Multi-Resolution Language Grounding
and Ali Farhadi
Paper, Dataset, and Code
The paper is available here
. The Professional Soccer Commentary Dataset with both fragment- and utterance-level gold annotations is available here
. A rough draft of the code is available here
Language is given meaning through its correspondence with a world representation. This correspondence can be at multiple levels of granularity or resolutions. In this paper, we introduce an approach to multi-resolution language grounding in the extremely challenging domain of professional soccer commentaries. We define and optimize a factored objective function that allows us to leverage discourse structure and the compositional nature of both langauge 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 grounding
An example of the different levels of granularity present in the soccer data. The dashed boxes on the left denote utterances made by the commentators. Solid boxes denote fragments that cannot be decomposed into finer resolution alignments. The table on the right is a portion of the detailed listing of game events.
The research was supported by the Allen Institute for AI, and grants from the NSF (IIS-1352249) and UW-RRF (65-2775).
R. Koncel-Kedziorski, Hannaneh Hajishirzi, and Ali Farhadi. Multi-resolution Language Grounding with Weak Supervision. EMNLP, 2014.