Scholarlink

In this research, we present BERT-LDA FusionNet, a novel hybrid model to improve text classification tasks by leveraging the strengths of Latent Dirichlet Allocation (LDA) for topic modelling and BERT (Bidirectional Encoder Representations from Transformers) for contextualized word embeddings, with a multi-modal approach to predict the topic of text as well as using the visual data. We apply the developed model to predict whether incident reports are ‘Related’ or ‘Unrelated’, which is a difficult task since such reports are unstructured and diverse. BERT-LDA FusionNet employs LDA to distil latent topics in the text and BERT to obtain semantic richness and understand the context of each word to better understand the textual data. Furthermore, due to the multi-modal integration, the model can join textual features of text with the image-based features that are extracted by Convolutional Neural Networks (CNNs) to improve the accuracy of classification. We also evaluate and comprehensively compare against other traditional and state-of-the-art models, such as Support Vector Machine (SVM), Random Forest (RF), Logistic Regression (LR), Naive Bayes (NB), Deep Neural Networks (DNN), Convolutional Neural Networks (CNN) and Long Short Term Memory (LSTM). Our results show that, although LDA FusionNet has higher accuracy in classification and topic coherence, for 92.4% accuracy it produces better results than these models. Fusion embeds BERT with LDA topics and preserves semantic structure, as well as thematic structure. Interpretability with the help of topic coherence and visualization of attention is made more. The model is geared towards big scale, does domain adaptation, and is applicable in real-time. By integrating multi-modal data with unsupervised topic modelling, we offer a novel practical and effective approach to solving complex text classifications of various types which arise in many real-world applications and confirm that the contextual word embeddings can be used effectively for any text classification problem.

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

BERT-LDA FusionNet: A Hybrid Approach for Text Classification in Incident Reports Combining BERT, LDA, and Multi-modal Data

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • M. Priyadharshini,
  • V. Murugesh,
  • Sanjib Kumar Raul,
  • G. Sangeetha,
  • T. P. Udhayasankar,
  • Sowjanya Ramisetty,
  • Sandeep Kumar Panda

摘要

In this research, we present BERT-LDA FusionNet, a novel hybrid model to improve text classification tasks by leveraging the strengths of Latent Dirichlet Allocation (LDA) for topic modelling and BERT (Bidirectional Encoder Representations from Transformers) for contextualized word embeddings, with a multi-modal approach to predict the topic of text as well as using the visual data. We apply the developed model to predict whether incident reports are ‘Related’ or ‘Unrelated’, which is a difficult task since such reports are unstructured and diverse. BERT-LDA FusionNet employs LDA to distil latent topics in the text and BERT to obtain semantic richness and understand the context of each word to better understand the textual data. Furthermore, due to the multi-modal integration, the model can join textual features of text with the image-based features that are extracted by Convolutional Neural Networks (CNNs) to improve the accuracy of classification. We also evaluate and comprehensively compare against other traditional and state-of-the-art models, such as Support Vector Machine (SVM), Random Forest (RF), Logistic Regression (LR), Naive Bayes (NB), Deep Neural Networks (DNN), Convolutional Neural Networks (CNN) and Long Short Term Memory (LSTM). Our results show that, although LDA FusionNet has higher accuracy in classification and topic coherence, for 92.4% accuracy it produces better results than these models. Fusion embeds BERT with LDA topics and preserves semantic structure, as well as thematic structure. Interpretability with the help of topic coherence and visualization of attention is made more. The model is geared towards big scale, does domain adaptation, and is applicable in real-time. By integrating multi-modal data with unsupervised topic modelling, we offer a novel practical and effective approach to solving complex text classifications of various types which arise in many real-world applications and confirm that the contextual word embeddings can be used effectively for any text classification problem.