2011 Presentations (Communicative Events)

Towards Strict Sentence Intersection: Decoding and Evaluation Strategies

McKeown, Kathleen; Thadani, Kapil

We examine the task of strict sentence intersection: a variant of sentence fusion in which the output must only contain the information present in all input sentences and nothing more. Our proposed approach involves alignment and generalization over the input sentences to produce a generation lattice; we then compare a standard search-based approach for decoding an intersection from this lattice to an integer linear program that preserves aligned content while minimizing the disfluency in interleaving text segments. In addition, we
introduce novel evaluation strategies for intersection problems that employ entailment-style judgments for determining the validity
of system-generated intersections. Our experiments show that the proposed models produce valid intersections a majority of the time and
that the segmented decoder yields advantages over the search-based approach.