The primary task of any goal-oriented dialogue agent is to accomplish a user's task with utmost user satisfaction. Existing virtual agents assume that the end-user will always have a predefined task goal, which will be served after filling the corresponding intent and slots. However, in real-life, users do not always have a generic pre-known task goal, i.e., they determine their precise task goal dynamically based on their utility value and the agent's serving capability. They may upgrade/downgrade/update their goal components in real-time to maximize their utility value. This assumption emphasizes the gap between real human assistance and virtual agent assistance, where users have a flexible goal to maximize their utility. This work attempts to develop a dialogue agent that can deal with dynamically changing user goals; Goal driven module (GDM) has been incorporated with the dialog manager (DM) to track user goals and update them accordingly. A Dynamic Goal Driven Dialogue Agent (DGDVA) has been developed by incorporating a Dynamic Goal Driven Module (GDM) on top of a deep reinforcement learning based dialogue manager. In the course of a conversation, the user sentiment provides grounded feedback about agent behavior, including goal serving action. Please see the ReadeMe for more details.
Data: Deviation adapted Virtual Agent (DevAV) Dialogue-Corpus
Conversational Dialogue Corpus featured with Dynamic Task Goal.
Domain : Sales Domain, Annotated Tag : Intent Slot and User Sentiment
The proposed system contains three sub-modules namely :
1) Natural Language Understanding (NLU)
2) Dynamic Goal adapted Dialogue Manager (DM)
3) Natural Langauage Generator (NLG)
The pipline structure is as follows : (User Response) NLU --> DM --> NLG (Agent Response)
The dialogue corpus (DevVA) contains a total of 1000 dialogues (25 sample conversations + 975 grounded dialogue conversations). It includes a total of 8335 utterances, seven categories, and 18 slots. We have utilized a subset of GSMAreana [1] as the knowledge base for the corpus creation, which contains 2697 unique phone samples and 18 attributes (slots).
1) In NLU, there are two sub-modules named a) Intent and Slot tagger and b )Sentiment classifier for predicting users response schematic.
a) Intent and Slot tagger: We have utilized the joint BERT Model [2] as our IC/STM module, which achieves the accuracy of 93.11% and 87.39% for intent and slot, respectively.
b) Sentiment Classifier : We experimented with different models such as LSTM, BiLSTM, pre-trained BERT, and pre-trained XLNet, and found that the XLNet is performing best (accuracy - 96.68%) among these models. So, we integrated the pre-trained XLNet fine-tuned on the corpus (DevVA) as the Sentiment Classieir (SC) module.
2) Dynamic Goal adapted Dialogue Manager (DM): DM consists of mainly two layers: i) Goal Manager and ii) Dialogue Policy Learner. Goal Manager (GM) tracks discrepancy (goal deviation) for the current dialogue state and formulates a new goal as the observed discrepancy and dialogue state. The dialogue policy with architecture as Fig 4 is optimized using a Reinforcement learning algorithm (DQN). Also, the training loop has been shown in Fig 7. For more details, please see the paper and the GDM Incorporated DPL algorithm.
3) For NLG, the proposed system utilizes a template-based NLG system with an average of 4 response templates per action.
Details will be provided soon.
abhisektiwari2014@gmail.com
- Baichoo A. Kaggle GSMArean; 2017 https://www.kaggle.com/arwinneil/gsmarena-phone-dataset
- Bert for joint intent classification and slot filling 2019 https://arxiv.org/abs/1902.10909