HJ-Ky-0.1: an Evaluation Dataset for Kyrgyz Word Embeddings

TL;DR

HJ-Ky-0.1 introduces the first silver-standard dataset for Kyrgyz word embedding evaluation by translating the Russian HJ dataset and validating it with non-contextual embeddings. The paper compares pre-trained and Leipzig-corpus–trained embeddings (word2vec SGNS, fastText, and compressed fastText), showing faster subword models and cleaner Leipzig data yield higher semantic-correspondence scores, while stemming degrades performance. The work demonstrates Kyrgyz embeddings can be effectively evaluated with non-contextual methods and highlights the importance of data quality and language-specific resources. Future work aims to refine annotations and expand the dataset to strengthen guidance for Kyrgyz vector semantic methods.

Abstract

One of the key tasks in modern applied computational linguistics is constructing word vector representations (word embeddings), which are widely used to address natural language processing tasks such as sentiment analysis, information extraction, and more. To choose an appropriate method for generating these word embeddings, quality assessment techniques are often necessary. A standard approach involves calculating distances between vectors for words with expert-assessed 'similarity'. This work introduces the first 'silver standard' dataset for such tasks in the Kyrgyz language, alongside training corresponding models and validating the dataset's suitability through quality evaluation metrics.