Formal concept matching and reinforcement learning in adaptive information retrieval
The superiority of the human brain in information retrieval (IR) tasks seems to come firstly from its ability to read and understand the concepts, ideas or meanings central to documents, in order to reason out the usefulness of documents to information needs, and secondly from its ability to learn from experience and be adaptive to the environment. In this work we attempt to incorporate these properties into the development of an IR model to improve document retrieval. We investigate the applicability of concept lattices, which are based on the theory of Formal Concept Analysis (FCA), to the representation of documents. This allows the use of more elegant representation units, as opposed to keywords, in order to better capture concepts/ideas expressed in natural language text. We also investigate the use of a reinforcement leaming strategy to learn and improve document representations, based on the information present in query statements and user relevance feedback. Features or concepts of each document/query, formulated using FCA, are weighted separately with respect to the documents they are in, and organised into separate concept lattices according to a subsumption relation. Furthen-nore, each concept lattice is encoded in a two-layer neural network structure known as a Bidirectional Associative Memory (BAM), for efficient manipulation of the concepts in the lattice representation. This avoids implementation drawbacks faced by other FCA-based approaches. Retrieval of a document for an information need is based on concept matching between concept lattice representations of a document and a query. The learning strategy works by making the similarity of relevant documents stronger and non-relevant documents weaker for each query, depending on the relevance judgements of the users on retrieved documents. Our approach is radically different to existing FCA-based approaches in the following respects: concept formulation; weight assignment to object-attribute pairs; the representation of each document in a separate concept lattice; and encoding concept lattices in BAM structures. Furthermore, in contrast to the traditional relevance feedback mechanism, our learning strategy makes use of relevance feedback information to enhance document representations, thus making the document representations dynamic and adaptive to the user interactions. The results obtained on the CISI, CACM and ASLIB Cranfield collections are presented and compared with published results. In particular, the performance of the system is shown to improve significantly as the system learns from experience.