A comparative analysis on question classification task based on deep learning approaches.

Area of Science:

• Natural Language Processing
• Machine Learning
• Computational Linguistics

Background:

• Deep learning approaches have shown success in various text-mining tasks.
• Question classification is crucial for automatic question answering systems.
• Turkish language presents unique challenges due to its highly inflected nature.

Purpose of the Study:

• To investigate deep learning approaches for question classification in Turkish.
• To evaluate the effectiveness of different deep learning architectures and word embedding techniques.
• To achieve high accuracy in classifying Turkish questions for improved NLP applications.

Main Methods:

• Trained and tested Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), and Convolutional Neural Networks (CNN) architectures.
• Implemented hybrid models: CNN-GRU and CNN-LSTM.
• Utilized Word2vec (skip-gram and CBOW) for word embedding with various vector sizes on a Turkish question corpus.
• Performed comparative analysis using test and 10-cross fold validation accuracy.

Main Results:

• Deep learning architectures combined with Word2vec demonstrated significant impact on classification accuracy.
• The study achieved a maximum accuracy of 93.7% for Turkish question classification.
• Different Word2vec techniques and vector sizes influenced the performance of the deep learning models.

Conclusions:

• Deep learning, particularly with effective word embeddings like Word2vec, is highly effective for question classification in morphologically rich languages like Turkish.
• The findings highlight the importance of choosing appropriate word embedding strategies to optimize NLP tasks.
• The developed models offer a robust solution for enhancing automatic question answering systems in Turkish.