Goal: To explore context modelling in dialogue systems and improve DA tagging accuracy through multi-modal sequential input design.
In this project, I developed a hybrid model combining Convolutional Neural Networks (CNN) and Bidirectional LSTMs (BiLSTM) to perform dialogue act (DA) tagging. The model utilises contextual information from previous utterances and their DA labels to make more accurate predictions.
A key focus was placed on understanding how incorporating previous dialogue acts improves current tagging performance. Attention masks were also analysed to interpret which parts of the input and past context the model focuses on during inference.
Model Design:
- CNNs used for local feature extraction from utterance embeddings
- BiLSTM layers incorporated for sequential context tracking
- Previous DA tags encoded and passed alongside utterance features
- Soft attention mechanism implemented to weight contextual relevance
Results:
- Demonstrated improvement in accuracy with contextual integration
- Visualised attention distributions over past utterances
- Performed qualitative analysis of model decisions
Tech stack: PyTorch, NLTK, matplotlib