An Approach to Text Summarization Sankar K Sobha L AU-KBC Research Centre AU-KBC Research Centre MIT Campus, Anna University MIT Campus, Anna University Chennai- 44. Chennai- 44. [email protected] [email protected] networks and more recently in text processing ap- Abstract plications (Mihalcea and Tarau, 2004), (Mihalcea et al., 2004), (Erkan and Radev, 2004) and (Mihal- We propose an efficient text summarization cea, 2004). These iterative approaches have a high technique that involves two basic opera- time complexity and are practically slow in dy- tions. The first operation involves finding namic summarization. Proposals are also made for coherent chunks in the document and the coherence based automated summarization system second operation involves ranking the text (Silber and McCoy, 2000). in the individual coherent chunks and pick- We propose a novel text summarization tech- ing the sentences that rank above a given nique that involves two basic operations, namely threshold. The coherent chunks are formed finding coherent chunks in the document and rank- by exploiting the lexical relationship be- ing the text in the individual coherent chunks tween adjacent sentences in the document. formed. Occurrence of words through repetition or For finding coherent chunks in the document, we relatedness by sense relation plays a major propose a set of rules that identifies the connection role in forming a cohesive tie. The pro- between adjacent sentences in the document. The posed text ranking approach is based on a connected sentences that are picked based on the graph theoretic ranking model applied to rules form coherent chunks in the document. For text summarization task. text ranking, we propose an automatic and unsu- pervised graph based ranking algorithm that gives improved results when compared to other ranking 1 Introduction algorithms. The formation of coherent chunks greatly improves the amount of information of the Automated summarization is an important area in text picked for subsequent ranking and hence the NLP research. A variety of automated summariza- quality of text summarization. tion schemes have been proposed recently. NeATS The proposed text ranking technique employs a (Lin and Hovy, 2002) is a sentence position, term hybrid approach involving two phases; the first frequency, topic signature and term clustering phase employs word frequency statistics and the based approach and MEAD (Radev et al., 2004) is second phase involves a word position and string a centroid based approach. Iterative graph based pattern based weighing algorithm to find the Ranking algorithms, such as Kleinberg’s HITS weight of the sentence. A fast running time is algorithm (Kleinberg, 1999) and Google’s Page- achieved by using a compression hash on each sen- Rank (Brin and Page, 1998) have been traditionally tence. and successfully used in web-link analysis, social 53 Proceedings of CLIAWS3, Third International Cross Lingual Information Access Workshop, pages 53–60, Boulder, Colorado, June 2009. c 2009 Association for Computational Linguistics This paper is organized as follows: section 2 Rule 2 discusses lexical cohesion, section 3 discusses the text ranking algorithm and section 4 describes the A 3rd person pronominal in a given sentence refers summarization by combining lexical cohesion and to the antecedent in the previous sentence, in such summarization. a way that the given sentence gives the complete meaning with respect to the previous sentence. 2 Lexical Cohesion When such adjacent sentences are found, they form coherent chunks. Coherence in linguistics makes the text semantical- ly meaningful. It is achieved through semantic fea- Rule 3 tures such as the use of deictic (a deictic is an expression which shows the direction. ex: that, The reappearance of NERs in adjacent sentences this.), anaphoric (a referent which requires an ante- is an indication of connectedness. When such adja- cedent in front. ex: he, she, it.), cataphoric (a refe- cent sentences are found, they form coherent rent which requires an antecedent at the back.), chunks. lexical relation and proper noun repeating elements (Morris and Hirst, 1991). Robert De Beaugrande Rule 4 and Wolfgang U. Dressler define coherence as a “continuity of senses” and “the mutual access and An ontology relationship between words across relevance within a configuration of concepts and sentences can be used to find semantically related relations” (Beaugrande and Dressler, 1981). Thus a words across adjacent sentences that appear in the text gives meaning as a result of union of meaning document. The appearance of related words is an or senses in the text. indication of its coherence and hence forms cohe- The coherence cues present in a sentence are di- rent chunks. rectly visible when we go through the flow of the All the above rules are applied incrementally to document. Our approach aims to achieve this ob- achieve the complete set of coherent chunks. jective with linguistic and heuristic information. The identification of semantic neighborhood, oc- 2.1.1 Connecting Word currence of words through repetition or relatedness by sense relation namely synonyms, hyponyms and The ACE Corpus was used for studying the cohe- hypernym, plays a major role in forming a cohesive rence patterns between adjacent sentences of the tie (Miller et al., 1990). document. From our analysis, we picked out a set of keywords such that, the appearance of these 2.1 Rules for finding Coherent chunks keywords at the beginning of the sentence provide a strong lexical tie with the previous sentence. When parsing through a document, the relationship The appearance of the keywords “accordingly, among adjacent sentences is determined by the again, also, besides, hence, henceforth, however, continuity that exists between them. incidentally, meanwhile, moreover, namely, never- We define the following set of rules to find co- theless, otherwise, that is, then, therefore, thus, herent chunks in the document. and, but, or, yet, so, once, so that, than, that, till, whenever, whereas and wherever”, at the begin- Rule 1 ning of the present sentence was found to be highly coherent with the previous sentence. The presence of connectives (such as accordingly, Linguistically a sentence cannot start with the again, also, besides) in present sentence indicates above words without any related introduction in the connectedness of the present sentence with the the previous sentence. previous sentence. When such connectives are Furthermore, the appearance of the keywords found, the adjacent sentences form coherent “consequently, finally, furthermore”, at the begin- chunks. ning or middle of the present sentence was found to be highly cohesive with the previous sentence. Example 1 54 1. a The train was late. In Example 4, the pronominal he in the second sen- 1. b However I managed to reach the wedding tence refers to the antecedent Ravi in the first sen- on time. tence. In Example 1, the connecting word however binds Example 5 with the situation of the train being late. 1. a He is the one who got the first prize. Example 2 1. a The cab driver was late. In example 5 the pronominal he is possessive and 1. b The bike tyre was punctured. it doesn’t need an antecedent to convey the mean- 1. c The train was late. ing. 1 .d Finally, I managed to arrive at the wed- ding on time by calling a cab. 2.1.3 NERs Reappearance Example 3 Two adjacent sentences are said to be coherent 1. a The cab driver was late. when both the sentences contain one or more reap- 1. b The bike tyre was punctured. pearing nouns. 1. c The train was late. 1. d I could not wait any more; I finally ma- Example 6 naged to reach the wedding on time by calling a 1. a Ravi is a good boy. cab. 1. b Ravi scored good marks in exams. In Example 2, the connecting word finally binds Example 7 with the situation of him being delayed. Similarly, 1. a The car race starts at noon. in Example 3, the connecting word finally, though 1. b Any car is allowed to participate. it comes in the middle of the sentence, it still binds with the situation of him being delayed. Example 6 and Example 7 demonstrates the cohe- 2.1.2 Pronominals rence between the two sentences through reappear- ing nouns. In this approach we have a set of pronominals which establishes coherence in the text. From our 2.1.4 Thesaurus Relationships analysis, it was observed that if the pronominals WordNet covers most of the sense relationships. “he, she, it, they, her, his, hers, its, their, theirs”, To find the semantic neighborhood between adja- appear in the present sentence; its antecedent may cent sentences, most of the lexical relationships be in the same or previous sentence. such as synonyms, hyponyms, hypernyms, mero- It is also found that if the pronominal is not pos- nyms, holonyms and gradation can be used (Fell- sessive (i.e. the antecedent appears in the previous baum 1998). Hence, semantically related terms are sentence or previous clause), then the present sen- captured through this process. tence and the previous sentences are connected. However, if the pronominal is possessive then it Example 8 behaves like reflexives such as “himself”, “herself” 1. a The bicycle has two wheels. which has subject as its antecedent. Hence the pos- 1. b The wheels provide speed and stability. sibility of connecting it with the previous sentence is very unlikely. Though pronominal resolution In Example 8, bicycle and wheels are related cannot be done at a window size of 2 alone, still through bicycle is the holonym of wheels. we are looking at window size 2 alone to pick guaranteed connected sentences. 2.2 Coherence Finding Algorithm Example 4 The algorithm is carried out in four phases. Initial- 1.