How Much is 131 Million Dollars? Putting Numbers in Perspective with Compositional Descriptions

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 578–587, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics How Much is 131 Million Dollars? Putting Numbers in Perspective with Compositional Descriptions Arun Tejasvi Chaganty Computer Science Department Stanford University chaganty@cs.stanford.edu Percy Liang Computer Science Department Stanford University pliang@cs.stanford.edu Abstract How much is 131 million US dollars? To help readers put such numbers in con- text, we propose a new task of automati- cally generating short descriptions known as perspectives, e.g. “$131 million is about the cost to employ everyone in Texas over a lunch period”. First, we collect a dataset of numeric mentions in news arti- cles, where each mention is labeled with a set of rated perspectives. We then pro- pose a system to generate these descrip- tions consisting of two steps: formula con- struction and description generation. In construction, we compose formulae from numeric facts in a knowledge base and rank the resulting formulas based on fa- miliarity, numeric proximity and seman- tic compatibility. In generation, we con- vert a formula into natural language us- ing a sequence-to-sequence recurrent neu- ral network. Our system obtains a 15.2% F1 improvement over a non-compositional baseline at formula construction and a 12.5 BLEU point improvement over a baseline description generation. 1 Introduction When posed with a mention of a number, such as “Cristiano Ronaldo, the player who Madrid ac- quired for [. . . ] a $131 million” (Figure 1), it is often difficult to comprehend the scale of large (or small) absolute values like $131 million (Paulos, 1988; Seife, 2010). Studies have shown that pro- viding relative comparisons, or perspectives, such as “about the cost to employ everyone in Texas over a lunch period” significantly improves com- prehension when measured in terms of memory re- tention or outlier detection (Barrio et al., 2016). Cristiano Ronaldo, the player who Madrid acquired for [. . . ] $131 million. $1.3e8 ≈ 71e3 $/per/yr � �� � 1 × 27e6 per � �� � 2 × 30 min � �� � 3 1 = cost of an employee 2 = population of Texas 3 = time taken for lunch about the cost to employ everyone in Texas over a lunch period. construction generation mention formula perspective Figure 1: An overview of the perspective gen- eration task: given a numeric mention, generate a short description (a perspective) that allows the reader to appreciate the scale of the mentioned number. In our system, we first construct a for- mula over facts in our knowledge base and then generate a description of that formula. Previous work in the HCI community has relied on either manually generated perspectives (Barrio et al., 2016) or present a fact as is from a knowl- edge base (Chiacchieri, 2013). As a result, these approaches are limited to contexts in which a rele- vant perspective already exists. In this paper, we generate perspectives by com- posing facts from a knowledge base. For example, we might describe …

Content truncated to 500 words