|
Drawbacks of MT systems.
Submitting source text to MT systems such as Google Translate for the purpose of translation and language learning often does not work very well. Although MT systems have made tremendous progress in recent years, they still make mistakes in grammar and word choices, and translators or language learners are forced to accept these erroneous translations and are given little opportunity to learn the target language or translation skills in the process. A promising learning framework is to incorporate the user into the translation process during which the user is provided with subsequent grammar and text choices for the translation prefix, entered by the user, of the source text.
Working principles of TransAhead system.
Once the word usage preferences for word sequences, that is, writing suggestion module and translation equivalents, that is, translation suggestion module, are learned, they are stored for the real-time and real-world CAT system, TransAhead. TransAhead then predicts/suggests the grammar and text which immediately follow a translation prefix of the source text.
In Step (1), we first split the source text S into character-based n-grams, denoted by {si}. On the other hand, we find the word-level n-grams of the translation prefix Tp the user has entered (Step (2)). But this time we concentrate on the n-grams ending with the last word of Tp since, intuitively, these n-grams predict well the grammatical contents that are likely to follow Tp, or the subsequent grammar patterns. Take the translation prefix “we play an important role” in Figure 1 for instance. sliceToWordNgramWithPivot(Tp) will identify the pivotal n-grams of “role,” “important role,” “an important role,” “an role,” and “play role.” Notice that both consecutive and inconsecutive n-grams have an influence on predicting representative contexts. Consequently, the sliced n-grams include skipped ones such as “play role” and “an role”, as before.
Step (3) acquires the translation equivalents of source n-grams in {si} using the translation suggestion module in Section 3.2. There are two aspects worth mentioning. First, to handle the word segmentation issue in MT [Bai et al. 2008; Ma et al. 2007], TransAhead breaks the word boundary in the source text, finds source character n-grams’ translations, and interacts with the user while he/she translates along to segment the source text in bilingual sense and to complete the translation. In other words, the word boundary in our system is at first nondeterministic and is determined only when the corresponding translation is chosen by the user. Consider the Chinese (translation: “important”) in Figure 1. Although it is possible to translate the characters and individually, after the user keys in the translation “important”, its counterpart emerges as a word. The system and the user benefit from each other’s knowledge and compensate for each other’s insufficiencies. Second, an out-of-vocabulary (OOV) module [Huang et al. 2011b] that exploits constituent or character translations can be incorporated in Step (3) to reduce the impact of OOV on the system and the user, language learner in particular. Since the user directly interacts with TransAhead, the OOV module is of great importance. While the active vocabulary, TranslationOptions, in Step (3) is used to translate the source text, n-grams in {tj} is used to extract syntactic patterns via the writing suggestion model described in Section 3.2. Notice that the translation prefix reflects source segments that had been translated. Whenever the user enters a translation, the active vocabulary is altered and shrunk accordingly. Also, TransAhead may exploit a user’s vocabulary of preference (due to the user’s domain of knowledge or errors of the system) to improve system performance and user experience [Nepveu et al. 2004; Ortiz-Martinez et al. 2010].
Standard MT submodels can be utilized to rank TransAhead writing and translation predictions. These include source-to-target and target-to-source translation models, target-language language model, word alignment model (i.e., considering alignment positions of source words and target translations), and distortion model (i.e., considering the word orders of translations). In our case, the quality of grammar patterns may also be used. Besides system performance, the success of a CAT or CALL assistant further lies in user satisfaction, which closely correlates with the response time during user typing while translating, namely, the time delay between a user trigger and the return of TransAhead suggestions. To juggle speed and performance, our framework exploits language and directional translation models, the essential models in MT, in candidate ranking.
Finally, the algorithm outputs representative grammar patterns with high confidence translations (e.g., play ∼ role Prep(in) V+ing [close, . . . ]), or lexically translated grammar patterns expected to follow the current translation prefix for the source text and help language learning during the translation process. The algorithm is triggered each time a word delimiter (i.e., space) is entered, and the word boundary and the translation of the source text are determined along the way until the user submits the complete translation. Figure 1 shows example responses to the source text and two translation prefixes “we” and “we play an important role” in our working prototype, as applied in an interactive CAT and CALL environment.
Do'stlaringiz bilan baham: |
