Discourse-aware Statistical Machine Translation as a Context-sensitive Spell Checker

Proceedings of Recent Advances in Natural Language Processing, pages 475–482, Hissar, Bulgaria, 7-13 September 2013. Discourse-aware Statistical Machine Translation as a Context-Sensitive Spell Checker Behzad Mirzababaei, Heshaam Faili and Nava Ehsan School of Electrical and Computer Engineering College of Engineering University of Tehran, Tehran, Iran {b.mirzababaei,hfaili,n.ehsan}@ut.ac.ir Abstract Real-word errors or context sensitive spelling errors, are misspelled words that have been wrongly converted into another word of vocabulary. One way to detect and correct real-word errors is using Statistical Machine Translation (SMT), which translates a text containing some real-word errors into a correct text of the same language. In this paper, we improve the results of mentioned SMT system by employing some discourse- aware features into a log-linear reranking method. Our experiments on a real-world test data in Persian show an improvement of about 9.5% and 8.5% in the recall of detection and correction respectively. Other experiments on standard English test sets also show considerable improvement of real-word checking results. 1 Introduction Kukich (1992) has categorized errors of a text into five categories: 1. isolated error 2. syntactic error 3. real-word error 4. discourse structure and 5. pragmatic error. In this paper, we focus on the third category, which is also referred as context- sensitive spelling error. This type of error includes misspelled words that are converted to another word of the dictionary (e.g., typing “arm” instead of “are” in the sentence “we arm good”). In order to detect and correct this kind of error, context analysis of the text is crucial. Here, we propose a language-independent method, which is based on a phrase-based Statistical Machine Translation (SMT). In this case, the input and output sentences are both in the same language and the input sentence contains some real-word errors. Phrase-based SMT is weak in handling long- distance dependencies between the sentence words. In order to capture this kind of dependencies, which affects detecting the correct candidate word, mentioned SMT is augmented with a discourse-aware reranking method for reranking the N-best results of SMT. Our work can be regarded as an extension of the method introduced by Ehsan and Faili (2013), in which they use SMT to detect and correct the spelling errors of a document. But here, we use the N-best results of SMT as a candidate list for each erroneous word and rerank the list by using a discourse-aware reranking system which is just a log-linear ranker. Shortly, the contributions of this paper can be summarized as follow: The N-best results of SMT are regarded as a candidate list of suspicious word, which is reranked by using a discourse-aware reranking system. Two discourse-aware features are employed in a log- linear ranker. The keywords in whole document surrounding the erroneous sentence are considered as the context window. We have achieved about 5% improvement over the SMT- based approach in detection and correction recall and 1% in precision on English experiment. The state-of-the-art results are achieved for Persian context-sensitive spell checker respect to F- measure and Mean Reciprocal Rank metrics. This …

Content truncated to 500 words