Semantic search as extractive paraphrase span detection

Abstract

In this paper, we approach the problem of semantic search by introducing a task of paraphrase span detection, i.e. given a segment of text as a query phrase, the task is to identify its paraphrase in a given document, the same modelling setup as typically used in extractive question answering. While current work in paraphrasing has almost uniquely focused on sentence-level approaches, the novel span detection approach gives a possibility to retrieve a segment of arbitrary length. On the Turku Paraphrase Corpus of 100,000 manually extracted Finnish paraphrase pairs including their original document context, we find that by achieving an exact match of 88.73 our paraphrase span detection approach outperforms widely adopted sentence-level retrieval baselines (lexical similarity as well as BERT and SBERT sentence embeddings) by more than 20pp in terms of exact match, and 11pp in terms of token-level F-score. This demonstrates a strong advantage of modelling the paraphrase retrieval in terms of span extraction rather than commonly used sentence similarity, the sentence-level approaches being clearly suboptimal for applications where the retrieval targets are not guaranteed to be full sentences. Even when limiting the evaluation to sentence-level retrieval targets only, the span detection model still outperforms the sentence-level baselines by more than 4 pp in terms of exact match, and almost 6pp F-score. Additionally, we introduce a method for creating artificial paraphrase data through back-translation, suitable for languages where manually annotated paraphrase resources for training the span detection model are not available.