Bimodal Corpora Terminology Extraction: Another Brick in the Wall

International Conference RANLP 2009 - Borovets, Bulgaria, pages 236–240 Bimodal Corpora Terminology Extraction: Another Brick in the Wall Claudiu Mih˘ail˘a∗ University of Wolverhampton Wulfruna Street Wolverhampton WV1 1LY claudiu.mihaila@info.uaic.ro Dalila Mekhaldi University of Wolverhampton Wulfruna Street Wolverhampton WV1 1LY dalila.mekhaldi@wlv.ac.uk Abstract This paper presents a new study on automatic terminology extraction in the context of bimodal corpora that were generated from lectures and meetings. More specifically, the study aims to observe to which extent written text (dis- cussed documents) and spoken text (dialogue transcript) share keywords. Using a hybrid ter- minology extraction approach, experiments have been performed on a collection of bimodal En- glish corpora, including one scientific conference presentations corpus and two decision-making meetings corpora respectively. The evaluation results highlight a difference between keywords extracted from written text and from spoken text. Moreover, the obtained results emphasise the importance of considering text obtained from different modalities in order to generate rich and consistent keyword lists for bimodal corpora. Keywords bimodal corpus, keyword extraction, keyword similarity, spoken document, written document. 1 Introduction Corpora, which are defined as large bodies of linguis- tic evidence composed of attested language use [9], are increasingly used in natural language processing field(nlp), such as in machine translation, automatic summarisation, etc. However, the diversity of nlp tasks that are based on corpora is proportional to the variousness of types of the latter. Therefore many cor- pora types have been distinguished [9], mainly mono- lingual corpora, parallel corpora and comparable cor- pora. Whilst most of the existing studies are based on these corpora types, bimodal corpora, which are origi- nally built in the context of multimodal and multime- dia applications, have recently appeared as a new cor- pora type that needs to be studied. Bimodal corpora are defined as pairs of texts either used or generated during multimedia events (e.g. meetings or lectures), each of them obtained through a different modality (e.g. text of written documents, speech transcript, text extracted from video recording frames). ∗The author is also affiliated to the Faculty of Computer Science, ”Al.I. Cuza” University of Ia¸si, Romania From another side, terminology extraction (te) con- sists in the identification and retrieval of the impor- tant lexical units, i.e. keywords, from a corpus. One of the challenges in te is the definition of a keyword, which greatly depends on the context of the applica- tion. Other challenges are morphological and lexical variation of keywords, the consideration of single-word and multi-word keywords, etc. Once these challenges are addressed, the extracted keywords are particularly useful for conceptualising a knowledge domain or for supporting the creation of domain ontologies, due to their high specificity and low ambiguity. The accuracy in the identification of keywords may influence nlp tasks like analysis, understanding, generation, trans- lation, automatic summarisation [12], and multiple- choice test item generation [5] of textual documents. Terminology extraction from bimodal corpora, which is studied in this paper, represents the basis of many nlp and nlp-related fields of study. The task of multimodal document alignment framework, defined in …

Content truncated to 500 words