NLI_project

A conversational agent trained on Restaurant and Hotel domains (Erasmus Mundus EMAI project)

Team

**Name	GitHub
Camille
Bernardo
Roberta
Nazanin

Implementation of the core functionalities of a CA

• Domain identification/ Dialog act prediction

• Content extraction from User utterances (semantic frame slot filling)

• Agent move prediction

Data set: https://huggingface.co/datasets/multi_woz_v2i (i=2,3,4)

Approach:

• Dialog Act Prediction Model: A classification model that takes user utterance as input and predicts the dialog act. This could be an LSTM or Transformer-based model. (Classification)
  • Classifies the type of action the user is attempting to perform (e.g., asking a question, making a statement).
• Semantic Slot Filling Model: This could be a sequence-to-sequence model or a named entity recognition (NER) model. (Classification)
  • Extracts key pieces of information (e.g., date, location) from the user's utterance.
• Agent Move Prediction Model: Use reinforcement learning or rule-based methods to decide the agent's next move based on the dialog history and current state. (Reinforcement Learning)
  • Determines the next action the conversational agent should take (e.g., provide information, ask for clarification).