Automatic Quality Control of Uzbek WordNet Synsets Using Rule-Based and Graph-Based Diagnostics
Abstract
This article examines quality control issues in the Uzbek lexical-semantic dataset UzWordNet-GoldCore, which consists of 11,178 source entries. We developed a two-stage verification system. In the first stage, deterministic rules are applied to the source table; in the second stage, candidate synsets and semantic relations are represented as a graph and their structure is checked. As a result of the conversion, 7,072 lexical items, 10,533 senses, 2,740 candidate synsets, 11,095 unique examples, and 9,086 relation entries were generated. Through exact matching, 2,500 relation targets, or 27.51 percent, were successfully linked. The remaining cases include lexical targets that do not exist, ambiguous matches, category names, part-of-speech mismatches, spelling variants, and relation fields composed of multiple units. We also examined how different grouping criteria affect the number of synsets. Gloss-based grouping alone yielded 2,706 groups, while resolving groups with metadata conflicts produced a minimum of 2,953–3,078 groups. Six cases were manually reviewed: five were assessed as potential metadata errors and one as an acceptable modeling exception. The proposed system does not automatically modify the resource but instead generates a list of cases to be monitored for specialist review.