Per-Anders Jande, KTH: Data-driven pronunciation variation modelling: annotation, induction and evaluation

The pronunciation of a word depends on the context in which the word is uttered. A model of pronunciation variation due to discourse context is interesting in a description of a language variety. Such a model can also be used to increase the naturalness of synthetic speech and to dynamically adapt synthetic speech to different areas of use and to different speaking styles. This seminar will include the presentation and discussion of a number of issues associated with the development of a pronunciation variation model from recorded speech data: 1) methods for obtaning context parameters for pronunciation variation modelling and a system for annotation 2) methods for model induction from context annotation and 3) methods and measures for evaluation of model prediction accuracy.

Mustapha Skhiri, LiU: Some investigations of facial and head movements in interactive dialogue systems

Dialogue is an interactive communication of information mainly based on speech, but also visual information such as gesture, facial expression and head movement clearly makes the conversations much smoother and more natural. In my coming seminar I will speak about my present research, firstly details the methodology that was designed and utilised during my investigations into this relatively unexplored area of quantifying facial gestures. Secondly I'll mention the process to capture a continuous natural dialogue and finally I will present some results.

Atelach Alemu, Stockholms universitet

In this presentation, I will discuss experiments and results from our participation in last year's Cross Language Evaluation Forum (CLEF 2004) where we participated in the bilingual run using Amharic queries to the English document collection.

I will then give an overview of our plans for participation in CLEF 2005. This year, our cross lingual experiments will be extended with the inclusion of Swedish (and possibly one other European language). Since resources for the language pair Amharic-Swedish are much less developed than they are for Amharic-English and English-Swedish, we intend to use English as a pivot language that will serve as the intermediary between Amharic and Swedish. I will present the challenges faced in the bilingual Amharic-English experiments, and describe how these challenges will be intensified in the Amharic-English-Swedish pivot CLIR. I will propose methods and techniques to handle these issues. Particularly I will discuss issues concerning out-of-dictionary words, and word sense disambiguation. We intend to use edit-distance based fuzzy matching to handle out-of-dictionary words (only proper names) and mutual information based word sense disambiguation for MRD lookup query translation.

Eva Forsbom: Rhetorical structure analysis as a basis for summarisation in information retrieval

Would you appreciate more coherent summaries when using information systems? Maybe a rhetorical structure analysis might help...

In information systems, some kind of content analysis of the texts in the system is made. Furthermore, on the basis of the analysis, a summary of each text found relevant for a search query is often created and shown as as a guide for relevance judgement. of the hits. The analyses are usually based on the authors' choice of wording, but more seldom on their way of presenting the content, e.g. by making some information bits more salient than others. A rhetorical structure analysis (Mann & Thompson 1988) should therefore give clues to what information bits the authors consider important. This could give a good generic summary.

But as a reader, you might have other information needs than the authors had in mind. What is less important to them might be important to you. For that purpose, a summary based on your information need is needed. Such summaries are already present in some systems, but they are seldom very coherent since bits and pieces from the original text are cut out and pasted together without much concern for their original context. With a rhetorical structure analysis, on the other hand, it should be possible to pay more attention to the original context, thus providing for a coherent specific summary based on your information need.

In the seminar, a short introduction to some rhetorically based methods for summarisation will be given. A presentation of my thesis work, which is just started, will also be given. The thesis topic is text linguistic methods in summarisation and information access in Swedish, where one of the methods will be a rhetorical structure analysis. In the presentation, there will also be a discussion on the kind of knowledge needed for making such an analysis, and how the knowledge could be extracted from annotated corpora with machine learning methods.

Reference:
William C. Mann and Sandra A. Thompson. 1988. Rhetorical Structure Theory: Toward a Functional Theory of Text Organization. Text 8:3, pp 243-281.

Leif Grönqvist: Evaluation of Latent Semantic Vector Models using a Swedish Word comprehension test

I will present an evaluation set (HP200) based on the word comprehension part of the Swedish "Högskoleprovet", a qualification test for future university students, and an evaluation of some vector models trained with different corpora and parameter settings, using latent semantic indexing and singular value decomposition for the dimension reduction.

One way to use HP200 to evaluate a vector model is to calculate vectors for each question and its alternatives. The alternative vector closest to the question vector is selected as an answer which is compared to the correct answer. How well could this work? I will show the current results.

Mikael Gunnarsson: From Document Structure Disambiguation to Genre Identification

The theoretical foundation for this approach to genre identification is that genre is a sociotechnical dynamic phenomenon that can be characterised at least at three levels: a) a discourse community (or domain), b) a communicative situation, and c) form and content of its artefacts (or documents). In information seeking contexts not only the topical content is of great importance for the user. On the contrary, in some cases the genre, or the type of document, is equally important in order to sort out obviously irrelevant documents, or at least as a way to refine the visualization of a large set of topically retrieved documents.

Research on computational document genre identification have mostly focused on document features derived from linguistic expressions irrespective of its structural placement. With a few exceptions, applied genre research in general have not recognized document structures realised in different kinds of markup schemes to any considerable extent as indicators on genre adherence. Moreover, web technology and its distributional and remediating character generate emergent and innovative genres that extend and transform the genre space to a considerable degree, thereby calling for a reevaluation of existing classification schemes and methods.

If existing markup is to be taken account of as cues to any kind of genre identification, the markup itself needs to be analysed in order to identify its stuctural semantics. This presentation focuses on how existing markup can be mapped to a set of predefined "genre modules" which then can be used for similarity clustering of documents based on structural (and linguistic) document features.