Probabilistic Verb Selection For Enhanced Data-to-Text Generation.pdf

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 direction and magnitude of a percentage change. Rather than simply using the same verbs again and again, we present a principled data-driven approach to this problem based on Shannon's noisy-channel model to bring variation and naturalness into the generated text. Our experiments on three large-scale real-world news corpora demonstrate 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.

Natural Language Generation (NLG) is a fundamental task in Artificial Intelligence (AI) that aims to automatically turn structured data into prose. The use of NLG in financial services has been growing very fast, and 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, as absolute numbers might be considered small in one case but large in another, depending on the unit and context.

The challenge is to select the appropriate verb for any percentage change. For example, in newspapers, 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 journalists writing such news stories use descriptive language such as the verbs "jump" and "climb" to express the direction and magnitude of a percentage change. It is possible to simply keep using the same neutral verbs, such as "increase" and "decrease", again and again, as in most existing data-to-text NLG systems. However, the generated text would sound much more natural if computers could use a variety of verbs suitable in the context like human authors do.

Expressions of percentage changes are readily available in many natural language text datasets and can be easily extracted. Therefore, computers should be able to learn from such expressions how people decide which verbs to use for what kind of percentage changes. In this paper, we address the problem of verb selection for data-to-text NLG through a principled data-driven approach. We show how to employ Bayesian reasoning to train a probabilistic model for verb selection based on large-scale real-world news corpora, and demonstrate its advantages over existing verb selection methods.

The proposed probabilistic model is based on Shannon's noisy-channel model and is trained on three large-scale real-world news corpora. The model is able to learn the pattern of usage around verbs and accurately imitate human authors' usage. The experiments demonstrate that the proposed model outperforms the state-of-the-art method significantly.