Neural Networks and Information Retrieval in a Libraries Context Amsterdam , The Netherlands Friday June 24, 1994 and Friday September 16, 1994 M.S.C. Information Retrieval Technologies BV, based in Amsterdam, the Netherlands, is currently undertaking a study on Neural Networks and Information Retrieval in a Libraries Context, in collaboration with the Department of Computational Linguistics of the University of Amsterdam and the Department of Information Technology and Information Science at Amsterdam Polytechnic. This study is funded by the European Commission as a complementary measure under the Libraries Programme In this study the general application of artificial neural net (ANN) technology to information retrieval (IR) problems is investigated in a libraries context. Typical applications of this technology are advanced interface design, current awareness, SDI, fuzzy search and concept formation. In order to discuss and disseminate the results obtained through this study, two one-day workshops will be organized by M.S.C. Information Retrieval Technologies BV, the first one after compilation of the State of the Art Report and the second one after completion of the prototyping and experimentation phase. During both workshops, there will be much room for discussions on how to commercialise such applications of ANN in a libraries context. Both workshops are open to participants from other organizations, commercial and academic, that are interested in various applications of ANNs in existing libraries systems. For who: Interesting for all: - Computer Companies - Information Management and Supply Companies - Government Agencies - Libraries - Universities and Polytechniques That are Interested in: - Neural Networks - Information Retrieval - Libraries Sciences - Natural Language Processing - Advanced Computer Science - Data compression For applications such as: - Current Awareness - Selective Dissemination of Information (SDI) - Information Filtering - Automatic Contents Based Information Distribution - Categorization - Advanced Interface Design - Fuzzy Retrieval (Information recognized by Optical Character Recognition and Speech Recognition). - Retrieval Generalization - Thesaurus Generation - Information Compression - Juke box staging General Information Costs per participant for both days: Commercial companies Dfl. 950,- Universities and non-profit institutions (*) Dfl. 500,- Students (*) Dfl. 150,- (*) Letter of university or non-profit institution must be shown at registration These costs include: Workshop Proceedings State of the Art report on Neural Networks in Information Retrieval as composed by MSC Achievements report on Neural Networks in Information Retrieval as composed by MSC Ongoing coffee & tea Lunch Diner Future mailings on progress Limited availability of travel grants for students (please apply) All other expenses such as traveling, hotels, short stays, etc. are not included in the fee. Payment The following payment methods are accepted: 1. Credit Cards 2. Prepayment by bank 3. Personal cheques More information: M.S.C. Information Retrieval Technologies BV Dr Johannes C. Scholtes Dufaystraat 1 1075 GR AMSTERDAM the Netherlands Telephone: +31 20 679 4273 Fax: +31 20 6710 793 Internet: 100322.250@compuserve.com or scholtes@msc.mhs.compuserve.com Compuserve: MHS: SCHOLTES@MSC or 100322,250 Background & Introduction Recent research of artificial neural networks (ANN) in the field of pattern recognition and pattern classification applications has provided successful alternatives of traditional techniques. Products applied for optical character recognition (OCR), speech recognition, hand-written character recognition and prediction of non-linear time series are good examples of commercialization of these ANN techniques. So far, the European Commission has funded more than 40 projects of different sizes under the ESPRIT and other programmes which involve research on or the application of ANN technology. The task of Information Retrieval (IR), that is the matching of a large number of documents against a query, can also been seen as a pattern recognition or pattern classification task. Therefore, there have been several approaches to the application of ANN in IR in order to increase the quality of the retrieval process. Despite the theoretical and practical evidence that ANN are good tools for pattern recognition tasks, it is still an open question whether they are appropriate tools within the specific domain of Bibliographic Information Retrieval. Apart from some minor studies it seems no real attempt has been made up until now to integrate an ANN as a main component of a bibliographical information retrieval system or an on-line library catalogue (OPAC). It is therefore not clear whether and how ANN techniques can be combined with more "classical" methods, for instance rule-based or statistical approaches. By the same token it is not clear either to what extent existing OPACs could benefit from ANN technology. Objectives The objectives of this study are: to ascertain the State-of-the-Art of the application of Artificial Neural Net (ANN) technology to Information Retrieval (IR), with particular emphasis on bibliographic information in a libraries context; to assess the (potential) quality of ANN-based approaches to IR in this particular domain of interest, in comparison with traditional practices. Here "quality must be understood in terms of both (measurable) efficiency and practical benefits; to stimulate interest in the practical application of ANN technology to bibliographic information retrieval in a libraries context. Information Retrieval It can be stated that Information Retrieval (IR) is the ultimate combination between Natural Language Processing (NLP) and Artificial Intelligence (AI). On the one hand there is an enormous amount of NLP data that needs to be processed and understood to return the proper information to the user. On the other hand, one needs to understand what the user intends with his or her query given the context of the other queries and some kind of user model. Most of these systems still use techniques that were developed over thirty years ago and that implement nothing more than a global surface analysis of the textual (layout) properties. No deep structure whatsoever is incorporated in the decision to whether or not retrieve a text. There is one large dilemma in IR research. The data collections are so incredibly large, that any method other than a global surface analysis would fail. However, such a global analysis could never implement a contextually sensitive method to restrict the number of possible candidates returned by the retrieval system. Information retrieval can also be a very frustrating area of research. Whenever one invents a new model, it is difficult to show that it works better (qualitatively and quantitatively) than any previous model. The addition of new dependencies often results in much too slow a system. Systems such as Salton's SMART exist for over 30 years without having any serious competition. The field of information retrieval would be greatly indebted to a method that could incorporate more context without slowing down. Since computers are only capable of processing numbers within reasonable time limits, such a method should be based on vectors of numbers rather than on symbol manipulations. This is exactly where the challenge lies: on the one hand keep up the speed, and on the other incorporate more context. Artificial Neural Networks The connectionist approach offers a massively parallel, highly distributed and highly interconnected solution for the integration of various kinds of knowledge, with preservation of generality. It might be that connectionism or neural networks (despite all currently unsolved questions concerning learning, stability, recursion, firing rules, network architecture, etc.), will contribute to the research in natural-language processing and information retrieval. Distributed data representation may solve many of the unsolved problems in IR by introducing a powerful and efficient knowledge integration and generalization tool. However, distributed data representation and self-organization trigger new problems that should be solved in an elegant manner. Current Problems in Information Retrieval The main objectives of current IR research can be characterised as the search for systems that exhibit adaptive behaviour, interactive behaviour and transparency. More specifically, these models should implement properties for: Understanding incomplete queries or making incomplete matches, Understanding vague user intentions, Ability to generalise over queries as well as over query results, Adapting to the needs of an evolving user (model), Allowing dynamic relevance feed-back, Aid for the user to browse intelligently through the data, and Addition of (language) context sensitivity. Different Approaches in Information Retrieval and Neural NetworksTwo main directions of neural network related research information retrieval can be observed. First, there are relatively static databases that are investigated with a dynamic query (free text search, also known as document retrieval systems). Next, there are the more dynamic databases that need to be filtered with respect to a relatively static query (the filtering problem also known as current awareness systems and Selective Dissemination of Information, SDI). In the first case the data can be preprocessed due to their static character. In the second case, the amounts of data are so large that there is no time whatsoever for a preprocessing phase. A direct context-sensitive hit-and-go must be made. Early neural models adapt well to the paradigms currently used in information retrieval. Index terms can be replaced by processing units, hyperlinks by connections between units, and network training resembles the index normalisation process. However, these models do not adapt well to the general notion of neural networks. In addition, it is difficult to imagine what to teach a neural information retrieval system if it is used as a supervised training algorithm. The address space will almost always be too limited due to the large amounts of data to be processed. A combination of structured (query, retrieved document numbers) pairs does not seem plausible either, considering the restricted amount of memory of (current) neural network technology. Nevertheless, most of the neural IR models found in literature are based on these principles. Also problematic are the so-called clustering networks. Due to the large amounts of data in free text databases, clustering is very expensive and is therefore considered irrelevant in changing information retrieval environments. More interesting are the unsupervised, associative memory type of models, that can be used to implement a specific pattern matching task. This type of neural networks can be particularly useful in a filtering application. Here, the memory demands of the neural network only need to fulfil the query (or interest) size, and not the size of the entire data base. It is in this area where neural networks are expected to be most useful and relevant for information retrieval. Especially topics such as fuzzy retrieval, current awareness, SDI, concept formation and advanced interface design are in the scope of the project. However, input from the workshops is very important for the final determination of the direction of the research. Program Day 1: June 24, 1994 9.15-9.30 Welcome and Introduction Dr Ir Johannes C. Scholtes, President of MSC Information Retrieval Technologies B.V. 9.30-11.00 Tutorial Neural Networks (Back Propagation Kohonen Feature Maps) Dr Ir Johan Henseler, Forensic Laboratories, Head of Section Computer Criminality 11.00-11.15 Break 11.15-12.30 Information Retrieval Application in Libraries Dr E. Sieverts, Professor at Amsterdam Polytechnique. Library Program 12.30-13.30 Lunch 13.30-15.00 Presentation Findings & State of the Art Report 15.00-15.15 Break 15.15-16.00 Directions for (Commercial) Applications Dr ir Johannes C. Scholtes 16.00-17.00 Panel Discussion 17.00-18.00 Reception 19.00-... Diner and evening program Day 2: September 16, 1994 9.15-9.30 Welcome and Introduction Dr Ir Johannes C. Scholtes. President of MSC Information Retrieval Technologies B.V. 9.30-11.00 Achievements Dr Ir Johannes C. Scholtes. President of MSC Information Retrieval Technologies B.V. & Dr E. Sieverts. Professor at Amsterdam Polytechnique Library Program 11.00 - 12.30 Hands on demonstrations 12.30-13.30 Lunch 13.30-15.00 Problem Issues by Dr E. Sieverts. Professor at Amsterdam Polytechnique. Library Program 15.00-15.15 Break 15.15-16.00 Commercial Implications by Dr Ir Johannes C. Scholtes. President of MSC Information Retrieval Technologies B.V. 16.00-17.00 Panel Discussion 17.00-18.00 Reception 19.00-... Diner and evening program During the day, demo's of the prototypes will be available to the participants of the workshop. Each demo will be guided by a specialist who demonstrates the software