TL;DR
This paper introduces a novel corpus-based learning method for automatically extracting rules to index compound nouns in Korean documents. By leveraging statistical learning from a large tagged corpus, their approach overcomes the limitations of manual rule creation and improves both the coverage and efficiency of compound noun indexing for information retrieval.
What Problem Does This Paper Address?
Compound nouns are crucial for precise information retrieval because they convey more specific meaning than simple nouns. However, identifying and indexing them accurately is challenging, especially in languages like Korean where compound nouns can take many forms. Existing methods fall into two camps: statistical approaches that often struggle with diverse noun structures and require massive co-occurrence data, and linguistic approaches that rely on extensive human-crafted rules, which are labor-intensive, prone to errors, and limited in coverage. This paper aims to bridge this gap by developing a method that automatically learns indexing rules from data, requiring minimal linguistic expertise and offering broader coverage.
Methods
The proposed system has two main modules:
- Compound Noun Indexing-Rule Extraction:
- Seed Collection: They start with a set of "compound noun seeds" (known compound nouns) gathered from encyclopedias and dictionaries.
- Statistical Information Collection: A large tagged Korean corpus (1,000,000 "eojeol" or spacing units) is used to collect statistical information about these seeds.
- Rule Extraction: A template-based algorithm, using Part-of-Speech (POS) tags, automatically extracts potential indexing rules from the tagged corpus. This process is guided by the POS tags of adjacent words and the structure of the compound noun seeds.
- Rule Filtering: Extracted rules are refined using frequency thresholds and by discarding rules containing "negative lexical items" (words that indicate exclusion or non-existence, which can lead to spurious compound nouns).
- Precision Learning: The precision of each extracted rule is learned and refined iteratively. This involves calculating a "modified frequency" (MF) by combining a rule's precision with the frequency of the compound nouns it generates. This iterative process helps to improve the statistical information and thus the rule precision.
- Compound Noun Indexing, Filtering, and Weighting:
- Indexing: Documents are processed using a natural language processing engine that performs morphological analysis and POS tagging. The tagged sequences are then compared against the learned compound noun indexing rules to identify and index compound nouns.
- Filtering: To reduce over-generation of spurious compound nouns and improve efficiency, seven different filtering methods are evaluated. These methods utilize mutual information (MI), frequency of compound nouns (FC), and frequency of compound noun elements (FE).
- Final Filtering Method (H): A novel filtering method is proposed, combining the precision of rules with the mutual information of compound noun elements, weighted by a coefficient.
Datasets
- Compound Noun Seeds: Collected from ETRI Kemong encyclopaedia and Internet dictionaries (totaling 10,368 seeds).
- Tagged Corpus: A 1,000,000 "eojeol" (Korean spacing unit) tagged corpus (Korean Information Base) was used for extracting statistical information and rules.
- Test Collections:
- KTSET2.0 test collections (4,410 documents, 50 queries)
- KRIST2.0 test collection (13,514 documents, 30 queries)
Key Insights and Ideas
- Automated Rule Learning: The core idea is that statistical learning from a large corpus can automatically discover compound noun indexing rules, mimicking and even surpassing the coverage of human-crafted rules. This significantly reduces the manual effort and linguistic expertise required.
- Hybrid Approach: The method combines statistical learning with linguistic information (POS tags) in a synergistic way, leveraging the strengths of both.
- Iterative Precision Learning: The iterative refinement of rule precision, by incorporating modified frequencies, is a key mechanism for improving the accuracy and robustness of the learned rules, especially in the face of data sparseness.
- Effective Filtering: The proposed filtering method (H), which combines rule precision with mutual information, demonstrates superior effectiveness and efficiency compared to simpler filtering techniques. This highlights the importance of a multi-faceted approach to managing spurious compound nouns.
- Broader Coverage: The automatically extracted rules cover more diverse types of compound nouns than traditional manual rules, as evidenced by the fact that manual rules are a subset of their generated rules.
Key Results
- Coverage: The automatic method generated 30,168 actual compound nouns, a 35.4% increase over the 22,276 generated by manual linguistic rules.
- Effectiveness:
- On KTSET2.0, the proposed method achieved an average recall of 82.66% and 11-pt average precision of 42.24%, slightly lower than manual rules (83.62% recall, 42.33% precision) but with a significantly smaller number of index terms (515,801 vs. 504,040).
- On KRIST2.0, the proposed method achieved an average recall of 87.50% and 11-pt average precision of 34.54%, outperforming manual rules (86.32% recall, 34.33% precision) and again with a more manageable number of index terms (1,282,818 vs. 1,242,458).
- Filtering Efficiency: Method B (using average MI of compound noun seeds) achieved the best efficiency by generating the smallest number of compound nouns. However, the proposed method H achieved the best balance of effectiveness (recall and precision) and efficiency.
- Rule Diversity: The automatically extracted rules generated 2,017 rules after filtering, significantly more than the 5 manual linguistic rules, indicating greater diversity.
Limitations and Future Work
- Language Dependence: While the method is presented for Korean, the authors suggest that it doesn't inherently require language-dependent knowledge and could potentially be applied to other languages, which would require verification.
- Scalability: The authors mention the need to perform experiments on much larger commercial databases to test the practicality of their method.
- Future Work: The paper suggests future work could involve applying the method to other languages and testing it on larger, real-world datasets. They also note that their method doesn't require a parser, making it robust for unrestricted texts.
Summary updated