A Simple Domain-Independent Probabilistic Approach to Generation

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 502–512, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics A Simple Domain-Independent Probabilistic Approach to Generation Gabor Angeli UC Berkeley Berkeley, CA 94720 gangeli@berkeley.edu Percy Liang UC Berkeley Berkeley, CA 94720 pliang@cs.berkeley.edu Dan Klein UC Berkeley Berkeley, CA 94720 klein@cs.berkeley.edu Abstract We present a simple, robust generation system which performs content selection and surface realization in a unified, domain-independent framework. In our approach, we break up the end-to-end generation process into a se- quence of local decisions, arranged hierar- chically and each trained discriminatively. We deployed our system in three different domains—Robocup sportscasting, technical weather forecasts, and common weather fore- casts, obtaining results comparable to state-of- the-art domain-specific systems both in terms of BLEU scores and human evaluation. 1 Introduction In this paper, we focus on the problem of generat- ing descriptive text given a world state represented by a set of database records. While existing gen- eration systems can be engineered to obtain good performance on particular domains (e.g., Dale et al. (2003), Green (2006), Turner et al. (2009), Re- iter et al. (2005), inter alia), it is often difficult to adapt them across different domains. Further- more, content selection (what to say: see Barzilay and Lee (2004), Foster and White (2004), inter alia) and surface realization (how to say it: see Ratna- parkhi (2002), Wong and Mooney (2007), Chen and Mooney (2008), Lu et al. (2009), etc.) are typically handled separately. Our goal is to build a simple, flexible system which is domain-independent and performs content selection and surface realization in a unified framework. We operate in a setting in which we are only given examples consisting of (i) a set of database records (input) and (ii) example human-generated text de- scribing some of those records (output). We use the model of Liang et al. (2009) to automatically induce the correspondences between words in the text and the actual database records mentioned. We break up the full generation process into a se- quence of local decisions, training a log-linear clas- sifier for each type of decision. We use a simple but expressive set of domain-independent features, where each decision is allowed to depend on the en- tire history of previous decisions, as in the model of Ratnaparkhi (2002). These long-range contextual dependencies turn out to be critical for accurate gen- eration. More specifically, our model is defined in terms of three types of decisions. The first type chooses records from the database (macro content selection)—for example, wind speed, in the case of generating weather forecasts. The second type chooses a subset of fields from a record (micro con- tent selection)—e.g., the minimum and maximum temperature. The third type chooses a suitable tem- plate to render the content (surface realization)— e.g., winds between [min] and [max] mph; templates are automatically extracted from training data. We tested our approach in three domains: ROBOCUP, for sportscasting (Chen and Mooney, 2008); SUMTIME, for technical weather forecast …

Content truncated to 500 words