The Automatic Translation of Discourse Structures

We empirically show that there are significant differences between the discourse structure of Japanese texts and the discourse structure of their corresponding English translations. To improve translation quality, we propose a computational model for rewriting discourse structures. When we train our model on a parallel corpus of manually built Japanese and English discourse structure trees, we learn to rewrite Japanese trees as trees that are closer to the natural English rendering than the original ones. 1 Motivation Almost all current MT systems process text one sentence at a time. Because of this limited focus, MT systems cannot regroup and reorder the clauses and sentences of an input text to achieve the most natural rendering in a target language. Yet, even between languages as close as English and French, there is a 10% mismatch in number of sentences-what is said in two sentences in one language is said in only one, or in three, in the other (Gale and Church, 1993). For distant language pairs, such as Japanese and English, the differences are more significant. Consider, for example, Japanese sentence (1), a word-byword "gloss" of it (2), and a two-sentence translation of it that was produced by a professional translator (3). [The Ministry of Health and Welfare last year revealed I ] [population of future estimate according to 2] [in future 1.499 persons as the lowest s] [that after *SAB* rising to turn that 4] [*they* estimated but s ] [already the estimate misses a point ~] [prediction became. 7] (2) [In its future population estimates'] [made (3) public last year, 2] [the Ministry of Health and Welfare predicted that the SAB would drop to a new low of 1.499 in the future, s) [but would make a comeback after that, 4] [increasing once again, s ] [However, it looks as if that prediction will be quickly shattered. 6] The labeled spans of text represent elementary discourse units (edus), i.e., minimal text spans that have an unambiguous discourse function (Mann and Thompson, 1988). If we analyze the text fragments closely, we will notice that in translating sentence (1), a professional translator chose to realize the information in Japanese unit 2 first (unit 2 in text (1) corresponds roughly to unit 1 in text (3)); to realize then some of the information in Japanese unit 1 (part of unit 1 in text (1) corresponds to unit 2 in text (3)); to fuse …