|
Template Based Machine Translation (TBMT)
Template-based machine translation commonly uses translation template as the translation knowledge representation. Translation template is translation knowledge representation between the “translation rule” and “translation example”, which can be understood as translation instance having variable part or translation rules formed by constants and variables. Because of large granularity of translation rules, it is too abstract, while translation examples granularity is small, so it is too specific. Translation template can combine the advantages of both and reduce their shortcomings. Translation templates can be used in a variety of machine translation systems especially suited for MT system mixing a variety of translation knowledge representation.
Statistics-based MT and example-based MT have different ways of using translation templates. Example-based MT by searching translation template learning from the corpus to translate and statistics-based MT introducing the source-channel model and the probability theory knowledge, establishing probability model for the MT, and probability model provides a solution for the translation template conflicts. Statistics-based MT uses the translation template with probability information and the information is estimated from the corpus using statistical methods.
According to the definition in “A Chinese-English translation template extraction method” (written by Yang Erbao and, Lv Xueqiang), the broad definition of translation template is as follows:
Translation template is a translation norm, and is the collection of translation examples conforming to certain translation rules. For each element of the collection, its source language part is translated into target language according to specific constraints.
The translation template in this article complies with the above broad definition, and further refined translation template is defined as follows:
Respectively using the source language and target language writes strings which include constant (specific word fragments) and variable (variable word fragments) in two languages, and the two strings have correspondent translation relations.
A template includes a mapping relationship translation between the above two strings and their variables.
A simple translation template example of Chinese-English is as follows:
这是X1.
This is X1 .
In this template, each Chinese constant is corresponding to the English constant in the corresponding position, that is “这”corresponds to “This” and“是” corresponds “is”. The Chinese X1 and the English X1 are variables and the translation of Chinese X1 is the translation of English X1. In fact, the correspondence of translation template is much more complex than this one-to-one correspondence.The complexity of this correspondence is one of the advantages of using translation templates, because we do not need to calculate the correspondence which has already been described by translation template in the process of translating. One only needs to find the correct translation template. Therefore, the quality of translation template has important influence on the quality of MT.
The simple translation template of the above example has been able to describe the generalization of many common sentence patterns, such as:
Character introducing: “This is这是小张”and its corresponding translation of “This is Xiao Zhang”;
Referring to an article “这是一个苹果” and its corresponding translation of “This is an apple”;
Explaining reasons “这是因为他没来”and its corresponding translation of “This is because he did not show up” and so on.
Translation template is especially suited for texts having a fixed format, such as texts describing the knowledge of specific fields. A translation template of Chinese medicine patent field is as follows:
一种具有X1 作用的X2 及其制备方法
A X2 having X1 effects, and its preparing method
This template can cover the main contents of the sentence to be translated, describing the order of the complex relationship between bilingualism and generalizing to the translation of a variety of structurally similar sentences. For example, “a medicine having enriching effects and its preparation method”, and “a medicinal granule having hair loss prevention effects and its preparing methods” etc. Because of the frequently occurring sentence patterns in their field, such translation template is very effective in improving the translation accuracy.
Translation templates can also have additional matching conditions to improve the accuracy of template matching. Such as limiting X1 to be a noun phrase.
In the process of translation, the translation template can be used to translate the text that does not appear in the corpus, as long as these texts have the same structure as the translation template constant parts and meet the conditions for matching the variable parts. This feature can make up for corpus-based translation dependence on the corpus. Because of the characteristics of template generalization, translation templates can be seen as the appropriate generalization of bilingual text. MT using translation templates is scalable; high-quality translation template is often able to accurately describe the translation of sentence backbone and the information of long distance reordering.
Translation template having the same source language parts can have different target language parts. It has something to do with paraphrasing of natural language, which is the same sentence can have different expressions and different translations. So the source language part and target language part of the translation template exist many to many correspondence. In the MT process, according to different contexts, selects the most appropriate translation template from candidate templates of to be translated part.
Dividing in accordance with extraction method, translation template can fall into two categories: manual written translation template and automatic extraction translation template. Manual written translation template is written by people, high obtaining cost, but the accuracy rate and generality is good. Automatic extraction translation template, based on some of algorithm, is automatically extracted from the corpus, low extracting cost, but the accuracy rate and generality is usually worse than manual written template.
When using translation templates in the MT system, related problems have the following:
1. Extraction problem. Because of the effect of the translation template quality on machine translation quality, the problem of how to extract the translation template accurately describing linguistic phenomenon and in accordance with the grammar and syntax limitation is faced translation template researching.
2. Evaluation problem. When the corpus is large, the number of translation templates automatically extracted from the corpus is large and the quality is not uniform and may affect the speed and quality of the machine translation system. Evaluate the translation templates and the translation template of lower evaluation is filtered out and retaining the translation template of higher evaluation can reduce resources use of MT system and improve the translation quality.
3. Application problems. When translating sentences, needing to find a suitable translation template and then apply, and the method of correct application has something to do with sentence structure. In the process of translation, eliminate translation template conflicts; find the application of correct translation template is one of the problems needed to be solved.
|
