Multi-Resolution Language Grounding with Weak Supervision

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 386–396, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Multi-Resolution Language Grounding with Weak Supervision R. Koncel-Kedziorski, Hannaneh Hajishirzi, and Ali Farhadi University of Washington {kedzior,hannaneh,farhadi}@washington.edu Abstract Language is given meaning through its correspondence with a world representa- tion. This correspondence can be at mul- tiple levels of granularity or resolutions. In this paper, we introduce an approach to multi-resolution language grounding in the extremely challenging domain of pro- fessional soccer commentaries. We define and optimize a factored objective function that allows us to leverage discourse struc- ture and the compositional nature of both language and game events. We show that finer resolution grounding helps coarser resolution grounding, and vice versa. Our method results in an F1 improvement of more than 48% versus the previous state of the art for fine-resolution grounding1. 1 Introduction Language is inextricable from its context. A hu- man language user interprets an utterance in the context of, among other things, their perception of the world. Grounded language acquisition algo- rithms imitate this setup: language is given mean- ing through its correspondence with a rich world representation. A solution to the acquisition prob- lem must resolve several ambiguities: the seg- mentation of the text into meaningful units (spans of words that refer to events); determining which events are being referenced; and finding the proper alignment of events to these units. Historically, language grounding was only pos- sible over simple controlled domains and rigidly structured language. Current research in grounded 1Source code and data are available at http://ssli. ee.washington.edu/tial/projects/multires/ Figure 1: An example of the multiple resolutions at which soccer commentaries refer to events: The utterance level alignments are shown in the black dashed boxes. The first utterance can be further broken into the fragment-level align- ments shown; the second cannot be decomposed further. language acquisition is moving into real-world en- vironments (Yu and Siskind, 2013). Grounding sports commentaries in game events is a specific instance of this problem that has attracted attention (Liang et al., 2009; Snyder and Barzilay, 2007; Hajishirzi et al., 2012), in part because of the com- plexity of both the language and the world repre- sentation involved. The language employed in soccer commentaries is difficult to ground due to its dense informa- tion structure, novel vocabulary and word senses, and colorful, non-traditional syntax. These chal- lenges conspire to foil most language processing techniques including automated parsers and word- sense disambiguation systems. In addition to the structural problems presented by the language of soccer commentaries, the prob- lem of reference is further complicated by the fact that for game events (and other real-world phe- nomena) there is no standardized meaningful lin- guistic unit. Utterances ranging from a single word to multiple sentences can be used to refer to a single event. For example, in Figure 1 the first four words of commentary (I) refer to a single event, as does the entirety of (II). 386 Figure 2: …

Content truncated to 500 words