Download files
Abstract
Automatic text summarization is the process of automatically creating a compressed version of a given text. Content reduction can be addressed by extraction or abstraction. Extractive methods select a subset of most salient parts of the source text for inclusion in the summary. In contrast, abstractive methods build an internal semantic representation to create a more human-like summary. The majority of summarizers are designed to be extractive due to the complex nature of abstraction. This thesis moves toward abstractive text summarization, and makes this task: (i) more adaptable to a wide range of applications; (ii) more dynamic to different sources and types of text; and (iii) better evaluated using semantic representations.
To make it more adaptable, we propose a word graph-based multi-sentence compression approach for improving both informativity and grammaticality of summaries, which shows 44% error reduction over state-of-the-art systems. Then, we discuss adapting this approach into query-focused multi-document summarization, focusing on semantic similarities between the input query and source texts. This approach satisfies the query-biased relevance, information novelty and richness criteria.
To make this task more dynamic, we appraise the coverage of knowledge sources for the purpose of abstractive text summarization, and found a decline in performance of summarizers that only rely on specific terminologies. Our approach integrates general and domain-specific lexicons for incorporating textual semantic similarities, and bridging the knowledge and language gaps in domain-specific summarizers.
To fairly evaluate abstractive summaries including lexical variations and paraphrasing, we propose an approach based on both lexical and semantic similarities, which highly correlates with human judgments. Furthermore, we present an approach to evaluate summaries on test sets where model summaries are not available. Our hypothesis is that comparing semantic representations of the input and summary content leads to a more accurate evaluation. We exploit the compositional capabilities of corpus-based and lexical resource-based word embeddings for predicting the summary content quality. The experiment results support our proposal to use semantic representations for model-based and model-free evaluation of summaries.