Machine Translation Discussed With Young People

The Perception Young People Have of Machine Translation

At different moments during the presentation, the young people were asked certain questions: Do you use machine translation? What use do you make of it? Which translators do you think use it? Do you know how machine translation works? Have you been satisfied with the results? How do you see the future of machine translation?

Practically all of the young people responded in the affirmative to the first question, although the frequency of such use oscillated between “sometimes” and “always.” The majority affirmed that they had used MT to translate texts that they did not understand, or to be surer of their content (class materials, textbooks, and exercises), or else for song lyrics, emails, Facebook messages, and so on. Numerous students added that they use such translators as dictionaries, because they are “more practical and faster” than using online dictionaries.

The young people at the Fair declared that they had used both direct translation (into their mother tongue) and inverse translation (into a foreign-language) with equal frequency. In fact, some students explained that given the “shortcomings” or “faults” obtained, they often carried out both operations on the same text (e.g., English → Catalan and then Catalan → English) as a system for testing accuracy.

The pairs of languages most used were English-Catalan and English-Spanish and vice versa. Some way behind were French-Catalan or Spanish and then Arab-Catalan or Spanish. Minority uses included German-Catalan or Spanish and Italian-Catalan or Spanish. The greater use of English is due to the fact that it is currently the language that is offered and demanded in the majority of educational centers outside universities.

The vast majority of students knew about and used Google Translate. Although a few did mention machine translators offered by generalist newspapers or other software online.

Despite their daily use of MTs, the students had little idea of how they worked. Some “imagined” that they are fitted with grammar rules and enormous dictionaries with “all the words in alphabetical order.” None of them had ever heard of statistical translation, so we explained it to them during the activity.

Responses to questions about the degree of satisfaction on the one hand and the future of MT on the other were probably the most interesting.

The great majority of young people told us that machine translators did not work well and they particularly noted their “lack of naturalness” and/or “coherence” in many phrases along with errors in verb, article, and proper nouns use. In other words, in general, they are conscious of the shortcomings and clearly identify errors. Despite this, they consider such translators as important tools and are convinced that “probably” or “surely” in the future all translation will be MT with the exception of “some literary works.” In this latter case, they noted that it was necessary to take into account “the context of the work.” It can be said that in general students prioritized the advantages of MT (speed, ease of use, multiplicity of languages) over the inconveniences. To improve the results they took into account preediting and postediting.

In short, these young people demonstrated considerable confidence that future research would “get over the problems” and “limitations” that MT currently presents. For the boys and girls who came to our stand, MT is not a myth but rather an everyday reality.

Simultaneously with this dialogue with the students, a practical session divulging the reality took place.

Ways to Explain Machine Translation and Postediting

The content of the posters, drawn up on the basis of a corporate template and following suggestions about content, had to be clear, didactic, and enjoyable.

The first poster dealt with MT as a communication experience related with HT (Human Translation) and other translation processes. The working of SMT was also explained. It was stressed that in human translation the translator has to understand the text in order to be able to translate it, while in SMT, the system “translates” without understanding what it is translating. In the latter, searches are made for the most frequent equivalences in documents introduced into databases, many of these being parallel texts (existing translations from one language into another). The automatic system statistically calculates the probabilities that a particular combination of words will be translated in a particular way into the other language. The SMT system used in our lab is Moses, an open-source project started in 2005 by Philipp Koehn and Hieu Hoang, and later EU funded, which incorporates contributions from many sources. For any language pair, translation models are trained automatically. Once we have a trained model, a search algorithm quickly finds the highest probability translation among the exponential number of choices. As is clearly explained on the Moses website, the two main components in Moses are the training pipeline and the decoder, though there are other elements. In the training pipeline, raw data (parallel and monolingual) are selected and turned into an MT model. Raw data are then tokenized and word-aligned. Word alignments are used to extract phrase–phrase translations and statistics are used to estimate probabilities. The decoder is a single C++ application that will translate the source sentence into the target language. Finally, in the tuning step the different statistical models used are weighted against each other to produce the best possible translations. These technical aspects were summarized to the students without going deeply into questions affecting SMT research such as referred by Turchi, De Bie, and Cristianini (2008):

How far can this representation take us towards the target of achieving human-quality translations? Are the current limitations due to the approximation error of this representation, or to a lack of sufficient training data? How much space for improvement is there, given new data or new statistical estimation methods or given different models with different complexities?

In the second poster, the activities of the research group were described along with the challenges of postediting machine translations. That is to say editing and revising text carried out by professional human posteditors on material churned out by an MT.

To exemplify this, we showed visitors an erroneous headline from a bilingual edition of a newspaper that was translated using MT and postediting. We explained to them that the example contained two types of problems: mistakes by the machine translator, unable to resolve semantic ambiguities, and negligence by the human posteditor, unable to detect the MT mistakes.

We also showed the students phrases from more distant languages such as Chinese or Russian, translated by Google Translate, into Spanish. In deciding which fragments were better translated, we could infer which were the more frequently recurring phrases within the texts stored in the translator’s database.

The third poster summarized the main linguistic obstacles to MT (polysemy, idiomatic expressions, terminology, etc.). It also proposed future lines of research such as increasing the presence of minority and less translated languages on the web, MT into many languages simultaneously, translating from audio, and so on.

Hands-On

The practical part that we had prepared involved four activities that we deemed appropriate for different profiles of students (age, interest, attention span, etc.).

The first was a 2-minute competition detecting linguistic errors and postediting problems in MT. This was to make students aware of the need for postediting in MT and to help them understand the responsibility of the translator in the final result. The examples, chosen from real-world texts, often made the audience laugh. The competition model we chose pleased the students because it appealed to their competitiveness.

The second activity consisted of noting eloquent examples of the dangers of a bad translation. Among others, we demonstrated three possible translations of a single medical prospectus.

In the third activity, lasting 5 minutes, we showed the translation of a fragment of a famous pop song. Side-by-side with the original chorus, in English, was the MT of it and the official version in Spanish, which was an adaptation that rhymed. The need for human intervention in adapting the latter was obvious. Students also appreciated the use of MT systems such as that known as human-aided machine translation, where the output of MT systems can save time and money by providing draft translations that are then postedited for publication (Hutchins, 2010).

The fourth activity consisted of demonstrating how we can all improve MT databases if the system allows feedback. In this way, users become active participants in improving the MT system. For instance, by sending correct postedited translations to Google Translate through the instruction “Not Correct?—Send.”

By the end of the experience, the students learnt how SMT really works, understood the value of human intervention, tested their SMT postediting skills, and learnt how to improve an MT system.