A Comparative Study of the Application of Different Learning Techniques to Natural Language Interfaces

A Comparative Study of the Application of Different Learning Techniques to Natural Language Interfaces Werner Winiwarter and Yahiko Kambayashi Dept. of Information Science Kyoto University Sakyo, Kyoto 606-01, Japan {ww I yahiko}©kuis, kyoto-u, ac. jp Abstract In this paper we present first results from a comparative study. Its aim is to test the feasibility of different inductive learn- ing techniques to perform the automatic acquisition of linguistic knowledge within a natural language database interface. In our interface architecture the machine learn- ing module replaces an elaborate semantic analysis component. The learning module learns the correct mapping of a user's input to the corresponding database command based on a collection of past input data. We use an existing interface to a produc- tion planning and control system as evalu- ation and compare the results achieved by different instance-based and model-based learning algorithms. 1 Introduction One of the main obstacles to the efficient use of nat- ural language interfaces is the often required high amount of manual knowledge engineering (see (An- droutsopoulos et al., 1995) for a recent survey). This time-consuming and tedious process is often referred to as "knowledge acquisition bottleneck". It may re- quire extensive efforts by experts highly experienced in linguistics as well as in the domain and the task (Riloff and Lehnert, 1994). Therefore, natural lan- guage interfaces represent a domain that is very well suited for the application of machine learning algo- rithms to automate the acquisition process of lin- guistic knowledge. So far, inductive learning has already been ap- plied successfully to a large variety of natural Jan- guage tasks. This includes basic linguistic problems such as morphological analysis (van den Bosch et al., 1996), parsing (Zelle and Mooney, 1996), word sense disambiguation (Mooney, 1996), and anaphora resolution (Aone and Bennett, 1996). Besides this, there also exists some research on applications, e.g. machine translation (Yamazaki et al., 1996), text categorization (Moulinier and Ganaseia, 1996), or information extraction (Soderland et al., 1996). The learning task in natural language interfaces is to select the correct command class based on seman- tic features extracted from the user input. There- fore, it can be modeled as classification problem, i.e. the machine learning algorithms construct a theory from the training data that is used for classifying unseen test data (Quinlan, 1996). So far, we con- sider only supervised learning so that each training case has to be labeled with the correct class. We apply different existing instance-based and model-based algorithms to this problem and com- pare the achieved results. In addition, we have also developed several new algorithms, which we present briefly in this paper. We have implemented all algorithms by means of the deductive object- oriented database system ROCK ~ ROLL (Barja et al., 1994). It solves the problem of updates in deductive databases in that it separates the declarative logic query language ROLL from the imperative data ma- nipulation language ROCK within the context of a common object-oriented data model. Besides this, ROCK ~ ROLL makes a clean distinction …

Content truncated to 500 words