MKPM: Multi keyword-pair matching for natural language sentences

Abstract Sentence matching is widely used in various natural language tasks, such as inference, paraphrase identification and question answering. For these we need to understand the logical semantic relationship between two sentences. Most current methods use all information within a sentence build model hence determine its another sentence. However, contained some sentences may cause redundancy or introduce noise, impeding performance of model. Therefore, propose method based on multi keyword-pair (MKPM), which uses keyword pairs represent them, avoiding interference noise. Specifically, first sentence-pair-based attention mechanism sp-attention select most important word pair from pair, then Bi-task architecture pairs. The follows: 1. In order at level sentences, design word-pair task (WP-Task), complete independently. 2. We sentence-pair (SP-Task) by denoising. Through integration our can more accurately granularities Experimental results show that achieve state-of-the-art several tasks. Our source code publicly available 1 .