[Paper Review] Inducing Constraint Grammars

Christer Samuelsson, Pasi Tapanainen+1 more|ArXiv.org|Jul 1, 1996

Natural Language Processing Techniques4 citations

TL;DR

This paper presents a method for automatically inducing Constraint Grammar rules from annotated corpora using local lexical biases and next-neighboring context, enhanced with barrier mechanisms to improve structural accuracy. The approach achieves 98.2% recall and 87.3% precision (1.12 tags per word), demonstrating strong performance compared to other induction methods, though still below hand-crafted rule systems.

ABSTRACT

Constraint Grammar rules are induced from corpora. A simple scheme based on local information, i.e., on lexical biases and next-neighbour contexts, extended through the use of barriers, reached 87.3 percent precision (1.12 tags/word) at 98.2 percent recall. The results compare favourably with other methods that are used for similar tasks although they are by no means as good as the results achieved using the original hand-written rules developed over several years time.

Motivation & Objective

• To develop an automated method for generating Constraint Grammar rules from annotated corpora without manual hand-crafting.
• To improve the accuracy of part-of-speech tagging by leveraging local lexical information and contextual constraints.
• To explore the use of barriers in constraint grammar induction to model syntactic boundaries and improve parsing precision.
• To evaluate the performance of induced rules against existing methods in terms of recall and precision.
• To assess the feasibility of replacing or accelerating the development of hand-written constraint grammars through data-driven induction.

Proposed method

• The method uses local lexical biases—specifically, the part-of-speech tags of neighboring words—to guide rule induction.
• It applies a context window around each word, analyzing the immediate left and right lexical neighbors to infer likely tag assignments.
• Barriers are introduced to prevent rule application across syntactic boundaries, such as clause or phrase breaks, improving structural accuracy.
• Rules are induced iteratively by analyzing co-occurrence patterns in the training corpus, favoring high-frequency, high-precision patterns.
• The system uses a greedy rule selection strategy, prioritizing rules that maximize precision while maintaining high recall.
• The final grammar is constructed by combining the most reliable rules derived from the corpus, with pruning to avoid overfitting.

Experimental results

Research questions

• RQ1 Can constraint grammar rules be effectively induced from annotated corpora using only local lexical and contextual information?
• RQ2 How does the inclusion of barriers affect the precision and recall of induced constraint grammars?
• RQ3 What level of performance can be achieved by an automatically induced grammar compared to manually crafted systems?
• RQ4 To what extent does local context alone suffice for reliable part-of-speech tagging in a rule induction framework?
• RQ5 Can the induced grammar generalize well across unseen data while maintaining high recall?

Key findings

• The induced constraint grammar achieved 98.2% recall, indicating strong coverage of the test data.
• The system reached 87.3% precision, corresponding to an average of 1.12 tags per word, which is competitive with other induction-based methods.
• The use of barriers significantly improved precision by preventing incorrect rule applications across syntactic boundaries.
• Despite strong performance, the results fell short of those achieved by hand-written constraint grammars developed over several years of manual tuning.
• The method demonstrated that local context and lexical biases alone can yield a robust, data-driven approach to grammar induction.
• The results suggest that rule induction from corpora is a viable alternative to manual rule creation, especially in low-resource settings.