TL;DR
This paper tackles the challenge of automatically grouping related terms within a specific domain (Polish economic texts) to aid in building domain models for information extraction. They propose a novel approach that combines WordNet's semantic relationships with morphosyntactic features derived from text, demonstrating that this hybrid method significantly improves clustering accuracy compared to using WordNet alone.
What Problem Does This Paper Address?
Building domain-specific knowledge bases, like ontologies or terminology lists, is crucial for many Natural Language Processing (NLP) tasks, including information extraction and question answering. However, creating these resources is often a manual, time-consuming, and error-prone process. While automatic term extraction methods exist, organizing these extracted terms into semantically coherent groups (clustering) remains a significant challenge. Existing approaches often rely heavily on general language resources, which may not adequately capture domain-specific nuances, or they neglect valuable linguistic information. This paper addresses the need for an automated method to cluster domain-specific terms, specifically focusing on Polish economic texts, by leveraging both semantic (WordNet) and morphosyntactic information.
Methods
The authors propose a multi-faceted approach to terminology clustering:
- Automatic Term Extraction: They begin by extracting candidate terms from Polish Wikipedia economic articles. This process is supported by linguistic knowledge, including morphological analysis and a custom domain dictionary.
- Term Similarity Definition: The core of their contribution lies in defining various types of term similarity. These include:
- Contextual Similarity: Analyzing the surrounding words (left/right contexts), nearest verbs, nouns, and prepositions.
- Coordination: Identifying terms that appear together in coordinated structures (e.g., connected by "and" or commas).
- Syntactic Patterns: Recognizing terms that appear in specific syntactic constructions indicating similarity (e.g., "such as," "for example").
- Lexical Similarity: Measuring similarity based on common head elements and common modifiers within terms.
- WordNet Similarity: Utilizing WordNet's hierarchical structure (synsets, hypernyms, hyponyms) to assess semantic relatedness.
- Overall Similarity Calculation: An overall similarity score for a pair of terms is calculated as a weighted sum of up to 19 different similarity coefficients derived from the above methods.
- Hierarchical Clustering: They employ hierarchical clustering using the MultiDendrograms tool, specifically using the unweighted average of similarity coefficients.
What's novel is the explicit combination of morphosyntactic features (context, coordination, syntax) with WordNet similarity, and the systematic evaluation of different weighting schemes for these features.
Datasets
- Polish Wikipedia Economic Articles: The primary dataset for term extraction and clustering consists of economic articles from Polish Wikipedia, collected in 2011. This dataset comprises 1219 articles and approximately 456,000 tokens.
- National Corpus of Polish (NKJP): A balanced one-million-word subcorpus of general Polish (NKJP) was used for comparison to filter out terms that are too general and not specific to the economic domain.
- Manual Clustering: A manually prepared version of the term partition was created by a human annotator (and verified by a second annotator) to serve as a gold standard for evaluating the automatic clustering results. This manual grouping resulted in 127 groups.
Key Insights and Ideas
- Morphosyntactic Information is Powerful: The study strongly suggests that morphosyntactic information, particularly term contexts and syntactic patterns, can be highly effective for initial term clustering, sometimes even outperforming WordNet alone. This is a crucial insight for low-resource languages or highly specialized domains where comprehensive semantic resources might be scarce.
- Hybrid Approach Yields Best Results: Combining WordNet's semantic information with morphosyntactic features leads to the most accurate clustering. This highlights the complementary nature of different linguistic information sources.
- Context Matters: The detailed analysis of contextual similarity (POS tags, base forms of adjacent words, nearest verbs/nouns) proves to be a valuable signal for term relatedness.
- Domain Specificity is Key: The comparison with the general Polish corpus (NKJP) underscores the importance of filtering domain-specific terms to avoid noise and improve the relevance of the clustering.
Key Results
- The proposed method, combining WordNet and morphosyntactic information, achieved a best F-measure of 75.1% when comparing the automatically generated clusters to the manually created groups (model W5).
- Using only WordNet similarity (model W1) resulted in a best F-measure of 68.8%.
- Using morphosyntactic information alone (model W2, where WordNet similarity was set to 0) yielded usable results at the level of about 62%.
- The experiments showed that adjusting the weights of different similarity coefficients could enhance the matching of clusters, with reasonable weight distributions yielding results about 5% better than using only WordNet data.
- The authors identified 400 terms for the clustering experiment, and the manual grouping resulted in 127 groups, with a maximum group size of 14.
Limitations and Future Work
- Quality of Lexical Tools: The authors acknowledge that the quality and coverage of Polish taggers and other lexical tools used for preprocessing are not always optimal, especially for specialized texts. This can impact the quality of the results.
- Manual Inspection for General Terms: Phrases with a greater C-value in the general corpus (NKJP) than in the economic corpus require manual inspection, indicating that perfect automatic filtering of general terms is challenging.
- Future Work: The paper suggests that the presented method can be adapted to texts in any domain or language, provided that appropriate lexical tools are available. Further research could explore more sophisticated methods for integrating semantic and morphosyntactic information, as well as investigate the impact of different clustering algorithms. The authors also note that the terminology selection stage can be less precise when economic texts share a large common part with general language.
Summary updated