Exploring Deep Neural Networks and Transfer Learning for Analyzing Emotions in Tweets

An LSTM-CNN that slightly beats the SemEval-2018 benchmark on multi-label emotion classification (Jaccard 58.63 vs. 57.88). For intensity regression, features transferred from several pre-trained models into XGBoost reach 0.798 average Pearson, essentially tied with the top system (0.799) and best on fear and sadness. Deep SHAP heat-maps show which words and emoji drive each prediction.

Problem

Emotion analysis usually stops at classification, but emotions vary in intensity. Labeled intensity data is scarce, and deep emotion models are rarely explained; attention weights give only a partial view.

Approach

• Data: SemEval-2018 Affect in Tweets. E-c has 11 emotions, multi-label, 10,983 tweets. EI-reg has per-emotion intensity for anger, fear, joy and sadness, about 3k tweets each (Table 1).
• ECCU (classification; Fig. 1): Twitter word2vec (frozen) → LSTM → CNN → max pooling → sigmoid. EIPU (intensity) uses the same design with a single output (Table 2).
• EITL (transfer learning for intensity; Fig. 2): concatenates features from ECCU, EIPU, DeepMoji and the OpenAI sentiment neuron, then trains an XGBoost regressor per emotion (Table 3).
• Explanations: Deep SHAP values normalized per tweet and rendered as word-level heat-maps (Eq. 2, Table 6).
• Preprocessing with ekphrasis (spell correction, segmentation).

Fig. 1: Embeddings → LSTM → CNN with max pooling → sigmoid emotion vector.Figure from Senarath et al., arXiv 2020.

Fig. 2: ECCU, EIPU, DeepMoji and sentiment-neuron features → XGBoost → intensity.Figure from Senarath et al., arXiv 2020.

Key results

• Classification (Table 5): ECCU scores Jaccard 58.63, micro-F1 71.92, macro-F1 52.8, vs. NTUA-SLP 57.88 / 70.1 / 52.8, the benchmark the paper uses.
• Intensity (Table 4), Pearson r:
  • EITL averages 79.81% vs. SeerNet 79.90% (state of the art) and NTUA-SLP 77.70%.
  • Per emotion: fear 78.67 (vs. 77.90) and sadness 79.99 (vs. 79.80) beat SeerNet; anger is 82.16 vs. 82.70 and joy 78.42 vs. 79.20.
  • EIPU alone reaches 70.83 avg, well above the BoW/NBoW baselines (52–64).
• Explanations (Table 6): keywords and emoji dominate the predictions, and the model takes context into account.

Contributions

1. Simpler but effective models for emotion classification and intensity prediction.
2. Applying state-of-the-art interpretation (Deep SHAP) to visualize and explain deep emotion-intensity models.

Future work

• Attention mechanisms (Transformers) to improve intensity prediction.

@article{Senarath2020,
  author   = {Senarath, Yasas and Thayasivam, Uthayasanker},
  journal  = {Preprint},
  title    = {Exploring Deep Neural Networks and Transfer Learning for Analyzing Emotions in Tweets},
  year     = {2020},
  keywords = {arXiv,first}
}