How Can a Dual-Channel Model Grasp Emotions in Video Comments?

Have you ever noticed the comments that fly across videos on platforms like Bilibili? Those messages, called "danmu," scroll in real time and are packed with slang, emojis, and quick jokes. Analyzing them is a challenge for AI because the words aren't used like in textbooks. A new study built a dual‑channel model, named ADCMix‑BiLSTM, to better understand these emotions.

The main trick is to switch from words to characters, so the model captures abbreviations and wordplay typical of danmu. Then, the model splits into two channels. The first combines self‑attention with a dilated convolutional neural network (ADCMix) and, on top of that, uses a BiLSTM with an attention mechanism to grasp context. The second channel balances possible biases, preventing the model from focusing on misleading signals.

Finally, a multi‑dimensional decoder translates the information into various sentiment categories, even distinguishing subtle nuances. Tests on two datasets showed the new approach outperforms previous methods by at least 2 % in accuracy and 3 % in F1‑score. In practice, the model better separates similar feelings, making real‑time comment analysis more reliable.

The ADCMix‑BiLSTM model is built on a hybrid architecture that merges the capabilities of self‑attention with those of dilated convolutional neural networks. The ADCMix (Attention‑Dilated CNN Mix) employs convolutional filters with increasing dilation, allowing it to capture long‑range dependencies without excessively increasing the parameter count. This is especially useful for danmu, where colloquial expressions can contain patterns of varying length. Self‑attention, on the other hand, enables the model to dynamically weight relationships among all tokens in the sequence, highlighting key words that most influence sentiment.

The BiLSTM channel with attention adds a bidirectional sequential perspective, processing context both left‑to‑right and right‑to‑left. Attention applied to the BiLSTM assigns higher weights to hidden representations most relevant for the final classification. The second channel, designed as a complementary encoder, uses a lighter structure based on character embeddings and a multi‑layer feed‑forward network. This channel acts as a "quality control," mitigating biases introduced by the first channel and providing an alternative data representation.

The multi‑dimensional decoder is implemented via a series of dense heads, each responsible for a sentiment dimension (e.g., joy, sadness, surprise, irritation). Using softmax activation for each head yields independent probabilities for each dimension, reflecting the multi‑label nature of danmu. Training employs a combined loss that sums the cross‑entropy of all heads, promoting simultaneous learning of multiple emotional aspects. Experimental results demonstrate that integrating these components yields a significant improvement over models based solely on word embeddings or monodimensional architectures, confirming the effectiveness of a dual‑channel approach for sentiment analysis of highly colloquial texts.