Multi-Task Learning Model for Kazakh Query Understanding

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Low-Resource Language Technologies

Background:

• Query understanding (QU) is crucial for question answering and dialogue systems, involving named entity recognition (NER) and question classification (QC).
• Traditional pipeline methods treat NER and QC separately, potentially missing task correlations.
• Kazakh, an agglutinative language, presents unique challenges for existing QU techniques due to its rich morphology.

Purpose of the Study:

• To propose a novel multi-task learning model for query understanding (MTQU) tailored for agglutinative, low-resource languages like Kazakh.
• To leverage the inherent correlation between NER and QC tasks for mutual performance enhancement.
• To develop and introduce new corpora for Kazakh query understanding (KQU) to facilitate research.

Main Methods:

• Developed a multi-task learning model (MTQU) that jointly trains NER and QC tasks.
• Incorporated a multi-feature input layer to effectively utilize Kazakh's linguistic characteristics (stem, suffixes, POS, gazetteers).
• Constructed the Kazakh Query Understanding (KQU) corpora for training and evaluation.

Main Results:

• The MTQU model demonstrated a simple yet effective approach to query understanding.
• The multi-feature input layer significantly improved model performance.
• The proposed model achieved competitive results on the newly created KQU dataset.

Conclusions:

• Jointly training NER and QC tasks within an MTQU framework is beneficial, especially for morphologically rich, low-resource languages.
• The MTQU model effectively captures and utilizes language-specific features, outperforming traditional methods.
• The introduction of the KQU corpora provides a valuable resource for advancing Kazakh NLP research.