Vector space and language models for scientific document summarization

Abstract

In this paper we compare the performance of three approaches for estimating the latent weights of terms for scientific document summarization, given the document and a set of citing documents. The first approach is a term-frequency (TF) vector space method utilizing a nonnegative matrix factorization (NNMF) for dimensionality reduction. The other two are language modeling approaches for predicting the term distributions of human-generated summaries. The language model we build exploits the key sections of the document and a set of citing sentences derived from auxiliary documents that cite the document of interest. The parameters of the model may be set via a minimization of the Jensen-Shannon (JS) divergence. We use the OCCAMS algorithm (Optimal Combinatorial Covering Algorithm for Multi-document Summarization) to select a set of sentences that maximizes the term-coverage score while minimizing redundancy. The results are evaluated with standard ROUGE metrics, and the performance of the resulting methods achieve ROUGE scores exceeding those of the average human summarizer.