Probabilistic Verb Selection for Data-to-Text Generation

Probabilistic Verb Selection for Data-to-Text Generation Dell Zhang†1, Jiahao Yuan‡, Xiaoling Wang‡2, and Adam Foster† †Birkbeck, University of London, Malet Street, London WC1E 7HX, UK ‡Shanghai Key Lab of Trustworthy Computing, East China Normal University, 3663 North Zhongshan Road, Shanghai 200062, China 1dell.z@ieee.org, 2xlwang@sei.ecnu.edu.cn Abstract In data-to-text Natural Language Generation (NLG) systems, computers need to find the right words to describe phenomena seen in the data. This paper focuses on the problem of choosing appropriate verbs to express the di- rection and magnitude of a percentage change (e.g., in stock prices). Rather than simply using the same verbs again and again, we present a principled data-driven approach to this prob- lem based on Shannon’s noisy-channel model so as to bring variation and naturalness into the generated text. Our experiments on three large-scale real-world news corpora demon- strate that the proposed probabilistic model can be learned to accurately imitate human authors’ pattern of usage around verbs, outperforming the state-of-the-art method significantly. 1 Introduction Natural Language Generation (NLG) is a fundamen- tal task in Artificial Intelligence (AI) (Russell and Norvig, 2009). It aims to automatically turn struc- tured data into prose (Reiter, 2007; Belz and Kow, 2009) — the opposite of the better-known field of Natural Language Processing (NLP) that transforms raw text into structured data (e.g., a logical form or a knowledge base) (Jurafsky and Martin, 2009). Being dubbed “algorithmic authors” or “robot journalists”, NLG systems have attracted a lot of attention in re- cent years, thanks to the rise of big data (Wright, 2015). The use of NLG in financial services has been growing very fast. One particularly important NLG problem for summarizing financial or business data is to automatically generate textual descriptions of trends between two data points (such as stock prices). In this paper, we elect to use relative percentages rather than absolute numbers to describe the change from one data point to another. This is because an absolute number might be considered small in one case but large in another, depending on the unit and the context (Krifka, 2007; Smiley et al., 2016). For example, 1000 British pounds are worth much more than 1000 Japanese yen; a rise of 100 US dollars in car price might be negligible but the same amount of increase in bike price would be significant. Given two data points (e.g., on a stock chart), the percentage change can always be calculated easily. The challenge is to select the appropriate verb for any percentage change. For example, in newspa- pers, we often see headlines like “Apple’s stock had jumped 34% this year in anticipation of the next iPhone ...” and “Microsoft’s profit climbed 28% with shift to Web-based software ...”. The journal- ists writing such news stories use descriptive lan- guage such as the verbs like jump and climb to express the direction and magnitude of a percent- age change. It is of course possible to simply keep using the same neutral verbs, e.g., increase and decrease for upward and downward changes re- …

Content truncated to 500 words