Scholarlink

In this research, an Assamese parts-of-speech (POS) tagger was created using advanced neural network architectures like Long Short-Term Memory (LSTM) with Conditional Random Fields (CRF) and Convolutional Neural Networks (CNN) with CRF. Our goal was to improve the accuracy of POS tagging for Assamese using the Bureau of Indian Standards (BIS) tag set. A large corpus of 70,000 manually tagged Assamese words was created and annotated according to BIS standards. The text underwent extensive preprocessing to ensure cleanliness and effective tokenization, and the words and tags were converted into numerical formats suitable for training our models. Our architecture included an embedding layer with pre-trained word embeddings to grasp linguistic subtleties, LSTM and CNN layers to recognize patterns in the text sequence, and a CRF layer to refine POS tag predictions. Our investigations focused on various configurations, and the LSTM-CRF model achieved an accuracy of 89.32%, while the CNN-CRF model achieved an accuracy of 95.52%. These findings represent a significant advancement in POS tagging for Assamese, providing valuable insights for linguistic research in related languages.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hybrid Approaches for Assamese POS Tagging: LSTM-CRF and CNN-CRF Model

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Nomi Baruah,
  • Pritom Jyoti Goutom

摘要

In this research, an Assamese parts-of-speech (POS) tagger was created using advanced neural network architectures like Long Short-Term Memory (LSTM) with Conditional Random Fields (CRF) and Convolutional Neural Networks (CNN) with CRF. Our goal was to improve the accuracy of POS tagging for Assamese using the Bureau of Indian Standards (BIS) tag set. A large corpus of 70,000 manually tagged Assamese words was created and annotated according to BIS standards. The text underwent extensive preprocessing to ensure cleanliness and effective tokenization, and the words and tags were converted into numerical formats suitable for training our models. Our architecture included an embedding layer with pre-trained word embeddings to grasp linguistic subtleties, LSTM and CNN layers to recognize patterns in the text sequence, and a CRF layer to refine POS tag predictions. Our investigations focused on various configurations, and the LSTM-CRF model achieved an accuracy of 89.32%, while the CNN-CRF model achieved an accuracy of 95.52%. These findings represent a significant advancement in POS tagging for Assamese, providing valuable insights for linguistic research in related languages.