A Normalized-Cut Alignment Model for Mapping Hierarchical Semantic Structures onto Spoken Documents

Proceedings of the Fifteenth Conference on Computational Natural Language Learning, pages 210–218, Portland, Oregon, USA, 23–24 June 2011. c⃝2011 Association for Computational Linguistics A Normalized-Cut Alignment Model for Mapping Hierarchical Semantic Structures onto Spoken Documents Xiaodan Zhu Institute for Information Technology National Research Council Canada Xiaodan.Zhu@nrc-cnrc.gc.ca Abstract We propose a normalized-cut model for the problem of aligning a known hierarchical browsing structure, e.g., electronic slides of lecture recordings, with the sequential tran- scripts of the corresponding spoken docu- ments, with the aim to help index and access the latter. This model optimizes a normalized- cut graph-partitioning criterion and considers local tree constraints at the same time. The ex- perimental results show the advantage of this model over Viterbi-like, sequential alignment, under typical speech recognition errors. 1 Introduction Learning semantic structures of written text has been studied in a number of specific tasks, which include, but not limited to, those finding semantic represen- tations for individual sentences (Ge and Mooney, 2005; Zettlemoyer and Collins, 2005; Lu et al., 2008), and those constructing hierarchical structures among sentences or larger text blocks (Marcu, 2000; Branavan et al., 2007). The inverse problem of the latter kind, e.g., aligning certain form of already- existing semantic hierarchies with the corresponding text sequence, is not so much a prominent problem for written text as it is for spoken documents. In this paper, we study a specific type of such a problem, in which a hierarchical browsing structure, i.e., elec- tronic slides of oral presentations, have already ex- isted, the goal being to impose such a structure onto the transcripts of the corresponding speech, with the aim to help index and access spoken documents as such. Navigating audio documents is often inherently much more difficult than browsing text; an obvi- ous solution, in relying on human beings’ ability to read text, is to conduct a speech-to-text conversion through automatic speech recognition (ASR). Im- plicitly, solutions as such change the conventional speaking-for-hearing construals: now speech can be read through its transcripts, though, in most cases, it was not intended for this purpose, which in turn raises a new set of problems. The convenience and efficiency of reading tran- scripts (Stark et al., 2000; Munteanu et al., 2006) are first affected by errors produced in transcrip- tion channels for various reasons, though if the goal is only to browse salient excerpts, recognition er- rors on the extracts can be reduced by consider- ing ASR confidence scores (Xie and Liu, 2010; Hori and Furui, 2003; Zechner and Waibel, 2000): trading off the expected salience of excerpts with their recognition-error rate could actually result in the improvement of excerpt quality in terms of the amount of important content being correctly pre- sented (Zechner and Waibel, 2000). Even if transcription quality were not a problem, browsing transcripts is not straightforward. When intended to be read, written documents are almost always presented as more than uninterrupted strings of text. Consider that for many written docu- ments, e.g., books, indicative structures such as sec- tion/subsection …

Content truncated to 500 words