Machine Translation (MT) is a technology that automatically translates text using termbases and advanced grammatical, syntactic and semantic analysis techniques.
The idea that computers can translate human languages is as old as computers themselves. The first attempts to build such technology in the 1950s in the USA were accompanied by a lot of enthusiasm and significant funding. However, the first decade of research failed to produce a usable system and the now-famous report by Automatic Language Processing Advisory Committee (ALPAC) in 1966 found that the ten-year-long effort failed to fulfill expectations. The next time the general public heard of MT was likely in the late 1990s when the internet portal AltaVista launched a free online translation service called Babelfish. Although the quality was often lacking, it became immensely popular and brought MT into the limelight again. Other internet giants presented similar services soon after, the most well-known of which is now Google Translate.
Despite great strides in technology and addition of dozens of new language pairs, these free services are usable for “gist” or casual translation, but usually not for commercial purposes. On the other hand, commercial providers of MT technology have worked on improving their paid offerings and with customization such Machine Translation engines are finding commercial use in limited areas. However, challenges with understanding context, tone, language registers and informal expression remain the reason why MT is not expected to replace human translators in the foreseeable future. The main use cases for machine translation are applications that require real-time or near real-time interaction, for assimilating texts and “chat”, and as a productivity tool supporting human translators.
Machine translation is not to be confused with Computer-Aided Translation (CAT) Tools.
What is MT suitable for?
The most common uses of MT technology are as follows:
Gisting - The results of MT are generally not as good as translations produced by humans, but are useful for understanding roughly what a text says. Such translation may be good enough depending on the purpose and target audience.
MT-human - In some cases, human translators edit machine translation results to produce final translations in what is called post-editing.
Instant need - It can also be used for providing translations of materials that are time-sensitive and which cannot wait for the time required for human translation, such as results from database queries.
Controlled language – For texts written in controlled language, customized MT engines can provide very high-quality translations, for example in translation of patents or technical specification sheets.
High volume - Content producers are generating exponentially increasing volumes of material, and in many cases, human translation is simply not economically or technically feasible.
Pseudotranslation – Localizers can use MT to translate source text to check for internationalization issues in the target languages before committing to professional translation.
Support for human translators – Modern CAT tools allow users to translate source segments with MT. Translators can decide to use the results as they are or edit them manually, which can speed up their work.
Types of Machine Translation
Rule-Based Machine Translation (RBMT)
RBMT, developed several decades ago, was the first practical approach to machine translation. It works by parsing a source sentence to identify words and analyze its structure, and then converting it into the target language based on a manually determined set of rules encoded by linguistic experts. The rules attempt to define correspondences between the structure of the source language and that of the target language.
The advantage of RBMT is that a good engine can translate a wide range of texts without the need for large bilingual corpora, as in statistical machine translation. However, the development of an RBMT system is time-consuming and labor-intensive and may take several years for one language pair. Additionally, human-encoded rules are unable to cover all possible linguistic phenomena and conflicts between existing rules may lead to poor translation quality when facing real-life texts. For example, RBMT engines don’t deal well with slang or metaphorical texts. For this reason, rule-based translation has largely been replaced by statistical machine translation or hybrid systems, though it remains useful for less common language pairs where there are not enough corpora to train an SMT engine.
Statistical Machine Translation (SMT)
SMT works by training the translation engine with a very large volume of bilingual (source texts and their translations) and monolingual corpora. The system looks for statistical correlations between source texts and translations, both for entire segments and for shorter phrases within each segment, building a so-called translation model. It then generates confidence scores for how likely it is that a given source text will map to a translation. The translation engine itself has no notion of rules or grammar. SMT is the core of systems used by Google Translate and Bing Translator, and is the most common form of MT in use today.
The key advantage of statistical machine translation is that it eliminates the need to handcraft a translation engine for each language pair and create linguistic rule sets, as is the case with RBMT. With a large enough collection of texts, you can train a generic translation engine for any language pair and even for a particular industry or domain of expertise. With large and suitable training corpora, SMT usually translates well enough for comprehension. The main disadvantage of statistical machine translation is that it requires very large and well-organized bilingual corpora for each language pair. SMT engines fail when presented with texts that are not similar to material in the training corpora. For example, a translation engine that was trained using technical texts will have a difficult time translating texts written in casual style. Therefore, it is important to train the engine with texts that are similar to the material that will be translated.
Example-Based Machine Translation (EBMT)
In an EBMT system, a sentence is translated by analogy. A number of existing translation pairs of source and target sentences are used as examples. When a new source sentence is to be translated, the examples are retrieved to find similar ones in the source, then the target sentence is generated by imitating the translation of the matched examples. Because the hit rate for long sentences is very low, usually the examples and the source sentence are broken down into small fragments.
This approach may result in high-quality translation when highly similar examples are found. On the contrary, when there is no similar example found, the translation quality may be very low. EBMT has not been widely deployed as a commercial service.
Neural Machine Translation (NMT)
NMT is based on the paradigm of machine learning and is the newest approach to MT. NMT uses neural networks that consist of nodes conceptually modeled after the human brain. The nodes can hold single words, phrases, or longer segments and relate to each other in a web of complex relationships based on bilingual texts used to train the system. The complex and dynamic nature of such networks allows the formation of significantly more educated guesses about the context and therefore the meaning of any word to be translated. NMT systems continuously learn and adjust to provide best output and require a lot of processing power. This is why this approach has only become viable in recent years.
All the above have their shortcomings, and many hybrid MT approaches have been proposed. The two main categories of hybrid systems are:
- rule-based engines using statistical translation for post processing and cleanup,
- statistical systems guided by rule-based engines.
- either of the above with some input from neural machine translation system.
In the first case, the text is translated first by a RBMT engine. This translation is then processed by an SMT engine, which corrects any errors made. In the second case, the RBMT engine does not translate the text but supports the SMT engine by inserting metadata (e.g. noun/verb/adjective, present/past tense, etc.)
Almost all the practical MT systems adopt hybrid approaches to a certain extent, combining rule-based and statistical approaches. Most recently, more and more systems also take advantage of NMT to different degrees.
Measuring the quality of MT
Measuring and benchmarking MT quality remains a difficult challenge. While standardized quality scales exist, they only provide a comparative and not absolute measure of quality. This is important because what’s really needed is an automated way to identify problem texts so they can be routed for human review and post-edit. At present, the standard practice is to have human reviews look at a certain percentage of texts, or spend an assigned amount of time reviewing a subset of a project.
The most reliable method of MT quality evaluation requires human evaluators to score each sentence, either within text translated by an MT engine or in comparison with others. The average score on all the sentences from all evaluators is the final score. The most common metrics for human scoring are adequacy and fluency of translation.
Human evaluation is expensive and time-consuming and thus unsuitable for frequent use during research and development of MT engines. Various automatic evaluation methods are available to measure similarity of MT translation and that from a human translator. Some examples:
- Word error rate (WER) is defined based on the distance between the system output and the reference translation at the word level.
- Position-independent error rate (PER) calculates the word error rate by treating each sentence as a bag of words and ignoring the word order.
- Bilingual Evaluation Understudy (BLEU) computes the n-gram precision rather than word error rate.
- Metric for Evaluation of Translation with Explicit Ordering (METEOR) takes stemming and synonyms into consideration.
Automatic translation quality evaluation plays an important role in MT research since it helps measure quality between iterations of an engine and between different engines. However, the correlation between automatic and human evaluation metrics is not satisfactory.