Combining Hand-crafted Rules and Unsupervised Learning in Constraint-based Morphological Disambiguation

Kemal Oflazer, Gökhan Tür

Kemal Oflazer, Gokhan Tur
Conference on Empirical Methods in Natural Language Processing (1996)

Introduction

Turkish is an agglutinative language where a single root can generate millions of potential word forms through the addition of suffixes. This "Lego-like" complexity presents significant challenges for computational linguistics, specifically in morphological disambiguation. Traditional Part-of-Speech (POS) taggers designed for languages like English often fail due to massive ambiguity and data sparsity. This paper presents a hybrid system that combines hand-crafted linguistic constraints with unsupervised transformation-based learning to achieve high-accuracy disambiguation.

Methodology and System Architecture

The authors developed a multi-stage pipeline to process raw Turkish text. The system begins with advanced preprocessing, identifying both lexicalized collocations (idioms) and non-lexicalized collocations (e.g., duplicated adverbs used as adverbs).

The multi-stage system architecture, showing the flow from raw text through preprocessing to the disambiguation and learning modules.

The core of the system relies on two types of hand-crafted constraints:

  1. "Choose" Rules: 288 rules that select a specific parse based on certain contextual evidence.
  2. "Delete" Rules: 43 rules that eliminate impossible parses.

These rules are designed to be conservative, firing only when linguistic evidence is nearly certain. To handle cases where rules do not apply, the system utilizes unsupervised learning based on Eric Brill’s transformation-based approach. It bootstraps from "islands of certainty"—unambiguous contexts created by the hand-crafted rules—to induce new rules from untagged text.

Hierarchical Feature Representation

A key innovation of this work is the treatment of words as hierarchical feature structures rather than flat tags. This approach allows the system to focus on relevant grammatical markers (like Case or Person) while abstracting away from specific roots when necessary.

Conversion of a flat feature sequence into a hierarchical morphological structure.

This hierarchical projection is crucial for the learning module, as it enables the system to generalize patterns across different words that share similar morphological properties. By ignoring irrelevant details and focusing on functional features, the system significantly reduces the data sparsity problem inherent in agglutinative languages.

Experimental Results

The system was evaluated on two primary corpora: ARK (archaeology text) and C2400/C270 (history text). The hybrid approach demonstrated significant improvements in both precision and ambiguity reduction:

  • Recall: 96% to 97%, indicating the correct parse was rarely discarded.
  • Precision: 93% to 94%.
  • Ambiguity Reduction: The average number of parses per token was reduced from approximately 1.8 to 1.03.
  • Unknown Word Handling: A secondary morphological processor for unknown words (assuming nominal roots) reduced the "unknown" rate to below 1%.

Discussion and Limitations

The authors concluded that unsupervised learning requires an initial set of hand-crafted rules to avoid getting "stuck" in local minima. However, some challenges remain:

  • Genuine Ambiguity: Words like o (which can be a pronoun or determiner) often require deep semantic or discourse knowledge beyond local syntax.
  • Long-distance Dependencies: Since the system uses a 5-word window, it can struggle with free-word-order dependencies that span larger distances.
  • Root Preference: Future work could benefit from incorporating statistical preferences for more common roots to resolve remaining ambiguities.

Summary updated

Abstract

Kemal Oflazer, Gokhan Tur. Conference on Empirical Methods in Natural Language Processing. 1996.
BibTeX
@inproceedings{oflazer1996combining,
  author = {Kemal Oflazer and Gökhan Tür},
  title = {{Combining Hand-crafted Rules and Unsupervised Learning in Constraint-based Morphological Disambiguation}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {1996},
  url = {https://openreview.net/forum?id=Cv_N2f4FtOf},
}
Figure