granted with fewer claims and/or narrower claims. This assumption means that we may underestimate the negative impact of translation difficulty. Second, we do not consider that translation difficulty may have a deterrent effect on the likelihood of a business to file or fully prosecute a patent application in a foreign jurisdiction.¹⁰ Thus, we may also underestimate the impact of translation difficulty. However, these assumptions have important practical benefits which may outweigh their costs. First, we do not have readily available information to identify which claims are granted. Second, by excluding withdrawn or deemed withdrawn applications, we can interpret our findings more cleanly on patent examination outcomes. Otherwise, the estimated effect may reflect a mix of patent office and applicant decisions.

Our main interest is in the size and significance of the coefficients $\beta$ and $\theta$, which reflects the extent to which ambiguous translation affects success and the ability of a bilingual team member to ameliorate the negative effects of this ambiguity. The benefit of having a bilingual team member could be due to factors other than reducing the increased level of translation difficulty arising from the presence of ambiguous words. This might include general cultural familiarity, homophily with the examiners and attorneys, and the trust and subtle communication skills that flow from these skills. To identify the part of their skill set that is most clearly associated with translation per se, we focus on the interaction term between local inventor and translation ambiguity.

Our construction of the key variable ${T}_{i}$ has been inspired by the computational linguistic literature to proxy for how prone the translation is to error (that is, translation difficulty). As the value of a patent is closely related to its scope of legal protection defined by the claims (Adams, 2016), our text analysis focuses on the text of the patent claims. Furthermore, the translation ambiguity problem is likely to be more significantly revealed by the text of the patent claims than the descriptive text given patentees’ self-interest in obtaining broader patent scope by choosing more vague words in the drafting of the claims (Chiang, 2015; Chien, 2016; Freilich, 2018). Specifically, we took the English text of the first claim of each patent application family in our sample and compute the proportion of words with potentially multiple meanings or ambiguous senses (that is, in the linguistic terminology, the proportion of polysemous words, Jurafsky & Martin, 2018).¹¹

${T}_{i}$ is measured by the proportion of words with multiple senses (i.e., polysemous words) in the text of the first patent claim as specified in Eq. (2) below:

$$T_{i} = \frac{{{\text{Number\,of\,Polysemous\,words\,in\,Patent}}\, i{'}{\text{s\,first\,claim}}}}{{{\text{Number\,of\,words\,in\,Patent}}\,i{'}{\text{s\,first\,claim}}}}.$$

(2)

The Appendix provides further details on the construction of ${T}_{i}$ and a validation analysis on how it is systematically related to measured translation difficulty.

Estimating Sample

We used data on all patent families which originally filed in English between 2000 and 2006 and had a corresponding filing in their national language at the Chinese Patent Office (CNIPA), Japanese Patent Office (JPO), and Korean Patent Office (KIPO). We need data on applications sent to multiple offices to enable us to control for the quality of the underlying invention and these three Asian offices represent the largest offices in the world requiring a translation from English to non-English text. The source of our data is de Rassenfosse et al. (2019), who constructed data on patent application families simultaneously filed in at least two of IP5 patent offices (European Patent Office, US Patent and Trademark Office, CNIPA, JPO, and KIPO) based on the Worldwide Patent Statistical database PATSTAT and, most importantly, accurate information on the examination outcome identified from INPADOC PRS table for PATSTAT, CNIPA’s on-line patent search platform, JPO’s public access on-line Industrial Property Digital Library Database, KIPO’s public access on-line IPR Information Service, and USPTO Public Pair on-line database. Into these data, we add detailed patent claims text data from the A1-A3 and B1-B2 xml files from the EPO Register Backfile.

Our estimating sample was derived from a population of 771,736 patent application families with priority year 2000–2006. Given the lengthy time interval between application and examination in many patent offices, these priority years allows for most applications to have had a final examination decision. These patent families have 1,264,735 patent application members filed in at least two IP5 offices. From these applications, we select patent families with English as the language of the priority application and which members are filed and examined at CNIPA, JPO, and/or KIPO. This sampling design is to ensure that all applications in our estimating sample need to be translated from English before they are examined for patentability in each of the focal offices. We note that the selected families with English priority filing may receive either grant or reject decision at the EPO and USPTO. The sampling restrictions leave us with 63,592 patent families with 115,429 patent application members. Ultimately, the resulting estimating sample is slightly smaller due to missing values in regression variables: 63,589 families; 115,423 applications. Table 1 summarizes the estimating sample.

Table 1 Sample descriptive

From Table 1, about 48% of the families have an application at the JPO (being the largest office among the three). Around 63% of the sample is granted a patent and the grant rates vary slightly across offices, with the JPO having the lowest grant rate (a finding consistent with many other studies). The first claim has an average length of around 120 words with insignificant variation across the application set examined by the offices despite the much more significant variation in the sample size. In terms of the identity of the inventors or the applicants, there are many more patent applications filed by local inventors in Japan than in the other two countries. The cross-county variation seems to be consistent with the variation in the extent of innovative activities.

RESULTS And DISCUSSION

Main Findings

Table 2 summarizes the estimated coefficients and their associated clustered standard errors based on linear regression (OLS) and logistic regression (logit) estimation of Eq. (1). We estimate the later to assess if we can use the simpler linear regression in subsequent analyses. Also, for each model, we estimate a baseline model with minimum standard control variables (local inventor, office fixed effects, technology class fixed effects,¹² priority year fixed effects, and PCT filing fixed effects) and an extended model with additional controls for patent characteristics (number of words in the first claims, number of polysemous words in the first claims, number of claims, and number of forward citations). For the logit models, Table 2 also provides the marginal effect estimates of our main variable of interest (Translation ambiguity ${{\varvec{T}}}_{{\varvec{i}}}$) and the control variable of interest (local inventor fixed effect).

Table 2 Baseline model regression coefficient estimates (OLS and logit)

Our main interest is on the estimated coefficients of ${T}_{i}$ which range from – 0.250 to – 0.195 based on the linear regression or from – 0.243 to – 0.186 based on the (marginal effects) of the logit regression. These estimates mean the probability of grant decreases by up to approximately 25 percentage points when the text of the first claim goes from having no ambiguous word to having all words as ambiguous. This effect appears to be significant given the average grant rate is about 64%, as shown in Table 1. The effect suggests that even for a highly codified text such as patent applications, a mechanical language strategy in which translation is viewed simply a walking dictionary is likely to be inadequate (Janssens et al., 2004).

Figure 1 shows graphically how the predicted probability of grant based on the logit model coefficient estimates in column (4) of Table 2 decreases as translation ambiguity increases from 0 to 1. The figure also shows the 95% confidence interval and the distribution of the estimating sample of patent applications.

Figure 1

Probability of grant as a function of translation ambiguity.

The negative effect of translation ambiguity on probability of grant appears to be robust to the inclusion of control variables such as number of words in the first claim and number of claims. The first controls for unobserved factors that may affect the relationship between translation ambiguity and patent grant. For example, all else equal, shorter claims might be regarded by patent examiner as broader (that is, less specific) claims and thus are less likely to be granted. The number of claims captures any heterogeneity in patent applications in terms of patent value/quality/scope/complexity that is not captured in the translation ambiguity measure which is constructed based only on the first claim. The robustness of the relationship between translation ambiguity and grant probability is further confirmed by the addition of another variable often used in studies of determinants of patent grant and patent value: forward citation counts. Studies have shown that forward citations are an important proxy for invention quality and hence patentability and patent value (Carpenter et al., 1981; Narin et al., 1987; Narin & Olivastrao, 1988; Trajtenberg, 1990; Karki, 1997; Albert et al., 1991; Sampat et al., 2003; Czarnitzki, Hussinger, & Schneider, 2011).

In Table 3, we present fixed-effects linear regression estimates to assess how much of the language translation effect can be accounted for by variation in the expertise to translate. Specifically, we extended the baseline model in Table 2 by including fixed-effect dummy variables for each patent attorney and each applicant (i.e., the business) to control for the quality of the patent attorneys who drafted and prosecuted the application and the inventing firm. Compared to the estimates in Table 2, the coefficient estimates of ${T}_{i}$ in Table 3 are about 25% lower suggesting that a portion of the language translation effect can be explained by cross-applicant or cross-attorney variation in ability to translate the patent application.¹³ However, there is still a major portion of the language translation effect which can be explained neither by the patent application characteristics nor the inventor/applicant/attorney characteristics.¹⁴

Table 3 Applicant and attorney panel fixed effect model estimates

Finally, Janssens et al. (2004) argue that language translation needs to account for cultural and political factors. If this theory is true, then the presence native speakers in the inventing team should improve the ability of the team in addressing any translation difficulty arising from the presence of polysemous words as they have a better grasp of the local cultural nuances in word ambiguity and can lessen technical misunderstandings which occur in mechanical translations.

The estimated linear regression coefficients in Table 4 show the interaction between local inventor effect and translation ambiguity across offices. First, from the first column of the table, we can see a positive interaction effect: the negative effect of translation ambiguity is 8.2 percentage points lower if a local inventor is named on the application (or, in relative terms, about 30%). This effect is over and above the positive impact of a local inventor being present which may indicate cultural factors. In other words, it appears that a part of local inventor bias found in the literature is due to language-related factors. It might be possible that local inventors are better able to solve the translation ambiguity problem because they are relatively knowledgeable about the local meaning of words (of the region and of the patent office). In other words, a common language between the sender of the message (the inventor) and the recipient (the examiner) appears to matter in international patent examination, contradicting the basic assumption of the mechanical language strategy. The rest of Table 4 shows that there is a relatively significant variation across offices, indicating that the language effect might be target language dependent (that is, the translation ambiguity effect is worse for translation from English into Korean and Chinese).

Table 4 Interaction between translation ambiguity and local inventor presence and separate office estimates

In addition to the OLS regressions of the interaction model presented in Table 4, we also estimated the corresponding logit regressions. The average marginal effects of translation ambiguity (${T}_{i}$) for local and foreign inventors based on the logit coefficient estimates are summarized in Table 5.¹⁵ Overall, the logit model provides similar evidence to the interaction effect OLS estimates summarized in Table 4. In all offices, foreign inventors suffer a larger negative effect from translation ambiguity when compared to local inventors. However, the difference varies across offices with local inventors in China showing much less significant effect.

Table 5 Average marginal effects of translation ambiguity, local and foreign inventors, all and separate office estimates

Instead of looking at how the average marginal effects of translation ambiguity differ between local and foreign inventors, we can use the positive interaction effects (${T}_{i}$ x Local inventor) to see how the local inventor bias identified by earlier studies such as de Rassenfosse et al. (2019) is related to translation ambiguity. That relationship is shown graphically in Figure 2 below. The figure shows that the marginal effect of local inventor status increases with the level of translation of ambiguity at a slightly faster rate for the Korean patent office (KIPO) and the Chinese patent office (CNIPA), than the Japanese patent office (JPO).

Figure 2

Marginal effect of local inventor as a function of translation ambiguity. The shaded areas represent the 95% confidence intervals.

Is There an Alternative Interpretation?

We now discuss several aspects of the analysis that may suggest a different interpretation of the main findings discussed earlier. First, it is possible that translation difficulty leads to a poorly written application which could then lead to amendments of the original application following the direction of the patent examiners. If that is the case, we should not see a significant translation ambiguity effect in our data analysis because the applicant would always be able to revise their application. If we do see a significant translation ambiguity effect, as is the case in this paper, then we may not be able to interpret the estimated effect as purely the effect of translation difficulty.

Without data on the extent of correspondence between applicant and examiner, it is not possible to test which one is the most likely explanation. However, there is an argument that incorrectly translated application may lead to straight refusal at the JPO instead of a process of amendments (Kennedy, 2000). Furthermore, correcting mistakes can be complicated due to issues related to the introduction of new matter and the timing restrictions (Holthaus, 2019). In his Law Lecture Series, Nakayama (2016) provides detailed discussions of the complexity related to translation and correction of translation errors when filing patent application in Japan with prior art claims originated in non-Japanese language. The timing window and other restrictions in providing correct translation can be too restrictive, leading to a refusal or deemed withdrawn outcome. In other words, it appears that if there is any interaction triggered by possible translation errors, it is unlikely to lead to amendments which nullify our ability to detect the effect of translation difficulty in the ultimate patent examination outcome as postulated above.¹⁶

Furthermore, the significant interaction effect estimate between the two measures of translation difficulty (as shown by Figure 2) suggests that the local inventor effect is greater when translation ambiguity is higher. This is consistent with an interpretation that the translation ambiguity measure captures the negative effect of translation difficulty and that local inventor partially alleviates the translation difficulty problem presented by the presence of ambiguous words. However, because both main-effect coefficients area (statistically) significant, the local inventor coefficient estimate may also reflect other local advantages effects unrelated to language translation difficulty such as (cultural) familiarity with the local patenting process (de Rassenfosse & Hosseini, 2020; Webster et al., 2014).

Secondly, we have analyzed the patent application outcomes of our sample of patent families with English priority patent application which are filed in the Chinese, Japanese, or Korean patent offices. However, we have ignored the possibility that patent applicants from either China, Japan, or South Korea may file their priority applications directly in an English-speaking country such as the United States due to its importance in the global market share. In such cases, it is not clear if we can interpret the translation ambiguity effect purely as representing an underlying translation difficulty. Intuitively, Korean applicants/inventors who filed their priority applications in the US, for example, may experience less “translation ambiguity” problem when “translating” their English priority filing into Korean patent application filing.

Hence, to obtain a cleaner interpretation of the translation ambiguity variable as representing a translation difficulty effect, we re-estimated our main regressions using four subsamples based on whether at least one of the applicants or the inventors is local in each respective patent jurisdiction. These subsamples are: local applicant, foreign applicant (i.e., no local applicants), local inventor, and foreign inventor (i.e., no local inventors). To deepen our understanding, we also conducted a further split sample analysis based on the patent office. The re-estimated regression coefficients are summarized in Table 6 below. The results show that the coefficient of translation ambiguity is almost three times stronger for foreign applicants and inventors. This confirms our interpretation of the translation ambiguity coefficient as representing translation difficulty. However, as shown for the case of the Japanese Patent Office, the local applicants/inventors are not necessarily free from the translation ambiguity problem. One possible reason is that translation ambiguity can work in the reverse direction, that is Japanese local inventors may have introduced “translation” errors when drafting and filing their English priority applications in the English-speaking patent office, which then manifested back in the Japanese version.

Table 6 Subsample analysis – Comparison of local and foreign applicants/inventors

Thirdly, patent applicants may use more ambiguous words in their patent applications as a cover up for a low-quality underlying invention or that lower quality patent applicants are more simply more likely to draft ambiguous patent applications for any reason. In our baseline analyses, we tried to account for such possible correlation between invention quality and translation ambiguity by including forward citation count as a control variable. However, this may not be enough, especially if patent examiners in the USPTO or EPO which examined the English priority filing incorrectly grant (that is, being more lenient) a patent for a lower quality invention whenever the applications contain more ambiguous words.

Although leniency at the EPO and USPTO examination practice is a possibility, we believe it is highly unlikely because the EPO is known as one of the patent offices with the highest bar in terms of non-obviousness, novelty, or inventive step requirement. In fact, our estimates of an extended baseline model summarized in column (1) of Table 6 below show that patents which are granted by EPO are more likely to be granted by the three Asian patent offices. Also, our results hold even after including control variables such as the whether the patent families received a USPTO and/or EPO grant and the number of forward citations received.

To test further whether our main finding is driven by correlation between underlying invention quality, translation ambiguity, and leniency in the English-speaking patent office, we constructed a ‘novelty’ measure, following Aghion et al. (2021), who followed Kelly et al. (2018), based on patent full-text embedded vectors from Google Patents Research Data. This ‘novelty’ measure is computed as the average cosine distance between the full text of each patent application in our sample of patent families and the full text of all other US patents filed in the preceding 58 years in the same three-digit CPC patent class. As argued by Aghion et al. (2021), this text-based novelty measure can serve as a good measure of technology breakthrough. By definition, a highly novel patent application would be more likely to get a grant decision. Thus, if a highly novel patent application is negatively affected by translation ambiguity, then the effect can be interpreted as the effect of translation difficulty rather than the effect of a low-quality underlying invention or patent examiner’s leniency.

In column (1) of Table 7, we added ‘novelty’ as an additional control variable into the baseline regression. The positive estimated coefficient of novelty is consistent with the interpretation that more novel patent applications are more likely to be granted. Also, the negative coefficient of Translation ambiguity can be interpreted as the negative effect of translation difficulty, since we have controls for invention novelty. Furthermore, we split the sample into families which are highly novel (defined as the top 25% patent applications in terms of novelty) and of low novelty (defined as the bottom 25%). The coefficient estimates in column (2) show that even among the highly novel patent families, a higher language ambiguity is associated with a lower probability of grant. In contrast, as shown in column (3), the language effect is not statistically significant for low novelty patent families.

Table 7 Baseline model regression coefficient estimates (OLS), controlling for EPO and USPTO grant status and novelty and split-sample based on EPO and USPTO grant status and novelty

Lastly, we conducted split sample analysis by focusing only on patent families which are rejected by USPTO (and, separately, by EPO) to rule out any possibility of lenient grant by the English-speaking patent offices. The coefficient estimates in columns (4) and (5) further suggest that the negative impact of language ambiguity is unlikely to be driven by regression-to-the mean.

CONCLUSION

As global value chains expand, the need for more seamless transactions across national and cultural boarders will grow. Costly and inaccurate translations can amount to a behind-the-border trade barrier, which reduces international business below its optimal level.

In this study, we quantified the effect which language ambiguity may have on the success of documents requiring language translation. We have used patent applications as an example. Because the patent prosecution system is nation-based, an inventor who seeks patent protection in a foreign jurisdiction with a different official language than the language used in the original priority document will need to file a translated version of the same document. This requirement not only increases the patenting cost to the inventor (an “innovation tax” according to Van Pottelsberghe, 2011), but, more seriously, it increases the potential for translation errors and unfavorable outcomes.

Translation matters because the language used affects people’s behavior. For example, Packard and Berger (2017) found implicit endorsement statements such as “I liked it” or “I enjoyed it” as much less persuasive than explicitly endorsement such as “I recommend it” because the latter is perceived as a stronger signal. If the original word to be translated has multiple senses and therefore multiple translation words, then which words are chosen during the translation may determine the examiner’s action.

Our estimates showed that applications with more ambiguous original English text, are up to 25 percentage points less likely to receive a grant after translation, all else equal. Foreign inventors faced double whammies in the international patent system: first, because they are foreigners with potential disadvantages from lack of familiarity of the foreign system; and, on top of that, because they must file a translated version of their original application. Theoretically, the establishment of the patent system is aimed at the promotion of innovation by encouraging investment in the generation and commercialization of new knowledge and ideas. This objective can be achieved if there is no distortion in the patent prosecution process. The results of our analysis suggest that language translation difficulty can serve as a potential source of distortion in the global patent system beyond the direct effect on each patent applicant. If translation difficulty reduces the ability to obtain patent protection in foreign language jurisdictions, it may ultimately reduce the level of investment in globally beneficial innovation. Thus, translation difficulty could lead to not only losses on the effected specific patent applicants but also to a potentially significant welfare loss.

These findings serve to illustrate why international businesses should have adequate language translation strategy to address any translation difficulty arising from the presence of ambiguous words even when the deal involves the cross-border transfer of highly codified knowledge. Inadequate attention to potentially ambiguous words in the source language can be costly if a purely mechanical translation strategy is adopted. However, businesses may be able to complement a mechanical strategy with a cultural perspective to ensure that words with potentially culturally dependent ambiguous meaning are correctly translated through the employment of native speakers.

We conclude with a discussion on the main limitation our study and a possible extension. Our analyses, based on the interaction effect with local inventor and various sub-samples, seem to suggest that the negative effect of our translation ambiguity measure on patent grant probability can be interpreted as the negative effect of translation difficulty. However, a higher translation difficulty does not necessarily mean a lower translation quality. Hence, the estimated negative relationship between translation difficulty and the probability of patent grant may still be open to an alternative interpretation. To address this limitation requires a construction of a translation quality measure. One possible way to construct such a measure is to borrow the technique from the machine translation quality evaluation literature, by comparing the original English application with a ‘re-translated’ English versions of the corresponding foreign language application. We believe the text data are available from the patent offices or Google Patents Public Data, and the main task to figure out is how to ‘re-translate’ the foreign language application. Finally, we have provided evidence on the potential negative impact of language translation difficulty on global IP protection using a relatively simple measure of the ability to obtain a patent. It would be interesting to analyze a more complex measure of the ability to protect an invention globally by looking at how translation difficulty or translation quality is related to the scope of the granted patent claims. Also, another potentially fruitful extension to our analysis could be aimed at answering the question of why or when language translation difficulty matters by looking at data on the interaction between the patent applicant and patent examiners.

Notes

¹We thank Professor Peter Lloyd of the Gruber-Lloyd Intra-Industry Trade index (Grubel & Lloyd, 1975) for sharing this personal anecdote.

²In this paper, we use the term translation difficulty and translation ambiguity interchangeably. This is because our measure of translation difficulty is based on polysemous words with potentially ambiguous meaning.

³https://mpep.uspto.gov/RDMS/MPEP/current#/current/d0e122292.html (Accessed on 7-31-2022).

⁴See Kassis Henderson (2005).

⁵In the US, the Federal Circuit reversed district court judge’s claim construction one third of the time (Osenga, 2006).

⁶Anecdotally, it is not uncommon for patent attorneys to seek the assistance of a bilingual local inventor in the process of translation of patent documents from the original to local (target) language. Compared to external translators, a local bilingual inventor would have a better underlying knowledge of the invention and at least as good understanding of the linguistic clues. Also anecdotally, highly resourced patent applicants such as Samsung may be able to employ separate teams who work closely with the inventors for drafting customized patent applications depending on the target language. This is an extreme example of taking a “cultural perspective” of translation.

⁷Or a bilingual patent attorney or translator who has the “cultural” knowledge about the meaning of the technical terms used.

⁸See Chen (2020) for more details about the case.

⁹We denote translation ambiguity as ${T}_{i}$ with an i-subscript instead of an ij-subscript as we assume that the only source of translation ambiguity is the presence of words with multiple senses in the original English text. One may argue that the target language can also serve as a source of translation ambiguity. Hence, our measure is an underestimate of the true level of translation ambiguity. We assume that such underestimation is random across patent applications with respect to our main variable of interest (locality of the inventor) at least within the single office estimation (in which case there is no variation in the target language). To a certain extent, if the variation in the target language effect is systematic, we shall see the effect of translation ambiguity varies significantly when we estimate Eq. (1) using data from different patent office separately (which is not what we found as discussed later).

¹⁰A recent study by Petit, van Pottelsberghe de la Potterie, and Gimeno-Fabra (2020) shows that an extension of our analysis which includes withdrawn applications can potentially reveal a different aspect of how language translation matters in global IP protection.

¹¹Our choice to focus on the first claim follows the argument of Kuhn and Thompson (2019) who showed that “the breadth of a patent’s scope can be measured by counting the number of words in its first claim” (p. 6). Hanson (2015) argued that “[o]ften the first claim lays out the broadest rights, with subsequent claims narrowing the invention’s scope”. Bar and Costello (2017) similarly argued that “[t]he first claim in a patent is likely to be one of particular importance, and breath.” Other studies have also focused their analysis on the first claim. For example, Righi and Simcoe (2019) used the length of the first claim as a proxy measure for patent scope to investigate examiner specialization. Okada, Naito, and Nagaoka (2017) found the breadth of the first claim as a significant predictor of Japanese patent’s knowledge impact on subsequent inventions. It is possible however that for the granted patents in our sample, the set of the granted claims does not include the first claim. In that case, our grant probability estimation should count such granted patents as rejected patents. Otherwise, it is possible that our choice of only looking at the first claim underestimates the effect of translation difficulty. Practically speaking, all patent applications have a first claim and all first claims are an independent claim. The alternative of randomly choosing which claim to use may introduce a possible source of variation through the mixing of comparing independent and dependent claims. Either way, it is not clear which is the better approach. Detailed claim changes data are only available at the USPTO and, in the future, this issue should be revisited.

¹²For the technology class fixed effects, we use 595 technology dummy variables representing four-digit International Patent Classification (IPC) codes in our sample. For patents with multiple IPC codes, we use the first listed IPC code following the common practice in the literature.

¹³de Rassenfosse et al. (2018) investigated and found a significant role of the quality of the attorney on patent examination outcomes.

¹⁴In separate regression results, not shown in the paper but are available upon request, instead of patent attorney fixed effect, we used two measures of patent attorney’s experience: the number of patent applications handled, and the number of patent applications handled and granted prior to the current one. The findings from these regressions are similar to the estimates shown in Table 3.

¹⁵The coefficient estimates of the logit model are summarized in Table A2 in the Appendix.

¹⁶Anecdotally, when we consulted with an expert patent attorney/lawyer with over 40 years of experience in all the major offices, we were told that (a) he never heard of an examiner asking the applicant to clarify the language and (b) the examiner would not know that the translation was poor unless they compared the translated document with the original – and they don’t do this (even were they to be bilingual). Also anecdotally, a German patent office examiner refused to even examine an application which appears to be a machine translation of an original Chinese priority application. At the JPO, and perhaps similarly in other offices, if there is any ambiguity in the interpretation of the translation provided by foreign patent application with foreign prior art claim, then the Japanese translation text prevails. The JPO examiners are unlikely to consult the foreign language materials. They may ask for translation correction or they may not (Chisum and Farmer (2009).

¹⁷That is, in Melitz and Toubal (2014)’s terminology, the efficiency in communicating indirectly.

¹⁸Even without ambiguous words, translators may be faced with the lack of equivalence arising from different causes such as concepts which are culture specific or differences in physical and interpersonal perspective and expressive meaning (Chifane, 2012). In fact, these factors may lead to severe consequences even when no translation is required due to people from different fields may interpret the same terms differently (Cremer, Garicano, & Prat, 2007).

¹⁹WordNet is freely and publicly available for download and its structure makes it a useful tool for computational linguistics and natural language processing. There are more than 14,000 citations in Google Scholar to WordNet guidebook (Miller, 1998), indicating the wide adoption of the tool. In our study, we use WordNet 2.1, released in March 2005, the latest version of WordNet for Windows system in the Natural Language Processing Toolkit Python package NLTK 3.4.5.

²⁰For more discussion about the source texts, machine translations, and human evaluation translation quality scores in each of the annual workshops, see also Bojar et al., (2017, p. 208) and Barrault et al. (2019).

²¹The sample consists of 1594 observations from WMT 2017, 3192 observations from WMT 2018, and 1956 from WMT 2019 workshops. We selected only the English-to-Chinese translation sample to match with the patent jurisdictions studied in this paper. Unfortunately, the English-to-Japanese and English-to-Korean translations are not available in the WMT Workshop data.

Appendix: Proxy measure of translation difficulty

Rationale

Our key variable as a proxy measure of translation quality is essentially a measure of source text’s translation ambiguity (${T}_{i}$). We are motivated to construct such measure by the “equivalence paradigm” which is often used in the international business literature on cross-language research methods (Chidlow et al., 2014). Janssens et al. (2004) identified the central role of the “equivalence” concept when discussing “translation as walking dictionaries” in the mechanical language strategy. Under this concept, we basically assume that the objective of the translation of a patent claim is to obtain a translated claim text in the target language that is equivalent or of “equal value” to the original claim text in the source language (Pym, 2007). In other words, translation ensures the function of the original language in enhancing the efficiency of communication of the claimed invention (Marschak, 1965; Lazear, 1999).¹⁷

The objective of equivalence can be achieved if both original and translated texts have “identical meaning” (Hult et al., 2008). Our proxy variable essentially measures any potential deviation in meaning between the two texts that can be inferred from the extent of words in the source text which have multiple meanings (that is, polysemous words). Specifically, we compute the proportion of ambiguous words in the English patent claim text (the source text) in our patent application sample which are also filed in non-English-speaking jurisdictions: China, South Korea, and Japan. We call this measure translation ambiguity, that is the ambiguity level for translating the original English patent claim text (i.e., the error level arising from how much the translated claim text of “equal value” satisfies the “equivalence” paradigm).¹⁸

Translation Ambiguity (${{\varvec{T}}}_{{\varvec{i}}}$)

To simplify the construction of ${T}_{i}$, we limit the amount of text to be processed by focusing only on the first claim, as opposed to all claims. It has been argued that the first claim is usually the most important independent claim of a patent application (Kuhn & Thompson, 2019). The focus on the first claim is also beneficial as all patent applications have at least one claim and the number of claims varies across applications. Because of the latter, it is harder to ensure we are comparing apples-to-apples if the comparison between patent applications is based on the full text of all claims. Implicitly, by only comparing the first claim, we are assuming that the extent of translation ambiguity is represented by the text of the first claim.

Our measure of translation ambiguity (${T}_{i}$) is based on the degree of word senses such as polysemy and homonymy (Jurafsky & Martin, 2018). This degree is measured by the proportion of words with multiple senses, in the text of the first patent claim (See Eq. (2)).

$${T}_{i} = \frac{\text{Number\,of\,polysemous\,words\,in\,Patent}\; i\text{s\,first\,claim}}{\text{Number\,of\,words\,in\,Patent} i\text{s\,first\,claim}}$$

(A.1)

This measure, defined in Eq. (2), is inspired by research in computational linguistic literature focusing on measuring translation difficulty, which has shown that sense disambiguity presents a serious problem in translations (Angelelli, 2009; Hale & Campbell, 2002; Secară, 2005; Sun & Shreve, 2014). As argued by Baker (1992) and more recent applied studies such as Wildsmith-Cromarty (2008), Ulanska (2015) and Polyakova et al. (2018), the efforts to ensure equivalence in translation beyond word level, which requires the understanding of single words in association with other words in a collocational range, are often faced with problems arising from the translation of polysemy words. Inadequate proficiency in distinguishing subtle differences in the set of relevant lexical items in the source language may hinder the translators in selecting words in the target language which convey the true sense that needs to be conveyed. Mishra et al. (2013), for example, found a positive relationship between linguistic features of a sentence to be translated including "average of words polysemy" on the difficulty of a sentence to be translated based on human translator experimental data.

In our implementation, we identify polysemous words using WordNet, an extensive resource available online in various languages to identify the many different word senses or meanings of a word or lemma (Miller, 1995).¹⁹ For example, from WordNet we can identify that the word ‘bank’ can mean a type of financial institution or a sloping river side (Jurafsky & Martin, 2018). WordNet is a manually constructed lexical database developed by George Miller at the Cognitive Science Laboratory at Princeton University. WordNet can be considered as a combination of dictionary and thesaurus. WordNet includes different sets of cognitive synonyms named synsets. These sets are made of nouns, verbs, adjectives, and adverbs, each representing a distinct concept. Synonym sets are interlinked based on conceptual-semantic and lexical relations. Interlinking is also by considering specific senses of words. Therefore, words that are in close vicinity to each other in the network are semantically disambiguated. In addition, WordNet labels the semantic relations among words, whereas the groupings of words in a thesaurus do not follow any explicit pattern other than meaning similarity. Each of WordNet’s 117,000 synsets is linked to other synsets by means of a small number of “conceptual relations”. Word forms with different meanings are demonstrated in as many distinct synsets. This means each form-meaning pair in WordNet is unique (Miller, 1995).

Validation

We still need to confirm that the translation ambiguity (${T}_{i}$) measure is positively correlated with translation difficulty, which is in turn negatively correlated with a measure of translation quality. Unfortunately, to our knowledge, there are no data on translation quality of patent applications available anywhere. Hence, the evidence we provide is based on non-patent related data. Our implicit assumption in doing so is that the negative correlation between source text translation ambiguity and target text translation quality does not vary systematically over the types of source text. In particular, the source text sample in the data we use for the validation analysis consists of single sentences extracted from English news articles. Even if we were to believe that the type of the source matters, given that the interpretation of patent claim text is potentially more complex than news text, our validation is more likely to underestimate the strength of the correlation between the measured source text’s translation ambiguity (${T}_{i}$) and the quality of the translation.

Specifically, the sample source text we used for the validation analysis described below comes from Araghi and Palangkaraya (2019), who constructed the sample from publicly available machine translation evaluation data of the annual Workshops on Machine Translation (WMT) over the period of 2017–2019.²⁰ In their analysis, Araghi and Palangkaraya (2019) investigate the correlation between three measures of source text translation difficulty (degree of polysemy, length, and structural complexity) and the quality of machine translation of the source text as judged by human evaluators. Based on all samples of English source texts machine translated into 11 different languages (Chinese, Czech, Estonian, Finnish, German, Gujarati, Kazakh, Latvian, Lithuanian, Russian, and Turkish), they confirmed Mishra et al. (2013)’s finding in very different settings that translation difficulty features are systematically related to the quality of translation.

As a validation to our measure of translation ambiguity (${T}_{i}$) in Eq. (3), we estimate the following linear regression:

$${H}_{i}={\beta }_{o}+{\beta }_{1}{T}_{i}+{\varepsilon }_{i}$$

(3)

where $i$ indexes the source text in the sample as the unit of observation, $H$ is a translation quality score of 0–100 assigned by a human evaluator who observed the source text, the translation, and the reference translation, and $\varepsilon$ is a normally independently identically distributed random error. The estimating sample consists of 6742 English source texts and their corresponding Chinese machine translations from the 2017, 2018, and 2019 WMT workshops. We also include workshop year dummy variables during estimation to control for any year-to-year variation in the sample.²¹ The estimated coefficients of Eq. (3) are provided in Table 8 below. As can be seen, translation ambiguity (${T}_{i}$) is negatively associated with translation quality. For example, from column 2, a 1-unit increase (i.e., 100%) increase in translation ambiguity is associated with 5.5 points lower translation quality or approximately 7% lower than the average. Since this estimate is based on a sample of relatively simple English news text, we conjecture that the strength of the relationship is likely to be higher for the case of patent claims (Table 9).

Table 8 The relationship between translation quality and translation ambiguity (${T}_{i}$)

Table 9 Interaction between translation ambiguity and local inventor presence and separate office estimates, logit model coefficient estimates