Contextualized Embedding-Guided Summarization for Multi-Review Unique Selling Points Extraction

Abstract

User-Generated Content (UGC) has grown so quickly, alongside practical approaches that summarize such content, that demand for efficient methods has grown tremendously and firms that extract Unique Selling Points (USPs), through the summarization of multi-user reviews. This task is inherently challenging due to the weak correspondence between succinct USP summaries and individual sentences in the source review material. A prior example, USEsum, used sentence embeddings to automatically select appropriate content during summarization, but the implementation of a generic encoder and one aspect were limitations. In this work, we present CEG-Sum, a two-phase hybrid model for improved extraction of USPs from reviews. CEG-Sum relies on domain-adapted contextualized embeddings, and a Multi-Aspect Content Selection module trained to predict multiple aspect vectors that allow for comprehensive summaries of multiple USPs. The model subsequentley uses an abstractive summary generation parameter to further improve content selection with an input word promotion and use of Candidate Summary Reranking both to optimize summaries for volubility and semantic relevance at the same time. To demonstrate how CEG-Sum can produce more informational, coherent, and comprehensive summaries, experiments plus human evaluations reported it to show significant performance improvement over prior multi-review versions based on multiple USPs.