Problem Statement Smart Customer Support Agent for Amazon Product Queries
Problem Description We intend to develop an intelligent customer support agent capable of efficiently answering customer queries related to Amazon products. Given a user's question about a product, the agent should generate accurate and informative responses by analyzing the available dataset of questions and answers. Currently, we are starting with developing an agent, which can answer Yes/No type responses for customer queries.
Constraints Dataset: We have decided to use Amazon Question and Answer Data from the given dataset having questions and answers about products. Furthermore, we have selected the Software category for implementation. https://cseweb.ucsd.edu/~jmcauley/datasets/amazon/qa/ {
'questionType': 'yes/no',
'asin': '0594033926',
'answerTime': 'Jan 5, 2015',
'unixTime': 1420444800,
'question': 'Does it fit Nook GlowLight?',
'answerType': 'N',
'answer': 'No. The nook color or color tablet'
}
Accuracy and Efficiency: The system must provide accurate answers while minimizing response time to enhance user satisfaction.
User Feedback: Incorporate a feedback loop to continuously improve the model's performance based on user interactions.
In future, we would like to develop a real-time and scalable model to handle live chat with a large number of concurrent users.
Model State Representation: In this context, the state could consist of the current user query, the context (previous interactions or questions), and the current session history. We are using embeddings to convert these into a suitable format for the DRL model.
Action Space: Actions represent various response strategies, such as:
Providing a direct answer from the dataset.
Asking for clarification if the query is ambiguous.
Offering alternative product recommendations.
Initiating a search for the answer if it's not readily available.
Reward Function: Our reward function evaluates the quality of the agent's responses. The reward will be based on:
The accuracy of the response (e.g., whether it answered the user's query correctly).
Response time (rewarding timely responses).
User satisfaction feedback (if available - can be added in the future).
