Abstract
This paper presents a two-tier model for identifying and normalizing multi-word term variations in specialized corpora. By decoupling structural transformations (syntagmatic level) from lexical relationships (paradigmatic level), the proposed FASTR parser achieves high precision in recognizing morphological, syntactic, and semantic variants. Evaluation on medical and agricultural datasets demonstrates the system's effectiveness in expanding search capabilities while highlighting the challenges of semantic drift in hybrid variations.
Introduction
In Information Retrieval (IR), a significant challenge is the linguistic diversity used to express a single concept. A standard search for "malignant tumor" might fail to retrieve documents containing variants like "malignancy in orbital tumors." These variations can be morphological (changing word forms), syntactic (altering sentence structure), or semantic (utilizing synonyms).
Existing systems often struggle to accurately conflate these variants. The primary difficulty lies in developing a model that is flexible enough to capture complex variations while remaining precise enough to avoid "semantic drift"—a phenomenon where the meaning of a variant shifts away from the original term's intent. This paper proposes a framework that addresses these issues by separating structural rules from lexical knowledge.
The Two-Tier Framework
The authors propose a framework that describes term variation through two distinct levels of representation:
- Syntagmatic Level: This level describes the "skeleton" or structure of the term using context-free rules and transformations. It defines how a specific structure, such as a Noun-Noun compound, can transform into a different structure, such as a Verb-Preposition-Noun phrase.
- Paradigmatic Level: This level describes the "flesh" or lexical items. It defines "families" of words related either by a common root (morphology) or by meaning (semantics).
To implement this model, the authors utilize FASTR, a unification-based transformational parser. Rather than seeking exact matches, FASTR uses the two-tier rules to recognize variants in raw text. For example, if the system identifies "measure" and "measurement" as morphologically related and "speed" and "velocity" as synonyms, it can automatically link "speed measurement" to "measuring velocity" using a pre-defined structural transformation rule.
Experimental Setup
The model was evaluated using two primary scientific corpora: AGRIC, a French corpus of agricultural scientific abstracts (1.2 million words), and MEDIC, an English corpus of medical scientific abstracts (1.3 million words). To support the paradigmatic level, several external knowledge sources were integrated, including MULTEXT and CELEX for morphology, and AGROVOC, WordNet 1.6, and the Microsoft Word97 thesaurus for semantics.
Results and Discussion
The system demonstrated high precision across various types of term variations. Syntactic variations reached a precision of 95.7% on the French corpus, while morpho-syntactic variations performed at 84.6%. Pure semantic variations also showed strong results, with precision ranging from 78% to 91% depending on the thesaurus used.
However, the study identified a "Semantic Shift" trap in hybrid variations—those combining structural changes with synonym swaps. Precision dropped as low as 29.4% in these cases, often due to polysemy. For instance, the French word former can mean "to train" or "to build," leading a system to incorrectly link "continuing education" (formation permanente) to "building a network" (former un réseau).
Despite these challenges, the impact on recall is substantial. The authors found that variants represent between 27% and 37% of term occurrences. Ignoring these variants means missing nearly a third of relevant mentions in a corpus, underscoring the importance of robust normalization techniques.
Conclusion
The decoupling of structure and lexicon allows for a highly generalizable model where a single transformation rule can apply to thousands of word pairs. While polysemy remains a hurdle for semantic precision, the two-tier framework significantly improves the recognition of term variants in specialized domains. Future work will focus on disambiguation techniques and expanding the model to handle longer, more complex multi-word expressions.
Summary updated