A Gradual Refinement Model for A Robust Thai Morphological Analyzer

Asanee Kawtrakul, Chalatip Thumkanon, Thitima Jamjanya, Parinee Muangyunnan, Kritsada PoolwanCOLING 1996

TL;DR

This paper introduces a novel hybrid approach for Thai morphological analysis, tackling word boundary ambiguity, POS tagging, and implicit spelling errors. Their "gradual refinement" model uses a three-stage pruning process (preference, syntactic, and semantic) to efficiently disambiguate and correct errors, achieving improved accuracy and speed compared to corpus-based methods.

What Problem Does This Paper Address?

Natural Language Processing (NLP) for languages like Thai, which lack explicit word delimiters (like spaces in English), presents a significant challenge. A crucial first step is morphological analysis, which involves segmenting text into words and assigning each word its correct part-of-speech (POS) tag. This paper addresses three core problems in Thai morphological analysis:

  1. Word Boundary Ambiguity: Thai text is a continuous stream of characters, making it difficult to determine where one word ends and another begins. This can lead to multiple valid segmentations, some of which might be meaningless.
  2. Tagging Ambiguity: Many Thai words can function as multiple parts of speech. This ambiguity, combined with word boundary issues, can create a vast number of potential syntactic structures for a single sentence.
  3. Implicit Spelling Errors: Unlike explicit spelling errors (e.g., typos easily caught by a dictionary), implicit errors result in unintended but correctly spelled words. These errors can significantly alter the meaning and syntactic structure of a sentence, making them hard to detect.

Existing corpus-based approaches, while often accurate, require extensive training data and can be computationally expensive. This paper aims to develop a more efficient and robust Thai morphological analyzer that can handle these challenges without relying heavily on large corpora.

Methods

The authors propose a "gradual refinement module" that employs a hybrid technique combining heuristic, syntactic, and semantic knowledge. This module operates in three stages:

  1. Preference-based Pruning (Step 1): This initial stage addresses word boundary ambiguity. It uses a "preferred words" list, often determined by word co-occurrence in the left or right context, to prune unlikely segmentations. If both shortest and longest segmentations are possible, this stage prioritizes based on preferences, deferring more complex disambiguation to later stages.
  2. Syntactic-based Pruning and Implicit Spelling Correction (Step 2): This stage tackles remaining word boundary ambiguities, POS tagging ambiguities, and implicit spelling errors. It utilizes "syntactic coarse rules" and a POS transition matrix (PM) to filter out improbable POS sequences. If no valid POS sequence is found, it flags potential implicit spelling errors. A word generation function then proposes candidate words, and the process may loop back to earlier stages for re-pruning.
  3. Semantic-based Pruning and Implicit Spelling Correction (Step 3): This final stage addresses implicit spelling errors that might have slipped through syntactic analysis. It leverages "semantic dependency strength" calculated from a Lexicon Matrix. Pairs of words with weak semantic connections are identified as potential implicit spelling errors. A word generation function is again used to propose alternatives, and the process may iterate back to Step 2 for re-evaluation.

The overall system starts with a "Word Segmenting & Explicit Spelling Error Pruning" step (Step 0), which uses lexicon bases and word information rules to generate initial possible word groupings and tags, also detecting explicit spelling errors.

Datasets

The paper mentions using a corpus of 20,000 sentences for obtaining the POS matrix (Table 3). For experimentation, they tested their system on a corpus of 200 sentences using a PC-486 DX2.

Key Insights and Ideas

  • Hybrid Approach is Key: The paper demonstrates that combining heuristic (preferred words), syntactic (coarse rules, POS matrix), and semantic (dependency strength) knowledge offers a powerful way to disambiguate complex linguistic phenomena like word boundaries and POS tags.
  • Gradual Refinement for Efficiency: Breaking down the disambiguation process into sequential stages (preference, syntactic, semantic) allows for efficient pruning of unlikely candidates early on, reducing the computational load for subsequent, more complex analyses.
  • Implicit Spelling Errors as a Morphological Problem: Recognizing implicit spelling errors as a problem solvable at the morphological analysis level, rather than solely a post-processing task, is a significant contribution.
  • Time-Efficiency Advantage: The proposed model shows promise in achieving high accuracy while being computationally less demanding than traditional corpus-based methods, which is crucial for practical NLP applications.

Key Results

The experimental results, based on a corpus of 200 sentences, show the proposed linguistic-based gradual refinement model outperforming a corpus-based approach (word filtering) in several metrics:

  • Word Segmentation Accuracy: 92.5% vs. 85.2%
  • POS Tagging Accuracy: 88.7% vs. 76.6%
  • Implicit Spelling Error Correction: 76.6% vs. 61.9%

Furthermore, the proposed model achieved significantly better time efficiency, processing each sentence in milliseconds compared to minutes for the corpus-based approach.

Limitations and Future Work

Limitations:

  • The paper acknowledges that while the corpus-based approach is efficient, it can be computationally costly and requires a large amount of training and validation data. The proposed model, while time-efficient, might still have room for improvement in accuracy.

Future Work:

  • Unknown Word Processing: The authors suggest extending the system to handle unknown words, which are words not present in the lexicon.
  • Increasing Accuracy: Further research is needed to enhance the accuracy of the gradual refinement method.

Summary updated

BibTeX
@inproceedings{kawtrakul1996gradual,
  author = {Asanee Kawtrakul and Chalatip Thumkanon and Thitima Jamjanya and Parinee Muangyunnan and Kritsada Poolwan},
  title = {{A Gradual Refinement Model for A Robust Thai Morphological Analyzer}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1996},
}
Figure