Incorporating word embeddings in the hierarchical dirichlet process for query-oriented text summarization

The ever-growing amount of textual data available online creates the need for automatic text summarization tools. Probabilistic topic models are able to infer semantic relationships between sentences which is a key step of extractive summarization methods. However, they strongly rely on word co-occurrence patterns and fail to capture the actual semantic relationships between words such as synonymy, antonymy, etc. We propose a novel algorithm which incorporates pre-trained word embeddings in the probabilistic topic model in order to capture semantic similarities between sentences. These similarities provide the basis for a sentence ranking algorithm for query-oriented summarization. The summary is then produced by extracting highly ranked sentences from the original corpus. Our method is shown to outperform state-of-the-art algorithms on a benchmark dataset.