Introduction
Word Sense Disambiguation (WSD) is a critical task in Natural Language Processing, focused on identifying the correct meaning of a word based on its context. While supervised machine learning models have achieved high accuracy in WSD, they are notoriously data-hungry, requiring extensive manually sense-tagged corpora. For many resource-poor languages, such as Marathi, Bengali, and Tamil, these resources are non-existent and prohibitively expensive to create. This research proposes a "parameter projection" method that allows these languages to leverage linguistic data from a resource-rich "pivot" language like Hindi.
The MultiDict Framework
The foundation of this cross-lingual approach is MultiDict, a synset-based multilingual dictionary. Unlike traditional dictionaries that link words directly, MultiDict first links "synsets" (groups of synonyms representing a single concept) across languages. It then establishes manual cross-links between specific words within those synsets. This structure allows the system to map a specific sense in Hindi directly to its equivalent in a target language like Marathi.
As shown in the figure above, the word saagar in Marathi can map to different Hindi equivalents depending on the intended sense (e.g., "water body" vs. "abundance"). By using Hindi as a pivot, target languages automatically inherit the complex semantic relations—such as hypernymy and meronymy—already established in the Hindi WordNet.
Parameter Projection Methodology
The core innovation of this work is the observation that within specific domains (e.g., Tourism or Health), word sense distributions remain largely invariant across languages. The researchers identified five key parameters for projection:
- Domain-specific sense distributions: The frequency of a sense within a field.
- Dominance of senses: Identifying primary branches in the WordNet hierarchy for a domain.
- Corpus co-occurrence: Probabilities of words appearing near one another.
- Semantic distance: The shortest path between synsets in the WordNet graph.
- Conceptual distance: A metric for nouns based on the depth of their common ancestor.
The validity of this projection is supported by empirical data. As seen in the table above, the probability of a sense given a word (P(S|word)) learned from a native Marathi corpus is remarkably similar to the probability projected from a sense-tagged Hindi corpus. For instance, the sense distribution for kimat (worth vs. price) shows only a marginal difference between native and projected values.
Disambiguation Algorithms
Two primary algorithms were employed to utilize these projected parameters:
- Iterative WSD (IWSD): A greedy algorithm that first disambiguates monosemous (single-sense) words and then uses them as context to resolve more complex polysemous words.
- Modified PageRank: A graph-based approach that evaluates all possible senses in a sentence simultaneously to find the most globally consistent interpretation.
Results and Discussion
The study evaluated these methods on Tourism and Health domain corpora across Hindi, Marathi, Bengali, and Tamil. The results demonstrate that exact statistics are less important than relative distributions; as long as the relative order of sense frequency is preserved during projection, performance remains high.
The projected models achieved F1-scores between 71% and 79%. While this represents an 8-11% drop compared to models trained on native data, it significantly outperforms the standard WordNet baseline (52-65%). This trade-off highlights the viability of parameter projection as a solution for WSD in resource-constrained environments, effectively bypassing the need for massive manual tagging efforts.
Summary updated