Abstract:
Artificial Intelligence (AI for short) and related topics, such as figurative language and humour research in contribution to translation machine evolution are very important topics in translation studies. Therefore, in this paper I would like to propose a discussion about the relationship between translation machines and translation studies. It is important to acknowledge that AI and translation machines have improved significantly during the last two decades and are still improving as of today. However, despite being helpful for translators, it is important to keep track of AI’s flaws and to not trust translation machines blindly. My paper is mostly based on the translation studies research done by Klaudy (1994) and the research conducted in the field of AI translation machines by Barnden (2008), Farahsani et al. (2021), Gwóźdź (2019), Li et al. (2021), Nakamura (2009), Ritchie (2009) and Ting et al. (2019), respectively.
Key words: AI, translation, Google Translate
1. The Perspective of Translation Studies:
First of all, I would like to introduce this problem from the perspective of translation studies: translating a text, even a basic one, proposes multiple obstacles that require context, culture and cognition to be dealt with. Meta-phrase – meaning word-to-word or literal translation – in most cases is not enough. However, according to Nakamura (2009, 39) and Farahsani et al. (2021, 433) most translation engines[1] at the moment work on the basis of databases – choosing the appropriate form by using pre-written algorithms that use input and other different factors –, which makes the engine work similarly as meta-phrase with extra steps. As a result, most of the time, AI translated texts need a human translator to edit them and to correct mistakes made by the engine. Klaudy (1994, 33-34, 162) argues that sometimes translators need to make compromises (based on lexical and grammatical principles discussed in her book), because translating certain elements from one language to another is not always possible or straightforward and this process requires context, culture and cognition to be solved (Klaudy 1994, 66-79), which the AI lacks and as Farahsani et al. (2021) found in their research, in this case, leads to total omission (without further compensation) – just as in its case with the understanding and production of figurative language and humour. As argued by Nakamura (2009), omission is the most notable transfer operation, in the case of AI translation. However, there are other lexical and grammatical transfer operations that the AI cannot yet utilize, such as sentence separation or fusion for better readability. In summary, omission in some cases is inevitable in translations, however, while a translator possesses the ability to compensate with different tools or with creativity further in the text, the AI lacks this ability and can only produce output based on strictly the input, the database and the pre-written algorithms.
To summarise the academic discussion of translation studies, both Klaudy (1994) and Cui (2012) argue that one of the most important aspects in translation is the solving of the problem of language barriers, which can prove difficult for many reasons. In detail, language barrier is when people try to describe something in a given language by choosing a word (signifier) that explains everything within the context and the culture. As a result, this chosen signifier tells the receiver exactly what the sender means, but that signifier in another language may not exist at all. To overcome this problem, the translator sometimes needs to add an explanation or use other transfer operations in order to solve the question of the language barrier.Currently, mostly because of the wide access to the internet and translation machines; translators have a lot of tools available to help them overcome these barriers by leading them to an understanding of the various cultural differences.
2. The Relationship between Language and AI:
Second, according to Nakamura (2009, 39) – who is a Doctor of Engineering, involved in speech and language processing, with a focus on speech translation and speech recognition – a translation machine or software most often uses the word-to-word translation method to simply override this problem. Often leaning on the users to do the required check in order to get the proper meaning. However, a software that is capable of learning – either through deep learning (which enables the AI to expand its database) or by users or admins adding exceptions and new phrases. In the case of deep learning, the translation AI is analysing conversations and other databases, which can result in more accurate translations. However, the AI is only capable of working with pre-written core algorithms and as of yet unable to successfully write new algorithms to solve unforeseen problems (Nakamura 2009, 38) – therefore, the AI can only create what it is programmed to create or for what it has an already created schematic, unlike humans, who possess creativity. This process is also supported by the results of the recent research conducted by Farahsani et al. (2021).
Therefore, this language barrier presents a problem in the field of computer science too, and as of now; there is no perfect method to solve this. In software development and artificial intelligence research on translation machines, the goal is to make translator engines, which are able to learn and adapt to these situations (Nakamura 2009, 38-39). Based on the research of Barnden (2008), Gwóźdź (2019) and Ritchie (2009) respectively, this requires an AI that can understand figurative language and humour (in both the source and target language) in order to make decisions when forming the text in the target language and to teach AI “creativity” or to program an algorithm that works similarly to human cognition in order to produce output without relying too heavily on the input.
On the other hand, as Kirkpatrick (2020, 16) argues advances are being made and translator machines are already used in second language acquisition as well. One of the greatest advances, as Kirkpatrick (2020, 17) argues is that AI is able to read and interpret texts from pictures as well, not only from strings[2]. In addition, the research conducted by Ting et al. (2019, e9) found that AI is more efficient in the case of analysing data and thanks to deep learning works more accurately in achieving better diagnostics than human graders.
3. The Silent Aspect of Culture
Third, it is also important to underline that culture has a great impact on language. Certain signifiers have a deeper and more complex metaphorical meaning behind them, which the same signifier in a different language does not possess. This derives from the cultural differences between the source and the target language (Cui 2012).According to Daniel Everett (2019), different cultures through their language have different perceptions about the world around them, based on their lifestyle and what carries greater importance in their culture. He studied the Piraha language within the Amazonian Jungles, and based on his findings he concluded that language is formed by the culture and culture is formed by the language (in the case of metaphors: culture acts as a filter (Littlemore 2003)). The Pirahas, therefore, have multiple words for different trees, because those are more important to them, while they cannot understand why Dan Everett has different “names” attached to his different books (Everett 2019).
Franz Uri Boas (1941) established the idea of cultural relativism. According to Boas (1941), cultures cannot be objectively ranked as higher or lower, or better or more correct, because all humans see the world through the lens of their own culture, and judge it according to their own culturally acquired norms. Through this idea, culture shapes language, and language shapes culture (Boas 1941). However, when programmers create algorithms for translation engines, it is inevitable that in the question of cultural norms the programmers will rely on their own culture and background and while they try to include schemas for different syntaxes, as of yet they cannot write any algorithm for cultural perception based on human cognition. And as Li et al. (2021, 2-3) argue that the connection between human cognition and culture is the key to perceiving a message, which is also supported by the findings of Barnden (2008), who research metaphors in this respect.
Similarly, a great analogue to cultural perception is the example of understanding and producing humour – simply, without cognition, culture and context, AI is not able to produce real humour. However, it is one of the most important research fields in this regard, because if an algorithm can be written for humour perception and production, then AI will become able to possess a fraction of the human cognition (Ritchie 2009, 78-80). In addition, as Barnden (2008, 334) argues, metaphor research regarding AI is also very important in order to model human cognition.
The thesis of cultural relativism by Boas (1941) is an important aspect of translation studies and an important aspect in the algorithm that makes a decision. However, in practice, the opposite of this thesis is used. For example, the ordering and ranking of grammatical, lexical and cultural elements of the database are influenced and written to work based on the belief of its creator(s), the same is true for the algorithm. For instance, the personal language attitude (of the creator(s)) towards variation in a language is represented in the AI. Although, Kirkpatrick (2020) argues in favour of the AI’s ability to perceive cultural elements and to solve the problem of the language barrier, because in the case of closely related cultures and languages due to deep learning translation accuracy and problem-solving in this regard improved during the last two decades.
4. Human Cognition:
Fourth, the problem of the way of thinking: among many issues, this can cause the problem of missing the main point in a conversation or in mental calquing. The syntax of the sentences, the form of the text and the reason behind chosen words in relation to others are only understood by the AI in limited ways, based on having an algorithm to analyse it or a schema to work with it or not. In case of a mispronunciation and variation in language (except for specialised translation machines (Kirkpatrick 2020, 16)) or errors in the text, the AI often misinterprets and mistranslates. In most cases, the AI has a limited way of correcting lesser errors, such as mistypes and minor syntax errors. These issues are present in both translation studies and second language acquisition as well. For instance, for non-native speakers, it can be difficult to determine whether the text contained irony or other figurative language or not. Also, a professional language user may opt to correct the mistakes of the original text. For example, Google Translate is advancing in this regard, because in many cases it can work with context (and may correct errors based on context) and due to deep learning it can understand more and more already (Caswell et al. 2020).
Moreover, on the basis of the language barrier connected with the way of thinking, both Cui (2012) and Klaudy (1994) argue that often translators have to make difficult choices, because it can happen that certain jokes, phrases or ironies cannot be translated to a given language – it will remain impossible because the thinking of the native language users is different in some aspects and the form, the syntax or other elements of the target language are missing or just inherently different. Therefore, the point will get lost in translation. For example, puns that rely on language perception or audio-visual contexts.
According to Cui’s research (2012, 826), wordplay and humour are the most difficult to translate, because their meaning is strongly affected by the original language, resulting often in untranslatable signifiers. Cui (2012, 827-30) compares Chinese with English to demonstrate how certain elements of the language cannot be translated in a given context. The grammatical system also plays a great role in these situations, because word order or a different alphabet can easily be the sole origin of the meaning. Those a translator can only convey through explanations or omit fully and compensate in very different ways. For instance, a translator might be able to make a new wordplay to convey the meaning of the original, but in some contexts it is impossible, even by explanations through additional sentences or words. However, this method can easily lead to ruining the joke or the pun (Cui 2012, 829). Therefore, the translator has to make compromises and even let some things go and compensate for the loss in a different part of the text. Moreover, as I have previously established based on the research of Nakamura (2009) and Farahsani et al. (2021), AI in these cases operates with total omission, mainly because it lacks human creativity to compensate.
5. Distance between Languages:
Fifth, an important aspect of AI translation is the relationship between languages and the distance of language families. For instance, in the case of English to German, Polish to Czech or vice-versa, Google Translate works with an over 90 per cent accuracy. The reason for high accuracy rates in these cases is that these languages are closely related (Farahsani et al. 2021, 433), which means that closely related languages have similar syntax and even lexicon, among other similarities. Also, it is important to mention that in the case of English or Spanish there is a high input rate due to their current status in the intercultural space. On the other hand, for instance, due to the gendered nature of Polish, some context is lost in the translation process when translating from Polish to English. Such as the words “karczmarz” and “karczmareczka”, both mean “innkeeper” in English, but in Polish “karczmareczka” is the female innkeeper and this context often gets lost in AI translation due to the fact that the database assigns both Polish words with the same English one. Or another example is when the speaker in English is talking to a female listener and, because English is not heavily gendered, the words used in this conversation would be the same in the case of a male listener as well; however, in Polish, the verbs used in the case of a female listener would receive the proper female gender particle (‘-a’), which indicate the direction of the given act and vice-versa (Hellinger et al. 2015,12 & 100). (This is similar in Portuguese as well.) In this case, when translating from English to Polish (or Portuguese) the translator machine usually picks one or the other form (usually the male or in other cases the gender-neutral form), because most of the time it is the default case in the algorithm (Nakamura 2009, 38-39).
6. Accuracy of Google Translate as an Example:
Therefore, texts containing or working with fixed registers or texts that do not rely heavily on context are more accurately translated by translation machines. For example, legal documents work with fixed syntax and fixed terminological registers, therefore resulting in a more accurate (nearly completely correct) translation. The engine can easily select the correct pair and make decisions. However, texts that rely on context more or texts with a high factor of uncertainty are often translated inaccurately by the AI – such as literature, figurative language, et cetera. It is important to mention that if a translation engine is programmed specifically for literature translation it can function relatively well and may produce a readable outcome, but it is heavily affected by and relies on the interpretation and the preferences of the programmers creating such engines and as of today they are incapable of working independently (Seligman 2019).
Moreover, the research of Farahsani et al. (2021, 427) focused directly on the accuracy rate of Google Translate regarding English to Indonesian translation. This further supports the above-mentioned notion that distance and relationship between different languages that are part of different language trees are important factors in the case of AI translation. They used four inaccuracy indicators in order to evaluate the results – omission, addition, different meaning and zero meaning[3]. Regarding the translation of the word level, Google Translate achieved a 70.73 per cent accuracy rate, however, regarding the phrase level the translation only achieved 46.37 per cent (Farahsani et al. 2021, 427). In addition, a high accuracy rate regarding the word level was achieved due to the number of loan words from English in most technical terminology (one of their test texts was a mechanical engineering paper). In some cases, the translations only achieved a 34.48 per cent of accuracy rate, especially at the phrase level (Farahsani et al. 2021, 430-431). The most important finding of Farahsani et al. was that Google Translate used the sentence structure of the source text and only the general meaning of words (2021, 433).
The translation result of the source text does not have significant meaning. Google Translate translates the sentence based on its structure. However, the sentence structure in the target text is not used. Therefore, when translating using Google Translate, translators should check the prevalence of the sentence. (Farahsani et al. 2021, 433)
7. Conclusion and Summary:
In conclusion, researching humour and figurative language in order to model human cognition and write (code) AI (programs) based on that knowledge is a very important goal. Great advances have been made during the last twenty years and are being made in this field at this very moment. Humour and figurative language research provide AI with practical benefits that can be used in the automation of translation made by or with AI. Nowadays, these translation engines are getting better and better thanks to the learning ability of the AI that is behind them (Caswell et al. 2020). Therefore, they require less and less attention from humans. However, these engines are far from being perfect (Nakamura 2009 & Farahsani et al. 2021).
The most important problem that artificial intelligence research regarding humour and figurative language presents us with is the aspect of human cognition, the context of the message and the cultural elements that affect human speech production and perception. These are the main obstacles to overcome before AI could correctly perceive and produce humour or figurative language. Also, in practice, these are the most important aspects of translation studies as well (Klaudy 1994). The moment when AI becomes able to correctly perceive a message – or perceive a message as humans do – it is going to be able to produce correct translations as well (Li et al. 2021).
Ultimately, it is still a long way before artificial intelligence researchers can create an AI that can fully translate a text while paying attention to the aspects of context, culture and cognition in order to convey the original intentions. However, in terms of progress, the metaphorical and extended meanings are more accurately translated thanks to the research conducted in humour and figurative language, which contribute a lot to the algorithms that are responsible for speech and text perception, speech and text production, and learning of the AI. If in the future, these obstacles would be solved, AI is going to be able to make the necessary decisions with more accurate or even correct outcomes and become able to translate and interpret in real-time. Finally, I would like to acknowledge that in my paper I did not aim to belittle or deny the advances of AI regarding translation machines.
Reference List:
Barnden, John A. 2008. “Metaphor and Artificial Intelligence: Why They Matter to Each Other.” In Gibbs, Raymond W. Jr. (Ed), The Cambridge Handbook of Metaphor and Thought. Cambridge: Cambridge University Press. 311-339.
Boas, Franz Uri. 1941 “Language and culture” in Studies in the History of Cultures. 178-184.
Caswell, Isaac and Bowen Liang. 2020. Recent Advances in Google Translate. Published online by Google Research on June 8, 2020. Available at: https://ai.googleblog.com/2020/06/recent-advances-in-google-translate.html
Cui, Jingjing. 2012. “Untranslatability and the Method of Compensation” in Theory and Practice in Language Studies Volume 2: N. 4, April 2012. pp. 826-831
Everett, Daniel. 2019. The Story of Language. Canguro English. April 2019. Podcast. Available on: https://canguroenglish.com/podcast/
Farahsani, Yashinta, Ika Puspita Rini and Patria Handung Jaya. 2021. “Google Translate Accuracy in Translating Specialized Language From English to Bahasa Indonesia: A Case Study on Mechanical Engineering Terminology.” Advances in Social Science, Education and Humanities Research, volume 51. Published by Atlantis Press SARL.pp. 427-435.
Gwóźdź, Maja. 2019. “Figurative Language Grounding in Humanoid Robots.” In Bi, Yaxin, Bhatia, Rahul, and Kapoor, Supriya (Eds), Intelligent Systems and Applications. Proceedings of the 2019 Intelligent Systems Conference (IntelliSys), Volume 2. 347-363.
Hellinger, Marlis and Heiko Motschenbacher. 2015. Gender Across Languages: Volume 4. Goethe University Frankfurt & Technical University of Braunschweig.
Kirkpatrick, K. 2020. “Across the language barrier” in Communications of the ACM, 63(3). 15–17. doi:10.1145/3379495
Klaudy, Kinga. 1994. Languages in Translation. Budapest: Scholastica. 2007.
Li, P., Ning, Y., & Fang, H. 2021. “Artificial intelligence translation under the influence of multimedia teaching to study English learning mode.” The International Journal of Electrical Engineering & Education.
Littlemore, J. 2003. “The Effect of Cultural Background on Metaphor Interpretation.” Metaphor and Symbol 18 (4): 273-288.
Nakamura, Satoshi. 2009. “Overcoming the Language Barrier with Speech Translation Technology” in Quarterly Review No.31. April 2009. 35-48.
Ritchie, G.D. 2009. “Can Computers Create Humor?” in AI Mag., 30. 71-81.
Seligman, Mark. 2019. “Extracting the essence.” Translating and Communicating Environmental Cultures. First edition. Routledge. Ebook.
Ting, D. S. W., Lin, H., Ruamviboonsuk, P., Wong, T. Y., & Sim, D. A. 2019. “Artificial intelligence, the internet of things, and virtual clinics: ophthalmology at the digital translation forefront.” The Lancet Digital Health, 2(1), e8–e9. DOI:10.1016/s2589-7500(19)30217-1
[1] In IT (or computer science) the core of any program (including AI) is called the engine that contains the database and the main algorithms. Throughout the paper, I will use the term ‘engine’ when referring to the core mechanisms of a translation machine.
[2] In most programming languages ’string’ is the (basic) variable that contains letters, words or sentences; therefore, it is a great leap for the AI to be able to recognize words from pictures, which are files stored with complex methods.
[3] Zero meaning occurs, when a translation seems syntactically correct, however, it is meaningless.
Written for the University of Szeged — Institute of English and American Studies — For Science’s Sake! Popular Topics in Contemporary Linguistics seminar in May 2022.