Scholarlink

<p>Short, noisy social-media messages make sentiment analysis particularly challenging: slang, emojis, and misspellings not only reduce predictive accuracy but also obscure why a model outputs a given polarity. This paper introduces a lightweight hybrid model that combines a convolutional block, a bidirectional long short-term memory network (BiLSTM), and a single multi-head self-attention layer with a novel noise-invariant contrastive head (NICH) that normalizes noisy tokens before encoding. The model is trained in two phases (frozen then fine-tuned word embeddings) and evaluated on two contrasting benchmarks: Sentiment140 (1.6M tweets) and IMDb (50k movie reviews). Across both datasets, our approach outperforms classical baselines (logistic regression, support vector machine, random forest) and an Attention-BiLSTM reference model, while remaining competitive with a compact Transformer-based baseline (DistilBERT), achieving 90.63% accuracy and 0.9079 <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\hbox {F}_1\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mtext>F</mtext> <mn>1</mn> </msub> </math></EquationSource> </InlineEquation> on Sentiment140 and 91.90% accuracy and 0.9203 <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\hbox {F}_1\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mtext>F</mtext> <mn>1</mn> </msub> </math></EquationSource> </InlineEquation> on IMDb. Attention scores and a perturbation-based metric show that the model consistently focuses on sentiment-bearing tokens, providing token-level explanations that are useful for real-time monitoring of customer feedback and brand perception in noisy social-media streams.</p>

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

Attention-enhanced BiLSTM for causal sentiment mining in noisy social-media streams

  • 期刊论文

  • 发表日期：

    2026年1月31日

查看全文
  • Miloud Mihoubi,
  • Meriem Zerkouk,
  • Belkacem Chikhaoui

摘要

Short, noisy social-media messages make sentiment analysis particularly challenging: slang, emojis, and misspellings not only reduce predictive accuracy but also obscure why a model outputs a given polarity. This paper introduces a lightweight hybrid model that combines a convolutional block, a bidirectional long short-term memory network (BiLSTM), and a single multi-head self-attention layer with a novel noise-invariant contrastive head (NICH) that normalizes noisy tokens before encoding. The model is trained in two phases (frozen then fine-tuned word embeddings) and evaluated on two contrasting benchmarks: Sentiment140 (1.6M tweets) and IMDb (50k movie reviews). Across both datasets, our approach outperforms classical baselines (logistic regression, support vector machine, random forest) and an Attention-BiLSTM reference model, while remaining competitive with a compact Transformer-based baseline (DistilBERT), achieving 90.63% accuracy and 0.9079 \(\hbox {F}_1\) F 1 on Sentiment140 and 91.90% accuracy and 0.9203 \(\hbox {F}_1\) F 1 on IMDb. Attention scores and a perturbation-based metric show that the model consistently focuses on sentiment-bearing tokens, providing token-level explanations that are useful for real-time monitoring of customer feedback and brand perception in noisy social-media streams.