Data-driven Approaches for Information Structure Identification

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 9–16, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics Data-driven Approaches for Information Structure Identification Oana Postolache, Ivana Kruijff-Korbayov´a University of Saarland, Saarbr¨ucken, Germany {oana,korbay}@coli.uni-saarland.de Geert-Jan M. Kruijff German Research Center for Artificial Intelligence (DFKI GmbH) Saarbr¨ucken, Germany gj@dfki.de Abstract This paper investigates automatic identi- fication of Information Structure (IS) in texts. The experiments use the Prague Dependency Treebank which is annotated with IS following the Praguian approach of Topic Focus Articulation. We auto- matically detect t(opic) and f(ocus), us- ing node attributes from the treebank as basic features and derived features in- spired by the annotation guidelines. We present the performance of decision trees (C4.5), maximum entropy, and rule in- duction (RIPPER) classifiers on all tec- togrammatical nodes. We compare the re- sults against a baseline system that always assigns f(ocus) and against a rule-based system. The best system achieves an ac- curacy of 90.69%, which is a 44.73% im- provement over the baseline (62.66%). 1 Introduction Information Structure (IS) is a partitioning of the content of a sentence according to its relation to the discourse context. There are numerous theo- retical approaches describing IS and its semantics (Halliday, 1967; Sgall, 1967; Vallduv´ı, 1990; Steed- man, 2000) and the terminology used is diverse — see (Kruijff-Korbayov´a and Steedman, 2003) for an overview. However, all theories consider at least one of the following two distinctions: (i) a Topic/Focus1 distinction that divides the linguistic meaning of the sentence into parts that link the sentence content  We use the Praguian terminology for this distinction. to the discourse context, and other parts that ad- vance the discourse, i.e., add or modify informa- tion; and (ii) a background/kontrast2 distinction be- tween parts of the utterance which contribute to dis- tinguishing its actual content from alternatives the context makes available. Information Structure is an important factor in de- termining the felicity of a sentence in a given con- text. Applications in which IS is crucial are text- to-speech systems, where IS helps to improve the quality of the speech output (Prevost and Steedman, 1994; Kruijff-Korbayov´a et al., 2003; Moore et al., 2004), and machine translation, where IS improves target word order, especially that of free word order languages (Stys and Zemke, 1995). Existing theories, however, state their principles using carefully selected illustrative examples. Be- cause of this, they fail to adequately explain how different linguistic dimensions cooperate to realize Information Structure. In this paper we describe data-driven, machine learning approaches for automatic identification of Information Structure; we describe what aspects of IS we deal with and report results of the performance of our systems and make an error analysis. For our experiments, we use the Prague Dependency Tree- bank (PDT) (Hajiˇc, 1998). PDT follows the theory of Topic-Focus Articulation (Hajiˇcov´a et al., 1998) and to date is the only corpus annotated with IS. Each node of the underlying structure of sentences in PDT is annotated with a …

Content truncated to 500 words