Semantically-oriented text planning for automatic summarization

Resumo

Text summarization deals with the automatic creation of summaries from one or more documents, either by extracting fragments from the input text or by generating an abstract de novo. Research in recent years has become dominated by a new paradigm where summarization is addressed as a mapping from a sequence of tokens in an input document to a new sequence of tokens summarizing the input. Works following this paradigm apply supervised deep learning methods to learn sequence to sequence models from a large corpus of documents paired with human-crafted summaries. Despite impressive results in automatic quantitative evaluations, this approach to summarization also suffers from a number of drawbacks. One concern is that learned models tend to operate in a black-box fashion that prevents obtaining insights or results from intermediate analysis that could be applied to other tasks -an important consideration in many real-world scenarios where summaries are not the only desired output of a natural language processing system. Another significant drawback is that deep learning methods are largely constrained to languages and types of summary for which abundant corpora containing human authored summaries is available. Albeit researchers are experimenting with transfer learning methods to overcome this problem, it is far from clear how effective these methods are and how to apply them to scenarios where summaries need to adapt to a query or to user preferences. In those cases where it is not practical to learn a sequence to sequence model, it is convenient to fall back to a more traditional formulation of summarization where the input documents are first analyzed, then a summary is planned by selecting and organizing contents, and the final summary is generated either extractively or abstractively --using natural language generation methods in the latter case. By separating linguistic analysis, planning and generation, it becomes possible to apply different approaches to each task. This thesis focuses on the text planning step. Drawing from past research in word sense disambiguation, text summarization and natural language generation, this thesis presents an unsupervised approach to planning the production of summaries. Following the observation that a common strategy for both disambiguation and summarization tasks is to rank candidate items --meanings, text fragments-- we propose a strategy, at the core of our approach, that ranks candidate lexical meanings and individual words in a text. These ranks contribute towards the creation of a graph-based semantic representation from which we select non-redundant contents and organize them for inclusion in the summary. The overall approach is supported by lexicographic databases that provide cross-lingual and cross-domain knowledge, and by textual similarity methods used to compare meanings with each other and with the text. The methods presented in this thesis are tested on two separate tasks, disambiguation of word senses and named entities, and single-document extractive summarization of English texts. The evaluation of the disambiguation task shows that our approach produces useful results for tasks other than summarization, while evaluating in an extractive summarization setting allows us to compare our approach to existing summarization systems. While the results are inconclusive with respect to state-of-the-art in disambiguation and summarization systems, they hint at a large potential for our approach.