P. Gayathri ¹, N. Jaisankar ¹

Affiliations
1 School of Computing Science and Engineering, VIT University, Vellore – 632014, Tamil Nadu, IN

Abstract

Background/Objectives: As all documents related to medical domain do not come with author written summary, the objective is to introduce a summarizer that exploits medical domain-specific knowledge. Methods/Statistical Analysis: Sentence ranking technique has been used to produce high quality summary. The features such as sentence position, length, cue-words (domain-related terms) and acronyms are extracted to assign sentence score. Sentences are ranked and arranged in the decreasing order of their normalized score. The existing summarization approaches in the literature use few or more sentence features but we have opted for few best sentence features. Pre-existing summarizers are used for performance evaluation. Findings: The few best features to be considered in developing medical domain-specific summarizers are sentence position, sentence length, number of cue-words and number of acronyms. Summary produced by any summarizer can be highly informative if and only if it contains dissimilar sentences. Therefore, similarity between sentences is an important feature to be considered for creating highly informative summary. The proposed summarizer is compared with the pre-existing summarizers. The evaluation is done by using traditional metrics such as precision and recall and ROUGE. Not all medical documents come with an author written abstract or summary. So, medical documents with author written abstracts are used to test the performance. Results reveals that the proposed summarizer performs better when compared with existing summarizers and attained ROUGE scores also reveals the same with respect to quality of summary produced. Thus, proposed summarizer provide highly acceptable summary to user. Application/Improvements: Summarization is one of the information retrieval tasks. It helps to determine whether the retrieved document is relevant for in-depth study or not.

Keywords

Feature Extraction, Medical Document Summarization, Sentence Feature, Sentence Ranking, Summarization

Full Text

   

P. Gayathri ¹, N. Jaisankar ¹

Affiliations
1 School of Computing Science and Engineering, VIT University, Vellore – 632014, Tamil Nadu, IN

Abstract

Background/Objectives: This paper gives an overview of some tasks involved in the retrieval process, corpus and resources of medical information retrieval. Methods/Statistical Analysis: Inverted file representation method is used in the retrieval process for associating documents in the corpus with various search terms. Conventional statistical ranking functions such as Jaccard, Okapi and Euclidean have been widely used for ranking retrieved medical documents. An extractive informative generic mono-lingual single-document summarizer is used to produce medical domain-specific summary. Sentence ranking method is used to include most appropriate sentences in the final summary. Findings: Studies reveal that people are searching the web and read medical related information in order to be informed about their health. In the medical domain, richest and most used source of information is MEDLINE. Because of frequent use of acronyms in the medical literature, using the term that appears in documents as keywords for document indexing would not be effective. Also, using Bag of Words representation could not capture the semantic meaning of terms. Some domain-specific thesauri like UMLS, MeSH and Gene ontology are available for biomedical retrieval. These domain-specific thesauri can provide synonyms, hypernyms and hyponyms of a specific term but it does not look into the context. Therefore, the retrieval results of using domain-specific thesauri are somewhat conflicting. It is possible to identify which lexical variant of specific term should be used under specific context by using Wikipedia as resource for biomedical retrieval. Conventional ranking functions fail to capture the inherent features of natural language text. Evolutionary algorithm based ranking can enhance the retrieval performance. Any domain-specific summarizer must consider similarity between sentences as essential feature for summarization. Applications/Improvements: Improvements in retrieval results is achieved by using context-aware keywords as indexing keywords and highly robust hybrid evolutionary algorithm based ranking function for ordering the retrieved documents.

Keywords

Information Retrieval, Medical Information retrieval, Medical Document Corpus, Resources, Retrieval Process.

Full Text