Focal points of the DRUID project

Speech Recognition

The long term objective is to enable the generation of textual representations of the speech elements that are part of the documents to be disclosed. The transcripts resulting from speech recognition can be used as input for indexing tools that are enhanced with natural language processing techniques. Particularly interesting is the fact that this approach enables the disclosure and retrieval of fragments of continuous objects. The envisaged results will therefore be of interest to radio and video archives (the broadcast world) and to archives for the storage of logged dialogues (e.g. from telephone conversations or video conference sessions). Other short-term project spin-offs are expected that can be integrated in user interfaces (as asked for by e.g. the financial world).

Cross Language Retrieval

With the gradual disappearing of the distinction between 'local' documents and documents anywhere on the web, the possibility to handle documents in several languages becomes of primary importance. Part of this functionality has recently been introduced by AltaVista. The kind of search engines DRUID intends to make available for Internet and/or Intranet applications will provide several tools. Producing shallow translations of foreign language documents without relying on full machine translation systems seems feasible, given a high quality lexical database like Van Dale, and statistically aligned parallel or comparable corpora for domain applications. TREC 6 (IR-benchmarking competition) has shown that this approach is promising.

Information Filtering & Extraction

This theme aims at the development of techniques for information filtering, structuring, discovery and navigation. Personalization or tailoring of information delivery is the key concept. Several application types will be improved or developed, partly on the basis of so-called "deep NLP" techniques. Research will be set up in the following areas:

Categorisation. Assignment of metadata to documents. Classification and clustering methods based on established similarity metrics can be applied to generate taxonomies of related terms and related documents. Navigation tools based on these techniques will be developed and evaluated in order to support the needs for advanced search & discovery techniques.
Agent technology. Can be deployed to support information delivery given a user profile. Agents can inspect a document base or an Intranet (the latter with added hyperlink structure) and could classify documents given a predefined set of classes or a user profile, automatically or with user relevance feedback.
Summary Generation. Document summaries play an important role in relevance assessment of text. The summaries can also serve as input for the partial translation module to help assessing the relevance of foreign language documents. Classical statistical sentence-extraction techniques will be evaluated, but also other more shallow document summaries will be evaluated. It will be investigated whether combining information from document structure (title/heading), specific lexical elements, statistical measures (tf.idf) and robust pattern filling can lead to adequate executive summaries or news skims.

Video Segmentation

For the isolation of textual material and other features from a video document, various types of video processing are required. Some of these techniques are already commercially available, for others prototypes are available at TNO and/or CTIT. In addition to the preprocessing task for text-based video retrieval there are several other application areas for these techniques such as summarizing these techniques can also be extended to video summarizing.
To adequately define the research tracks for 1999 and onwards, a study of the added value of the various techniques for the disclosure of multi-media archives, and an assessment of the feasibility and integratability of available tools are needed.

Database Integration

In terms of scope and potential usability, DRUID has a considerable overlap with the Digital Multimedia Warehouse project which is coordinated by CWI. Joint experiments will be set up by CWI and TNO-TPD to assess the performance of (some of) the modules from TNO's multimedia retrieval workbench on a huge database and to generate a perspective on novel routes in combining NLP and Image Processing.
Scalable NLP based indexing architectures will also be studied in the particular setting of standard word based indices like Oracle Context.

Last modified April 2001 by Thijs Westerveld