Scholarlink

This paper presents a novel weakly-supervised text classification model that uses the class labels as the only source of supervision for unlabeled documents. The need for humongous amounts of training data for supervised text classification and need for domain expertise in case of semi-supervised text classification are overcome by the proposed approach. It constructs (1) class-space matrix for each class by concatenating contextual vectors of m pre-defined words that closely represent the class label, (2) generates psuedo labels for each document, and (3) generalizes the model using the concept of self-training. It does not use seed words as deployed by the current weakly-supervised approaches. Gensim’s Word2Vec model is used to obtain the top 100 similar words for each label name in advance, and then m words are manually selected from this list, based on their meaning, for each class. The proposed model’s performance is evaluated on AGNews, Twitter HateSpeech, and Amazon corpora without relying on the labeled data. The number of parameters of the proposed model is significantly lower than the state-of-the-art weakly-supervised text classification models, while the F1-score is comparable.

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

Weakly-Supervised Text Classification Using Class-Space Models and Language Model Self-training

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Ashish Reddy Chevva,
  • Somanadh Anirudh Ganapati,
  • Chetan Sai Potumarti,
  • Hima Bindu Kommanti

摘要

This paper presents a novel weakly-supervised text classification model that uses the class labels as the only source of supervision for unlabeled documents. The need for humongous amounts of training data for supervised text classification and need for domain expertise in case of semi-supervised text classification are overcome by the proposed approach. It constructs (1) class-space matrix for each class by concatenating contextual vectors of m pre-defined words that closely represent the class label, (2) generates psuedo labels for each document, and (3) generalizes the model using the concept of self-training. It does not use seed words as deployed by the current weakly-supervised approaches. Gensim’s Word2Vec model is used to obtain the top 100 similar words for each label name in advance, and then m words are manually selected from this list, based on their meaning, for each class. The proposed model’s performance is evaluated on AGNews, Twitter HateSpeech, and Amazon corpora without relying on the labeled data. The number of parameters of the proposed model is significantly lower than the state-of-the-art weakly-supervised text classification models, while the F1-score is comparable.