PDF, English Download (2MB) | Terms of use

Abstract

Automatic text summarization is one of the eminent applications in the field of Natural Language Processing. Text summarization is the process of generating a gist from text documents. The task is to produce a summary which contains important, diverse and coherent information, i.e., a summary should be self-contained. The approaches for text summarization are conventionally extractive. The extractive approaches select a subset of sentences from an input document for a summary. In this thesis, we introduce a novel graph-based extractive summarization approach. With the progressive advancement of research in the various fields of science, the summarization of scientific articles has become an essential requirement for researchers. This is our prime motivation in selecting scientific articles as our dataset. This newly formed dataset contains scientific articles from the PLOS Medicine journal, which is a high impact journal in the field of biomedicine. The summarization of scientific articles is a single-document summarization task. It is a complex task due to various reasons, one of it being, the important information in the scientific article is scattered all over it and another reason being, scientific articles contain numerous redundant information. In our approach, we deal with the three important factors of summarization: importance, non-redundancy and coherence. To deal with these factors, we use graphs as they solve data sparsity problems and are computationally less complex. We employ bipartite graphical representation for the summarization task, exclusively. We represent input documents through a bipartite graph that consists of sentence nodes and entity nodes. This bipartite graph representation contains entity transition information which is beneficial for selecting the relevant sentences for a summary. We use a graph-based ranking algorithm to rank the sentences in a document. The ranks are considered as relevance scores of the sentences which are further used in our approach. Scientific articles contain reasonable amount of redundant information, for example, Introduction and Methodology sections contain similar information regarding the motivation and approach. In our approach, we ensure that the summary contains sentences which are non-redundant. Though the summary should contain important and non-redundant information of the input document, its sentences should be connected to one another such that it becomes coherent, understandable and simple to read. If we do not ensure that a summary is coherent, its sentences may not be properly connected. This leads to an obscure summary. Until now, only few summarization approaches take care of coherence. In our approach, we take care of coherence in two different ways: by using the graph measure and by using the structural information. We employ outdegree as the graph measure and coherence patterns for the structural information, in our approach. We use integer programming as an optimization technique, to select the best subset of sentences for a summary. The sentences are selected on the basis of relevance, diversity and coherence measure. The computation of these measures is tightly integrated and taken care of simultaneously. We use human judgements to evaluate coherence of summaries. We compare ROUGE scores and human judgements of different systems on the PLOS Medicine dataset. Our approach performs considerably better than other systems on this dataset. Also, we apply our approach on the standard DUC 2002 dataset to compare the results with the recent state-of-the-art systems. The results show that our graph-based approach outperforms other systems on DUC 2002. In conclusion, our approach is robust, i.e., it works on both scientific and news articles. Our approach has the further advantage of being semi-supervised.