Text-to-Text Generation for Summarizing Informal Genres

Columbia University, New York NY

Investigators

Abstract

This project aims at the generation of coherent and on-target summaries and answers through the use of text-to-text generation, an approach which generates new sentences from the input text, fusing relevant phrases and discarding irrelevant ones. A syntactic, statistical framework for text-to-text generation is being developed that can be applied to informal genres, such as transcribed speech and email, where sentences are not guaranteed to be either complete or grammatical. It is exactly these genres that stand to benefit the most from this approach; for them, summarization using sentence extraction alone is not an option. The aim is a fully developed, syntactic statistical framework for text-to-text generation which features the use of a full syntactic grammar within a statistical framework for compression and combination, a model for incorporating constraints from pragmatics and semantics into the generation system, the ability to produce fluent, grammatical sentences from fragmentary and ungrammatical input, and the ability to generate sentences that make high level abstractions from input document sentences. The project features the integration of compression and language models into a lexicalized head-driven framework, enabling the generator to keep the sentence grammatical and avoid wording changes that dramatically alter meaning. Its framework can incorporate an arbitrary number of features beyond syntax that are important for summarization. A new dynamic programming technique allows the automatic extraction of large amounts of training data from a summary/document corpus. Information about who speaks to whom and paraphrasing rules will increase the range of revisions that can be addressed.