📄 Tweet Stance Detection Using an Attention based Neural Ensemble Model

Summary

The paper tackles stance detection on Twitter: deciding whether a tweet supports, opposes, or is neutral toward a given target, such as a political candidate. This is harder than sentiment analysis because the label depends on the target, and tweets are short, noisy, and full of hashtags and abbreviations. The authors propose PNEM, a neural ensemble model. The target and tweet word vectors are joined into one input matrix, and multi-kernel convolution extracts higher-level features from it. These features go to two parallel recurrent modules: a densely connected Bi-LSTM (DC-Bi-LSTM) and nested LSTMs (NLSTMs). Each module has a feed-forward attention layer that weights the important hidden states. The two outputs are concatenated and classified with softmax. PNEM was tested on the SemEval-2016 Task 6-A single-target data and on the multi-target 2016 US election dataset of Sobhani et al. (2017). On both, it beat a CNN+LSTM baseline and earlier deep learning systems. This suggests that combining newer LSTM variants with attention gives better tweet representations for target-dependent opinion mining.

Key points

• Single-target results (SemEval-2016 Task 6-A): The dataset has 2914 training and 1249 test tweets across 5 targets. PNEM reached F avg 72.11, compared with 67.15 for CNN+LSTM, 68.98 for SVM-ngrams, and 67.82 for MITRE (the best SemEval system). Against other deep models such as TGMN-CR, T-PAN, AS-biGRU-CNN and TAN, it gained between 1.07 and 3.44 points.
• Multi-target results: The dataset has 4455 tweets (3119 train, 446 dev, 890 test) over three candidate pairs. PNEM improved on Seq2Seq by 3.91 points and on DMAN by 1.99 points.
• Ablation study: Removing each component lowered F avg as follows:
  • multi-kernel convolution (MKC): −4.98
  • NLSTMs+ATT: −4.14
  • attention: −1.68
  • DC-Bi-LSTM+ATT: −1.11
  • target embedding: −0.99
• Setup: The model uses 300-d fastText skip-gram embeddings trained on Wikipedia, kernel sizes 2–5 with 600 filters, a 5-layer DC-Bi-LSTM and a 2-layer NLSTMs module. Training used Adam (learning rate 0.001), batch size 32, and up to 45 epochs.
• Preprocessing: NLTK stop words were removed, special characters were stripped, and hashtags were segmented following Baziotis et al. (2017).