Abstract
Saif Mohammad, Iryna Gurevych, Graeme Hirst, and Torsten Zesch present a method for estimating semantic distance in resource-poor languages by leveraging the conceptual structure of a resource-rich language's thesaurus. By creating cross-lingual distributional profiles of concepts (DPCs), they achieve state-of-the-art results on German benchmarks without requiring parallel corpora or sense-annotated data.
Introduction
Semantic distance—the degree of relatedness between words or concepts—is a cornerstone of natural language processing (NLP), powering applications from machine translation to spelling correction. However, computing this distance typically requires either high-quality structured knowledge sources (like WordNet) or massive text corpora. For the vast majority of the world's languages, such resources are scarce or non-existent. This paper addresses this "resource gap" by proposing a method to "borrow" the semantic structure of a resource-rich language like English to measure semantic distance in a target language like German.
Methodology: Cross-Lingual Mapping
The core innovation is the Cross-lingual Distributional Profile of Concepts (DPC). Traditional distributional measures track word-word co-occurrences, which often suffer from "sense conflation"—the mixing of different meanings of a polysemous word (e.g., "bank" as a river edge vs. a financial institution). The authors instead track word-concept co-occurrences.
The workflow involves three main steps:
- Bilingual Mapping: Words in the target language (German) are translated into the source language (English) using a bilingual lexicon.
- Concept Assignment: These English translations are mapped to categories in an English thesaurus (the Macquarie Thesaurus), which serves as a coarse-grained sense inventory.
- Matrix Construction: A Word-Category Co-occurrence Matrix (WCCM) is built from a raw target-language corpus. This matrix counts how often target words co-occur with other words that represent specific thesaurus concepts.
Addressing Ambiguity through Bootstrapping
A significant challenge in this process is the ambiguity introduced by both the bilingual dictionary and the thesaurus itself. A single German word might translate to multiple English words, each belonging to multiple thesaurus categories.
To mitigate this, the authors employ a bootstrapping technique. An initial pass over the corpus creates a "rough" WCCM. This matrix is then used to disambiguate word occurrences in a second pass, creating a refined, more accurate distributional profile. This allows the system to "self-correct" the noise introduced by the bilingual dictionary.
Evaluation and Results
The method was evaluated using a 240-million-word German newspaper corpus (taz), the Macquarie Thesaurus, and the BEOLINGUS German-English lexicon. Performance was measured against human-annotated word-pair datasets (Gur65 and Gur350) and a large-scale vocabulary task from Reader's Digest (RDWP).
The results were striking: the cross-lingual DPC approach outperformed monolingual methods that relied on GermaNet (the German version of WordNet). On the Gur65 dataset, the cross-lingual ASD measure achieved a correlation of 0.794 with human judgments, surpassing the best monolingual score of 0.764. In the RDWP task, the cross-lingual approach demonstrated significantly higher coverage and superior F-scores (0.32 vs. 0.29) compared to native German resources.
Conclusion
This research demonstrates that the semantic structure of a resource-rich language can be effectively "exported" to resource-poor languages. By using thesauri as coarse-grained sense inventories and leveraging bilingual lexicons, researchers can compute high-quality semantic distance measures without the need for expensive, manually-curated native resources or parallel corpora. This opens new avenues for NLP in low-resource linguistic environments.
Summary updated