Knowledge Extraction and Recurrent Neural Networks: An Analysis of an Elman Network trained on a Natural Language Learning Task

II I | / I! I! | I II II | II II Knowledge Extraction and Recurrent Neural Networks: An Analysis of an Elman Network trained on a Natural Language Learning Task Ingo Schellhammer*#, Joachim Diederich*, Michael Towsey*, Claudia Brugman** * Neurocomputing Research Centre, Queensland University of Technology, QLD, 4001, Australia # Dept of Information Systems, University of Muenster, D-48149 Muenster, Germany **School of Languages, University of Otago, NewZealand j oachim@fit, qut. edu. au Abstract We present results of experiments with Elman recurrent neural networks (Elman, 1990) trained on a natural language processing task. The task was to learn sequences of word categories in a text derived from a primary school reader. The grammar induced by the network was made explicit by cluster analysis which revealed both the representations formed during learning and enabled the construction of state-transition diagrams representing the grammar. A network initialised with weights based on a prior knowledge of the text's statistics, learned slightly faster than the original network. In this paper we focus on the extraction of grammatical rules from trained Artificial Neural Networks and, in particular, Elman-type recurrent networks (Elman, 1990). Unlike Giles & Omlin (1993 a,b) who used an ANN to simulate a deterministic Finite State Automaton (FSA) representing a regular grammar, we have extracted FSA's from a network trained on a natural language corpus. The output of k-means cluster analysis is converted to state-transition diagrams which represent the grammar learned by the network. We analyse the prediction and generalisation performance of the grammar. 1. Introduction Since their renaissance in the mid-1980s, Artificial Neural Network (ANN) techniques have been successfully applied across a broad spectrum of problem domains such as pattern recognition and function approximation. However despite these capabilities, to an end user an ANN is an arcane web of interconnected input, hidden, and output units. Moreover an ANN solution manifests itself entirely as sets of numbers in the form of activation function parameters and weight vectors. As such a trained ANN offers little or no insight into the process by which it has arrived at a given result nor, in general, the totality of "knowledge" actually embedded therein. This lack of a capacity to provide a "human comprehensible" explanation is seen as a clear impediment to a more widespread acceptance of ANNs. In order to redress this situation, recently considerable effort has been directed towards providing ANNs with the requisite explanation capability. In particular a number of mechanisms, procedures, and techniques have been proposed and developed to extract the knowledge embedded in a trained ANN as a set of symbolic rules which in effect mimic the behaviour of the ANN. A recent survey conducted by Andrews et al. (1995) offered an insight into the modus operandi of a broad cross-section of such techniques. 2. Methods 2.1. The data The data for these experiments were obtained from a first-year primary school reader published circa 1950's (Hume). To keep this initial task simple, sentences with embedded structures (relative clauses) and a length of more …

Content truncated to 500 words