Introduction
In Natural Language Processing (NLP), identifying that different words represent the same underlying concept is a fundamental challenge. For instance, a query for "development" should ideally retrieve documents containing "develop." Traditionally, researchers have relied on two primary methods: reductionist stemming (e.g., the Porter stemmer) and expansionist generation. However, stemming often suffers from over-stemming (linking unrelated words) or under-stemming (failing to link related words like adhere and adhesion). Conversely, expansionist systems frequently generate non-existent words.
Furthermore, existing resources like WordNet, while effective at linking synonyms, often fail to connect words across different parts of speech (e.g., the verb feed and the noun food). This paper introduces CatVar, a large-scale database designed to bridge this "cross-categorial" gap by mapping words that share a root but differ in part-of-speech.
Methodology
CatVar (version 2.0) was developed using a hybrid approach that combines human-curated linguistic data with automated clustering algorithms. The process began with Seed Extraction, pulling explicit links from high-quality resources such as NOMLEX (for nominalizations) and LDOCE (for verb-noun pairs).
These seeds were then used to guide an automated Clustering process. While the Porter stemmer was used to group words from the Brown Corpus and Penn Treebank, the "seed" links ensured that these clusters remained linguistically valid. For example, a cluster might include {hunger (N), hunger (V), hungry (Adj), hungriness (N)}.
Key Results and Evaluation
CatVar 2.0 comprises 96,368 unique lexemes organized into 62,232 clusters. Human evaluation by native speakers demonstrated high reliability, with a precision of 91.82% and an F-score of approximately 89%.
A significant finding is the "linkability" of the database. The authors found that 30% of the links within CatVar are invisible to standard stemmers. In practical applications, CatVar has shown tangible benefits:
- Machine Translation: Since 98% of translation "divergences" involve categorial variation, CatVar serves as a critical alignment tool.
- Headline Generation: Using CatVar as an "event filter" to ensure headlines contain informative verbal or nominalized roots improved Bleu scores from 0.1687 to 0.1740.
Discussion and Future Work
Despite its scale, CatVar faces certain limitations. As shown in the distribution of cluster sizes, approximately half of the database currently consists of single-word clusters. Future efforts will focus on "absorbing" these into larger, related groups. Additionally, while CatVar links words by shared roots, it does not yet distinguish between specific semantic relationships (e.g., distinguishing an action like construction from an agent like constructor).
Future work also includes integrating inflectional analysis to handle plurals and tenses, as well as expanding the database to support cross-linguistic tasks and bilingual alignment.
Summary updated