Interpretation and Transformation for Abstracting Conversations

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 894–902, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Interpretation and Transformation for Abstracting Conversations Gabriel Murray gabrielm@cs.ubc.ca Giuseppe Carenini carenini@cs.ubc.ca Department of Computer Science, University of British Columbia Vancouver, Canada Raymond Ng rng@cs.ubc.ca Abstract We address the challenge of automatically ab- stracting conversations such as face-to-face meetings and emails. We focus here on the stages of interpretation, where sentences are mapped to a conversation ontology, and transformation, where the summary content is selected. Our approach is fully developed and tested on meeting speech, and we subse- quently explore its application to email con- versations. 1 Introduction The dominant approach to the challenge of auto- matic summarization has been extraction, where in- formative sentences in a document are identified and concatenated to form a condensed version of the original document. Extractive summarization has been popular at least in part because it is a binary classification task that lends itself well to machine learning techniques, and does not require a natural language generation (NLG) component. There is ev- idence that human abstractors at times use sentences from the source documents nearly verbatim in their own summaries, justifying this approach to some ex- tent (Kupiec et al., 1995). Extrinsic evaluations have also shown that, while extractive summaries may be less coherent than human abstracts, users still find them to be valuable tools for browsing documents (He et al., 1999; McKeown et al., 2005; Murray et al., 2008). However, these same evaluations also indicate that concise abstracts are generally preferred by users and lead to higher objective task scores. The limitation of a cut-and-paste summary is that the end-user does not know why the selected sentences are important; this can often only be discerned by exploring the context in which each sentence origi- nally appeared. One possible improvement is to cre- ate structured extracts that represent an increased level of abstraction, where selected sentences are grouped according to phenomena such as decisions, action items and problems, thereby giving the user more information on why the sentences are being highlighted. For example, the sentence Let’s go with a simple chip represents a decision. An even higher level of abstraction can be provided by generating new text that synthesizes or extrapolates on the in- formation contained in the structured summary. For example, the sentence Sandra and Sue expressed negative opinions about the remote control design can be coupled with extracted sentences containing these negative opinions, forming a hybrid summary. Our summarization system ultimately performs both types of abstraction, grouping sentences according to various sentence-level phenomena, and generat- ing novel text that describes this content at a higher level. In this work we describe the first two components of our abstractive summarization system. In the in- terpretation stage, sentences are mapped to nodes in a conversation ontology by utilizing classifiers relating to a variety of sentence-level phenomena such as decisions, action items and subjective sen- tences. These classifiers …

Content truncated to 500 words