Detecting Content-Heavy Sentences: A Cross-Language Case Study

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1271–1281, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Detecting Content-Heavy Sentences: A Cross-Language Case Study Junyi Jessy Li University of Pennsylvania ljunyi@seas.upenn.edu Ani Nenkova University of Pennsylvania nenkova@seas.upenn.edu Abstract The information conveyed by some sen- tences would be more easily understood by a reader if it were expressed in multi- ple sentences. We call such sentences con- tent heavy: these are possibly grammatical but difficult to comprehend, cumbersome sentences. In this paper we introduce the task of detecting content-heavy sentences in cross-lingual context. Specifically we develop methods to identify sentences in Chinese for which English speakers would prefer translations consisting of more than one sentence. We base our analysis and definitions on evidence from multiple hu- man translations and reader preferences on flow and understandability. We show that machine translation quality when translat- ing content heavy sentences is markedly worse than overall quality and that this type of sentence are fairly common in Chinese news. We demonstrate that sen- tence length and punctuation usage in Chi- nese are not sufficient clues for accurately detecting heavy sentences and present a richer classification model that accurately identifies these sentences. 1 Introduction To generate text, people and machines need to de- cide how to package the content they wish to ex- press into clauses and sentences. There are multi- ple possible renderings of the same information, with varying degrees of ease of comprehension, compactness and naturalness. Some sentences, even though they are grammatical, would be more accessible to a reader if expressed in multiple sen- tences. We call such sentences content heavy sen- tences, or heavy sentences for brevity. In the established areas of language research, text simplification and sentence planning in dia- log and generation systems are clearly tasks in which identification of content-heavy sentences is of great importance. In this paper we introduce a novel flavor of the task in the cross-lingual setting, which in the long term may guide improvements in machine translation. We seek to identify sentences in Chinese that would result in heavy sentences in English if translated to a single sentence. Example I in Table 1 shows a Chinese sentence and its two English translations A and B. Transla- tor A used three English sentences to express all the information. Translator B, on the other hand, used a single sentence, which most readers would find more difficult to read. Example II illustrates a case where a translator would be hard pressed to convey all the content in a sentence in Chinese into a single grammatical English sentence. Here we provide an operational characteriza- tion of content-heavy sentences in the context of Chinese-English translation. Instead of establish- ing guidelines for standalone annotation, we re- purpose datasets developed for evaluation of ma- chine translation consisting of multiple reference translations for each Chinese sentence. In this cross-lingual analysis, sentences in Chinese are considered content-heavy if their content would be more felicitously expressed in multiple …

Content truncated to 500 words