Lost in machine translation: contextual linguistic uncertainty

Accuracy testing of main machine translation approaches. In order to examine the different types of machine translation methods, we used a number of phrases from Russian, English, Greek, Malay, and Japanese. The result was that Google (NMT) and Yandex (HMT) translation services showed the highest degree of accuracy, compared to PROMT (RBMT). However, PROMT had several results better than Google and Yandex.

Below are the samples of the most illustrative results of our online machine translation tests (see Tables 1-5).

These online translation examples illustrate the high degree of literal translation persistent in machine translation systems. When the machine translation is juxtaposed to human translation, it can be seen that the online translation procedure lacks conceptual meaning even though the semantic and syntactic systems are integrated into it the online translation procedure. Surprisingly, our testing indicated that all translation systems encounter problems with the recognition of both full and shortened names. In Table 4 Timah, a shortened name, was translated as `tin' and Greek name Егрцш) (Irene in English) also was not recognized as a name and got translated literally by its meaning (`peace') in another tested phrase va sixs Sikio n Eip^vn xskiKaf».

In most cases, idiomatic phrases were not detected and were translated literally by the systems (see Tables 1, 2). Also, this pair of examples demonstrates that the accuracy of a translation depends on the syntactical complexity of the source phrase. Whereas the outcomes in the first example basically maintain minimum readability, those in the second example are broken in terms of the sentence structure. All three systems are obviously weak in translating adverbial syntax; the adverbial clause straight up seems to make them particularly confused and this results in poor performance. Likewise, additional information in the fourth example «with her fair skin» and simple modifier today in Table 5 respectively cause the same type of mistranslation. Colloquialisms are also misrepresented: the phrase «Geram melihat anak comel ini» should convey a feeling of affection when seeing the child, and means `What a cute child!'. However, with Google Translate it loses its meaning in the given context, becoming `Greedy saw this cute kid'.

The ambiguity problems, which can be easily resolved by a human, largely contribute to wrong output, as is the case with avvno^ovtfoia: the Greek word describing the feelings of impatience and excitement caused by an unknown situation (Table 3). In the same example, the translation systems fail to identify the gender of the subject and even change the verb into an imperative form. Cases regarding Japanese are even more complicated: the third-person singular pronoun is replaced by the first-person plural we in one case, and in the other case, the subject is dropped as seen in the participial construction. This instance suggests that ambiguity of the action can result in the misplacement of the verb causing a wrong form as well as a fragmentary phrase never aligned in a complete sentence.

As mentioned above, the Malay language is full of idiomatic expressions which reflect the various cultural aspects of the language. This feature of the language has led to many inconsistencies in machine translations. As has been shown by previous research in an accuracy analysis during the translation of 200 Malay sentences containing proverbs into English, more than half (55.0%) of them were wrongly translated by Google Translate, and 34.0% were correct while only 9.6% were translated accurately into similar idioms [Abd Rahman, 2013]. The challenges encountered during machine translation were mostly rooted in the use of affixes in words and the additional stopwords in phrases during translation (both of which were used to reflect the grammatical structure of the language), as well as the use of different words with the same meaning [Abd Rahman, 2013]. In the first issue, the example memilih kasih can be translated when the affixes are removed: «pilih kasih». Hence, stemming, which is the detection and filtering of the proverb to exclude affixes, needs to be done before translation [Abd Rahman, 2013]. Secondly, proverbs may have stopwords such as in the following phrases: «sedikit-sedikit lama-lama jadi bukit» and «sedikit-sedikit lama-lama akan jadi bukit». Hence, the removal of superfluous words such as akan would enhance accuracy of the translation [Kwee, Tsai, Tang, 2009; Abd Rahman, 2013]. Thirdly, there may be different words used to represent the same proverbial meaning - «Ada angin, ada pokoknya» is similar to «Ada angin, ada pohonnya» (meaning anything that happen has a cause) [Abd Rahman, 2013]. In the case «bagai kera mendapat bunga» for someone who does not appreciate the value of a gift, beruk and monyet can replace kera to mean the same thing.

Contemporary machine translations have yet to solve the above-mentioned problems [Yusoff, Jamaludin, Yusoff, 2016]. As language does not exist in isolation but is part of a society and culture, one would need to be familiar with Malay to be able to translate the rich and colourful cultural contexts of the language. Nevertheless, there are studies of Semantic-based Translation using N-Grams that could deal with ambiguous sentences by identifying words with multiple (ambiguous) meanings [Yusoff, Jamaludin, Yusoff, 2016].

Therefore, our analyses and results of previous works in this field show that for machine translation to be effective, it first needs to incorporate three levels of meanings processing: the literal, the conceptual, and contextual. Secondly, a corpus of metaphors, idiomatic phrases and proverbs with equivalences from different cultures need to be constructed. Finally, the recognition of proper names and their shortened versions should be improved.

Machine translation and linguistic relativity theory

The act of translating from one language to another, apart from being a fairly complex problem when implemented by machines, can pose difficulties even for a human translator. A good example of this fact is the book «English As She Is Spoke: the new guide of the conversation in Portuguese and English» by Pedro Carolino [Da Fonseca, Carolino, 2002] which is full of grammatical and stylistic mistakes that are surprisingly similar to the ones made by the machine translation systems.

The linguistic relativity theory and theories similar to it explain some of these translational difficulties. They show that many languages differ in the amount of words they have, some words describe unique feelings, a person's characteristics, professional jargon, and many other realities specific to a culture [Whorf, 1956; Kovecses, 2005; Deutscher, 2010]. The surrounding environment, natural resources, and specific activities of a region also bootstrap vocabularies or slangs that have no equivalents in other languages [Wierzbicka, 1992; Durdureanu, 2011; Sanders, 2014]. Because of this cultural and geographical specificity, many of them can be translated only with the help of a contextual explanation, rather than with a distinct word.

A good example is a word sevdah, whose root comes from Turkish language, and is commonly used by people living in Bosnia. The standardly offered translations - «melancholy», «lovesickness», «yearning for love» - do not really capture the essence of sevdah. A more precise translation would be `enjoying your state of sorrow as a very special (sorrow-ish) kind of «pleasure» '. Perhaps, the English term «wallowing in your sorrow» would not be a bad way to understand sevdah but only because a better concept does not exist. This specific kind of «emotion» simply seems to be «reserved» for people from the Balkans who would find sevdah-kind of happiness in singing songs describing and glorifying the heroic death of their most loved ones. A possible explanation is that the people from the Balkans in the most difficult times in which they had little to look forward to simply evolved to learn how to enjoy in their sorrow. Today, mainly due to the changed historic circumstances, even many young people from the Balkans would struggle to understand word sevdah itself and the state of being in sevdah.

Such types of words do not have their counterparts in other languages and this kind of untranslatability very often leads to their borrowing: we adopt words from one language (the donor language) and incorporate them into another language without or with minor modifications. Sometimes this happens on a critical cultural level because borrowings tend to overflood national languages around the globe (e.g. economic and computer terms such as market, poster, billboard, slogan, hashtag, etc.). In France, this has even caused cultural resistance [Styblo, 2007; Caruso, 2012]. However, the question whether simply introducing a foreign word into a language entails introducing the relevant concept in that language remains discussable. Suppose we incorporate word sevdah into English in a way in which we can incorporate poster into Serbo-Croatian: it is unlikely that the former borrowing will yield the same result as the latter.

Difficulties in translations of unique words also could be overcome by «adaptation» or «free translation» wherein the social or cultural reality (idea) in the source language is replaced with new realities that are closer and more natural to the audience in the target language. Such «domestication» [Lawrence, 1995] of the source text is very artistic and vulnerable to criticism, and for the moment cannot be implemented by machine translation because of its creative complexity. The opposite side in the art of translation - «foreignization» - strives to save the source language and culture, and translate text into the target language with minimal changes using, for instance, comments and explanations about original realities. However, many researchers, as in the case of verbally expressed humour based on wordplay, agree that such methods are sometimes ridiculous because additional comments and explanations destroy the amusement [Low, 2011; Hoffman, 2012]. Humour comprehension requires implicit and explicit knowledge of specific cultural and linguistic realities, and their explanations could be too long or inappropriate for translation. For example, the joke «A priest, a rabbi, and a nun all walk into a bar, and the bartender says, `What is this, some kind of joke?'» requires knowledge about jokes that begin with «A, B and C walk into a bar…».

In this regard, we can see that the problem «how to translate» is a unique and disputable task that can be solved differently by different translators and with a variety of methods. This complexity / relativity of languages, cultures and methods makes current machine translation systems just an auxiliary tool. However, the more successful, socially accepted or standardised examples of human translations we will have, the more data could be borrowed, formalised and used by specialists and technical systems for «neutral translation» [Razlogova, 2017]. Thus, despite variabilities in the rendering of the same text, some formal invariants or typical examples can be extracted and be practically used in the machine and human translation.

Lost in Internet translations

The on-going technology boom has further created an additional problem in terms of language translation. Internet slangs, acronyms, hidden meanings, letter and number combinations in words and intentional mistakes are, generally, an accepted way of communicating online.

Additionally, foreign languages are heavily influenced by English in this field with many hybrid words being coined as a result of mixing two languages together. For instance, incorporating English in German to create Denglish, or English in Malay for Manglish, or writing Greek characters in English (Greeklish) as a means to not constantly switch between keyboard layouts. In this case, translations need to consider processing the literal, the conceptual, and the content/context meanings, referring to contextological dictionaries for styles and idiolects of the users.

Combined with the many abbreviations used online, it can be hard for speakers who are learning the language to understand these foreign compressed messages even if they might understand colloquial speech. For instance, the numbers 55 sounds as `go go' in Japanese and are used to convey the English meaning. In Malaysia, fuyoh is commonly used online and may be equivalent to the OMG in the internet slang. An example in Manglish: «Fuyoh, so cheap!»

The difficulty and novelty of this language style or idiolect is high enough that a reader outside a given network community and culture would be frustrated to understand anything. Even YouTube comments are becoming more and more outlandish for «strangers». These are the reasons why the origins of «netspeak» and «digital natives» - competent communicators in cyber contexts - are being postulated by researchers in modern culture or in so-called «generation Z» [Crystal, 2001; Pasfield-Neofitou, 2012; Sharifian, 2017, p. 108].

Therefore, «netspeak» is a refined example of an untranslatable language system via machine translation processes because the online translation tool requires to be trained to recognise style, deliberate typos, abbreviations and acronyms. In addition, the community-specific vocabulary predominantly depends on the contexts within which it is utilised. This social reality in interlocutor exchange situations directly affects the translation programs. It is essential to have machine training based on corpora that include words, concept, and content. However, such a problem was partly solved by emergence of a new branch in translation called «crowdsourced translation». Its two main forms are: 1) nonprofessional community-based systems such as the Google Translate community that corrects automated translations or Luis von Ahn's «Duolingo» language approach that uses a learning platform where people translate websites as a part of the learning process and 2) crowdsourced translation service platforms such as TM-Town, Gengo, Smartling and others with professional translators providing their services (https://www. morningtrans. com/crowdsourced-translation - does-it-work/).

Because of the international boom of social networks and the need for translating usergenerated content with its slang variations and informal language, Facebook and Twitter have also launched crowdsourced translation platforms in order to create multilingual posts. It must be noted, however, that crowdsource methods do not restrict themselves to translation problems only. For instance, project reCAPTCHA uses the input words or selected images to bring old books to the digital realm and to gather data for artificial intelligence, improving the accuracy of maps. This project is realised in expecting users to decipher distorted words or to identify particular pictures online to reach a successful registration. Crowdsourced translation for verification of machine translations may be the solution to accurate translation for languages in which words and phrases are heavily reliant on culture and context, such as in the Malay language. However, the difficulty may be in getting a sufficiently dedicated and sufficiently informed crowd to contribute to the translations.

Conclusion

In retrospect, looking at the problems discussed in previous works [Kotov Marchuk, Nelyubin, 1983; Novozhilova, 2014; Arestova, 2015; Dulov, Shmeleva, Boronkinova, 2017], we still see that:

1) At the present stage and in the near future, it is impossible to exclude the human editor from machine translation. The latest and most advanced Neural Machine Translation so far needs human corrections even with simple tasks [Marchuk, 2016; Nguyen, Chiang, 2017]. Therefore, linguists and other scholars still have to contribute to the further development of such systems, for example, through the construction of metaphors corpora and contextological dictionaries that could be used for translation (interpretation) of the most difficult literary texts.

For the further development of machine translation, it is also crucial to include the various pragmatics aspects of language into the process of computer-based translation. For the moment, «crowdsourced translation» could be a solution of that problem. It may provide resources for the current translation issue of fast-growing «netspeak». An attempt was also made by a recent project called «SenseTrans», a tool that adds contextual information to posts in social media using Al-analytics [Lim, 2018; Lim Cosley, Fussell, 2018].

The idea of mobilizing a crowd of translators of a variety of texts - both amateurs and professionals - into a sort of collective wisdom for trans-linguistic communications seems also to become a computational materialization of the «translation norms» or «the social reality of correctness notions» [Bartsch, 1987, p. xii]. According to the two pioneer contributors to the theory, G. Toury and Th. Herman, each individual's interaction with the multi-layered socio-cultural norms circumscribing her/his verbal operations informs the sense of accuracy and quality of a translation result [Toury, 1995; Hermans, 1996]. Crowdsourcing processes in tandem with platforms of professional translators able to verify them can potentially construct such a norm that helps make translation machine a usable apparatus.

2) We still do not have an integrated typology of various types and styles of translation. Indeed, translations of technical texts differ from translations of newspaper texts or informal online conversations. Among the systems tested in our research, only PROMT, with some limits, can translate with pre-set writing styles but still cannot detect such styles or eliminate stylistic mistakes. However, some grammar checker software (like Grammarly, StyleWriter, WhiteSmoke, etc.) have addressed this problem, revealing, for instance, colloquialisms in documents written in an academic style.

Many studies show that there are cultural and aesthetic differences between men and women in the use of vocabulary, syntax, and communication [Na, 2016; Okamoto, 2013]. In the Japanese language, «onnarashii hanashikata» (feminine ways of speech) and «otokorashii kotobazukai» (masculine ways of speech) is common. This kind of speech reflects specific forms of politeness and cultural norms represented by language. Although there is no masculine or feminine way of speaking in the Malay language, specific words may be used to describe a feminine or masculine trait. «Hitam manis» is always used for women but for men, they are «berkulit gelap» or dark complexion. In that context, wrong word choice could lead to ridiculous or impolite results in machine translation.

Nowadays, Google Search considers the original location of a search query, as well as the user's previous requests. Probably in the future, machine translation will also become less abstract and universal, and more personalised and situated, by constantly learning from the interests and idiolects of its users.

3) Modern ecolinguistics, the distributed language theory, the dynamic and adaptive systems approaches to language, systemic functional linguistics, and cognitive linguistics have shown that language (or process of «languaging») is dynamic, interactive, situated, ecologically and culturally embedded. All these aspects of language complicate its formalisation and machine translation. However, it may be still possible to find a common, formalizable core («universal grammar») for a group of languages by means of theoretical (linguistic, mathematical) and machine analysis of languages. Furthermore, dynamic, learning and evolving models (software) can be designed by programmers and linguists that could adapt the machine translation to the dynamic nature or the evolvability of natural language and to dialects/idiolects of their users.

References

1. Abd Rahman K., Md Norwawi N., 2013. The Challenges of Handling Proverbs in Malay-English Machine Translation. 14th International Conference on Translation 2013. Penang, University Sains Malaysia, pp. 27-29. URL: www.scribd.com/doc/ 163669571/Khirulnizam-The-Challenges-of - Automated-Detection-and-Translation-of - Malay-Proverb.

2. Arestova A.A., 2015. Sravnitelnyy analiz sistem mashinnogo perevoda [Comparative Analysis of Machine Translation Systems]. Vestnik Volgogradskogo gosudarstvennogo universiteta. Seriya 9, Issledovaniya molodykh uchenykh [Science Journal of Volgograd State University. Young Scientists' Research], no. 13, pp. 105-109. Babhulgaonkar A.R., Bharad S.V., 2017. Statistical Machine Translation. Intelligent Systems and Information Management. 1st International Conference on Intelligent Systems and Information Management (ICISIM), Aurangabad, pp. 62-67.

3. Bartsch R., 1987. Norms ofLanguage: Theoretical and Practical Aspects. London, New York, Longman. 348 p.