A Numeracy-Enhanced Decoding for Solving Math Word Problem

Rao Peng, Chuanzhi Yang, Litian Huang, Xiaopan Lyu, Hao Meng, Xinguo Yu

Abstract: Deep neural models have achieved promising progress in solving Math Word Problems (MWPs) recently. This paper presents a deep neural solver by adopting numeracy-enhanced decoding to promote the performance of expressions generation. It leverages numerical properties to enhance the capabilities of the decoder, primarily focusing on two aspects: token embedding and target prediction. For token embedding, this paper proposes a numeracy-enhanced token embedding method, which fuses the explicit numerical feature with the contextual feature for number tokens, enabling the decoder to perceive numerical properties during the inference. For target prediction, this paper proposes a dynamic target prediction method, which utilizes a numerical attention network to identify the mathematical category of the problem and adaptively invokes category-aware parameter matrices to generate diverse expressions for different problems. Experimental results demonstrate that the proposed method not only achieves competitive performance on the Chinese MWP dataset but also achieves state-of-the-art results on the NLPCC Shared Task 3 dataset.