Learning Script Participants from Unlabeled Data

Proceedings of Recent Advances in Natural Language Processing, pages 463–470, Hissar, Bulgaria, 12-14 September 2011. Learning Script Participants from Unlabeled Data Michaela Regneri ∗ Alexander Koller † Josef Ruppenhofer ‡ Manfred Pinkal ∗ ∗ Department of Computational Linguistics, Saarland University {regneri,pinkal}@coli.uni-saarland.de † Department of Linguistics, University of Potsdam koller@ling.uni-potsdam.de ‡ Department of Information Science and Language Technology, University of Hildesheim Josef.Ruppenhofer@uni-hildesheim.de Abstract We introduce a system that learns the par- ticipants of arbitrary given scripts. This system processes data from web experi- ments, in which each participant can be re- alized with different expressions. It com- putes participants by encoding semantic similarity and global structural informa- tion into an Integer Linear Program. An evaluation against a gold standard shows that we significantly outperform two in- formed baselines. 1 Introduction Scripts (Schank and Abelson, 1977) represent commonsense knowledge about the events that stereotypically constitute a certain activity. For in- stance, the “restaurant” script might specify that the patron enters, the waiter shows the patron to their seat, eventually the patron eats a plate of food, and so forth. There has always been agree- ment that script knowledge can be highly use- ful for a variety of applications in artificial intel- ligence and computational linguistics, including commonsense reasoning for text understanding (Cullingford, 1977; Mueller, 2004), information extraction (Rau et al., 1989) and automated sto- rytelling (Swanson and Gordon, 2008). But there is hardly an area where the discrepancy between the felt importance of a type of knowledge and the inability to provide any substantial amount of this knowledge for serious applications is greater. Recently, several groups have tackled the prob- lem using unsupervised methods for learning script-like knowledge from text corpora or data obtained through web experiments (Chambers and Jurafsky, 2008; Chambers and Jurafsky, 2009; Regneri et al., 2010). For the first time, they open up a perspective to wide-coverage resources of script knowledge. However, each of these ap- proaches handles only specific aspects of script information: Chambers and Jurafsky (2009) learn narrative schemas and their participants; they group verbs into schemas by virtue of shared par- ticipants assuming that this is an indicator for be- ing part of the same stereotypical activity, with- out knowing the actual scenarios. The system of Regneri et al. (2010) learns the temporal order of events occurring in specific stereotypical scenar- ios, but does not determine participants. In this paper, we present a system that automat- ically learns sets of participants associated with specific scenarios. We take the approach of Reg- neri et al. as our starting point. In this earlier work, several experimental subjects described what hap- pens in a given scenario in a web experiment; the system then learns what event descriptions from different subjects refer to the same event, and how they are temporally ordered, using Multiple Se- quence Alignment (Durbin et al., 1998). The spe- cific problem we consider is to group the differ- ent noun phrases occurring throughout a script into equivalence classes, resulting in one class for each participant. Our …

Content truncated to 500 words