Generating Natural Language Descriptions from Tables

Abstract

This paper proposes a neural generative architecture, namely NLDT, to generate a natural language short text describing a table which has formal structure and valuable information. Specifically, the architecture maps fields and values of a table to continuous vectors and then generates a natural language description by leveraging the semantics of a table. The NLDT architecture adopts a two-level neural model to make the most of the structure of a table to fully express the relationship between contents. To deal with the problem of out-of-vocabulary, this paper develops a simple and fast word-conversion method that replaces rare words appearing in texts with common field information in tables and directly replicates contents from table to the output sequence according to the field information. Besides, this paper adds the concept of theme to adapt the NLDT architecture to open domain and improves beam search algorithm to strengthen the results in the inference stage. On the WEATHERGOV dataset, the NLDT architecture improves the state-of-the-art BLEU-4 score from 61.01 to 62.89 and the current state-of-the-art F1 score from 73.21 to 78. On the WIKIBIO and WIKITABLE datasets, the NLDT architecture achieves a BLEU-4 score of 45.77 and 38.71 respectively which also outperform the state-of-the-art approaches. Furthermore, this paper introduces a Chinese dataset WIKIBIOCN including 33,244 biographies with corresponding tables. On the WIKIBIOCN dataset, the NLDT architecture achieves a BLEU-4 score of 38.87 and fairly good manual evaluation.