TL;DR
This paper proposes using decision trees to explicitly define domain-specific terms, moving beyond simple frequency-based methods. By modeling the hierarchical decision-making process of human terminologists, their approach achieves higher accuracy in identifying terms within specialized texts, offering a more robust way to build knowledge bases for information systems.
What Problem Does This Paper Address?
Knowledge-intensive information access systems, like those used for information extraction, often struggle with a "knowledge bottleneck." This means they are limited to specific applications because they lack a deep understanding of the specialized vocabulary (terms) within a given domain. Traditional methods for Terminology Acquisition (TA) often rely on defining what a term is and then searching for candidates in a corpus. However, these methods can be too simplistic, often using just the frequency of a word or phrase as a primary indicator of its "termhood." This paper argues that this approach misses crucial contextual and relational information that human experts use to identify terms. The challenge is to computationally capture this more nuanced, domain-specific understanding of what constitutes a term.
Methods
The core innovation is the use of decision trees to model the process of identifying terms. Instead of relying solely on frequency, the authors propose an "intensional term definition" (itd) represented by a decision tree. This tree learns to classify candidate terms by evaluating a set of features derived from the text.
Here's a breakdown of their approach:
- Implicit Domain Model (IDM): The system starts with a corpus of text and, optionally, a pre-existing list of known terms (a seeding terminology). This IDM contains the raw material for learning.
- Feature Engineering: They extract features that capture the "exogenous grammatical behavior" of terms. This includes:
- Syntactic Lexicalized Features: These focus on the grammatical relationships between a candidate term and other words in its context. For example, is the candidate a subject of a verb, or a prepositional modifier? Crucially, they identify the specific governing word (e.g., "equation" as the subject of "governs").
- Frequency-Based Features: Traditional measures of how often a term or phrase appears.
- Decision Tree Learning: A decision tree algorithm (specifically mentioning Quinlan's C4.5) is used to learn the classification rules. Each internal node in the tree represents a test on a feature (e.g., "Is the candidate the subject of the verb 'to be'?"), and leaf nodes represent the final decision (TERM or NON-TERM).
- Incremental Learning: To simulate how human terminologists refine their understanding, the learning process is incremental. New examples of terms and non-terms are added, and the decision tree is retrained, allowing the model to adapt and improve its "awareness" of the domain.
Datasets
The experiments were conducted on a corpus of approximately 250,000 words from Italian Civil Laws. This corpus was paired with a thesaurus of 600 term equivalence classes, built by expert terminologists. The thesaurus served as the ground truth for identifying valid terms. The authors processed the corpus using the CHAOS parser to generate candidate term equivalency classes, of which about 1/4 were found to be valid based on the thesaurus.
Key Insights and Ideas
- Decision Trees Mimic Human Expertise: The paper's central idea is that decision trees can effectively model the hierarchical, rule-based reasoning that human terminologists employ. This is a more sophisticated approach than simple statistical measures.
- Context is King (Especially Grammatical Context): Syntactic and lexicalized contextual information is far more discriminative for term identification than pure frequency. Understanding how a word or phrase is used grammatically within its domain is key.
- Negative Evidence is Crucial: Explicitly learning from "non-terms" (words or phrases that are not terms) significantly improves the model's ability to distinguish between actual terms and common words. This is analogous to how humans learn by exclusion.
- Incremental Learning Reflects Real-World Refinement: Simulating the incremental nature of human learning allows the model to evolve its understanding of termhood, mirroring how experts refine their judgment over time.
- Combining Information Sources is Superior: Integrating syntactic/lexicalized features with frequency-based measures yields the best results, highlighting the benefit of a multi-faceted approach.
Key Results
- Improved Accuracy: The proposed decision tree approach, particularly when combining syntactic lexicalized features with frequency (the $\Phi \times \Lambda$ space), achieved the lowest error rate of 13.88%. This is a significant improvement over methods relying solely on frequency ($\Phi$ space, 16.99% error) or purely syntactic features ($\Lambda$ space, 14.25% error).
- Domain Specificity Matters: The syntactic lexicalized approach ($\Lambda$) outperformed the pure frequency approach ($\Phi$), demonstrating that domain-specific grammatical patterns are highly effective.
- Evolution of Decision Rules: The study showed that as more negative examples were introduced (simulating increased terminologist scrutiny), the decision trees evolved. General rules based on common verbs like "to be" were replaced by more specific, domain-relevant cues, indicating a growing "awareness" of the domain.
Limitations and Future Work
- Computational Cost of Parsing: While the CHAOS parser is described as robust, the initial processing of the corpus to extract features can be computationally intensive.
- Ambiguity in Parsing: The authors acknowledge that syntactic ambiguity in parsing can affect feature extraction and potentially lead to irrelevant feature values. They address this by using a "plausibility" score for grammatical relations.
- Future Work: The authors suggest that this approach has potential for inducing ontological knowledge from texts, moving beyond just term identification to understanding deeper semantic relationships. They also see it playing a relevant role in future research on approximating terminologist behavior.
Summary updated