Text similarity semantic calculation based on deep reinforcement learning
Guanlin Chen, Xiaolong Shi, Moke Chen, Liang Zhou
Abstract
Semantic analysis is a fundamental technology in natural language processing. Semantic similarity calculations are involved in many applications of natural language processing, such as QA system, machine translation, text similarity calculation, text classification, information extraction and even speed recognition, etc. This paper proposes a new framework for computing semantic similarity: deep reinforcement learning for Siamese attention structure model (DRSASM). The model learns word segmentation automatically and word distillation automatically through reinforcement learning. The overall architecture LSTM network to extract semantic features, and then introduces a new attention mechanism model to enhance semantics. The experiment show that this new model on the SNLI dataset and Chinese business dataset can improve the accuracy compared to current base line structure models.