[1D5-GS-9] Natural language processing, information retrieval: Models and meaning acquisition

[1D5-GS-9-05] Consistent Data-to-Text Generation with Topic Sequences

〇Soichiro Murakami^1,2,3, Sora Tanaka^1,4, Masatsugu Hangyo⁵, Hidetaka Kamigaito¹, Hiroya Takamura^1,3, Manabu Okumura¹ (1. Tokyo Institute of Technology, 2. NTT DOCOMO, INC., 3. National Institute of Advanced Industrial Science and Technology, 4. Gurunavi, Inc., 5. Weathernews Inc.)

Keywords:Weather Forecast Commentary, Natural Language Generation, Consistency, Data-to-Text

Data-to-Text is the task of generating text that describes the contents from various types of data, such as weather forecast maps and time-series stock data. With the recent advances in neural networks, the data-to-text models have been making remarkable progress in terms of the ability to capture the characteristics of such complicated data and generate text that accurately describes their contents. However, concerning the generation of a document-scale text that includes multiple sentences, the data-to-text model often generates descriptions that mention duplicated contents and lacks the consistency of their topics due to the over-generation problem. In this paper, we focus on generating descriptions having not only accuracy but also consistency regarding their contents to tackle the problem mentioned above, which often causes in the data-to-text task. We propose a method to generate consistent text by predicting a sequence of topics from data and assigning it to the generation model to control the topics and their order. In the experiment, we show that the data-to-text model produces text containing consistent topics by specifying a sequence of topics to be mentioned and that it helps to relieve the over-generation problem.