Logical-Linguistic Model and Experiments in Document Retrieval

[摘要] Conventional document retrieval systems have relied on the extensive use of the keyword approach with statistical parameters in their implementations. Now, it seems that such an approach has reached its upper limit of retrieval effectiveness, and therefore, new approaches should be investigated for the development of future systems. With current advances in hardware, programming languages and techniques, natural language processing and understanding, and generally, in the field of artificial intelligence, there are now attempts being made to include linguistic processing into document retrieval systems. Few attempts have been made to include parsing or syntactic analysis into document retrieval systems, and the results reported show some improvements in the level of retrieval effectiveness. The first part of this thesis sets out to investigate further the use of linguistic processing by including translation, instead of only parsing, into a document retrieval system. The translation process implemented is based on unification categorial grammar and uses C-Prolog as the building tool. It is used as the main part of the indexing process of documents and queries into a knowledge base predicate representation. Instead of using the vector space model to represent documents and queries, we have used a kind of knowledge base model which we call logical-linguistic model. A development of a robust parser-translator to perform the translation is discussed in detail in the thesis. A method of dealing with ambiguity is also incorporated in the parser-translator implementation. The retrieval process of this model is based on a logical implication process implemented in C-Prolog. In order to handle uncertainty in evaluating similarity values between documents and queries, meta level constructs are built upon the C-Prolog system. A logical meta language, called UNIL (UNcertain Implication Language), is proposed for controlling the implication process. Using UNIL, one can write a set of implication rules and thesaurus to define the matching function of a particular retrieval strategy. Thus, we have demonstrated and implemented the matching operation between a document and a query as an inference using unification. An inference from a document to a query is done in the context of global information represented by the implication rules and the thesaurus. A set of well structured experiments is performed with various retrieval strategies on a test collection of documents and queries in order to evaluate the performance of the system. The results obtained are analysed and discussed. The second part of the thesis sets out to implement and evaluate the imaging retrieval strategy as originally defined by van Rijsbergen. The imaging retrieval is implemented as a relevance feedback retrieval with nearest neighbour information which is defined as follows. One of the best retrieval strategies from the earlier experiments is chosen to perform the initial ranking of the documents, and a few top ranked documents will be retrieved and identified as relevant or not by the user. From this set of retrieved and relevant documents, we can obtain all other unretrieved documents which have any of the retrieved and relevant documents as their nearest neighbour. These unretrieved documents have the potential of also being relevant since they are 'close' to the retrieved and relevant ones, and thus their initial similarity values to the query will be updated according to their distances from their nearest neighbours. From the updated similarity values, a new ranking of documents can be obtained and evaluated. A few sets of experiments using imaging retrieval strategy are performed for the following objectives: to search for an appropriate updating function in order to produce a new ranking of documents, to determine an appropriate nearest neighbour set, to find the relationship of the retrieval effectiveness to the size of the documents shown to the user for relevance judgement, and lastly, to find the effectiveness of a multi-stage imaging retrieval. The results obtained are analysed and discussed. Generally, the thesis sets out to define the logical-linguistic model in document retrieval and demonstrates it by building an experimental system which will be referred to as SILOL (a Simple Logical-linguistic document retrieval system). A set of retrieval strategies will be experimented with and the results obtained will be analysed and discussed.

[发布机构] University:University of Glasgow

[关键词] Computer science