Topic and Sentiment in Phrase-Based Statistical Machine Translation
Maryam Habibi, Nikolaos Pappas, Andréi Popescu-Belis · Infoscience (Ecole Polytechnique Fédérale de Lausanne) · 2017
In this paper, we model two textual properties, topic and sentiment, at the sentence and document levels, with the goal of improving the performance of machine translation by taking into account this information in source and target sentences.In the topical similarity approach, we augment the source sentence with the keywords extracted from its adjacent sentences and re-rank the candidate target sentences (hypotheses of a phrase-based statistical MT system, here Moses) in terms of their topical similarity to the augmented source sentence.The advantage of the model is being independent from the baseline MT system, similarly to IR re-ranking techniques.We model sentiment using the sentiment categories of words and sentences as factors in the same MT system.The results on English-French MT show that topic modeling improves lexical choice with respect to the baseline for about 5% of the lexical items that differ between the two systems.We observe that although the improvement obtained using topical information is not significant in terms of BLEU score, there is an improvement on the choice of terms in the target language based on topical information.As for sentiment information, it leads to an increase in BLEU scores of up to 1.5%.