Discourse Structures to Reduce Discourse Incoherence in Blog Summarization

Proceedings of Recent Advances in Natural Language Processing, pages 479–486, Hissar, Bulgaria, 12-14 September 2011. Discourse Structures to Reduce Discourse Incoherence in Blog Summarization Shamima Mithun Concordia University Computer Science & Software Eng. Montreal, Canada s mithun@encs.concordia.ca Leila Kosseim Concordia University Computer Science & Software Eng. Montreal, Canada kosseim@cse.concordia.ca Abstract Discourse incoherence is an important and typical problem with multi-document ex- tractive summaries. To address this issue, we have developed a schema-based sum- marization approach for query-based blog summaries that utilizes discourse struc- tures. In our schema design, we tried to model discourse structures which are typ- ically used by humans in their summary writing in response to a particular question type. In our approach, a sentence instan- tiates a specific slot of the schema based on its discourse structures. To validate our approach, we have built a system named BlogSum and have evaluated its perfor- mance through 4 human participants using a likert scale of 1 to 5. The evaluation re- sults show that our approach has signifi- cantly improved summary coherence com- pared to the summaries with no discourse structuring without compromising on con- tent evaluation. 1 Introduction Research on text summarization dates back since the 1950s and with the growth of the Internet it has become a popular research topic in last decade. Text summarization reduces text search time by providing the most relevant information from the documents which enables users to comprehend more quickly the main ideas of a set of docu- ments. Over time, different summarization tech- niques have been developed and evaluated. Al- though significant improvement continues to be made, the summaries generated automatically are by no means of the same quality as their human created counter parts. The area in which auto- matic summaries differ most from human gener- ated summaries is text coherence (Otterbacher et al., 2002; Conroy and Dang, 2008; Genest et al., 2009). Coherence problems can be the result of dif- ferent phenomena: discourse incoherence, redun- dancy, temporal incoherence, grammatical mis- takes or many other linguistic problems. In a man- ual analysis of 15 summaries, (Otterbacher et al., 2002) showed that coherence problems are caused mostly by discourse incoherence (34%) where the main concern is the lack of relations between sen- tences as well as in the overall summary. Recently, (Genest et al., 2009) demonstrated that the performance of automatic summarizers in term of linguistic quality is significantly weaker compared to that of a baseline consisting of sen- tences extracted from the source documents by 5 human extractors and added to the summary with- out any modification. This result indicates that there is still much space to improve coherence of summaries even for pure extractive summaries. 1.1 Discourse Incoherence Computational theories on discourse coherence were introduced by (Hobbs, 1985; Mann, 1988). According to (Mann, 1988), a discourse is coher- ent if the hearer knows the communicative role of each of its portion; that is, if the hearer knows how the speaker intends each clause to relate to each other. As …

Content truncated to 500 words