Transfer Learning Enhances Chatbot Models Across Domains

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Aug 9 2024

In a paper published in the journal Knowledge-Based Systems, researchers explored the feasibility of using transfer learning to improve chatbot models for customer service across various domains. Training models on one domain and transferring knowledge to another showed that most models showed significant improvements, particularly in data-scarce areas.

The study involved interactions with 19 companies from industries like e-commerce and telecommunications, with 16 domains showing statistically significant results. Additionally, the paper discussed the potential deployment of these models on physical robots like Softbank's "Pepper" and Temi.

Background

Past work in chatbot development has focused on enhancing customer service through techniques like deep learning (DL), reinforcement learning (RL), and attention-based modeling. Research has demonstrated that chatbots can improve customer interactions, especially when integrated with human-in-the-loop strategies. They also benefit when used with social robots, such as Pepper. This background highlights the potential of transfer learning in advancing chatbot capabilities.

Methodology and Implementation

This section details the methodology of the experiments, beginning with data collection and processing to create datasets for training chatbot models. A dataset of 3,003,124 tweets and responses from 19 different customer support accounts was used, with non-English text removed and data cleaned to eliminate profanity. The dataset covers various domains, including airlines, consumer technology, telecommunications, and more, with each domain contributing unique conversations and file sizes after preprocessing.

The team implemented machine learning (ML) and transfer learning approaches. First, a self-attention neural network was trained on data from each domain to forecast the subsequent words in customer service dialogues. This setup aimed to refine the model's ability to handle various customer service interactions.

Transfer learning was implemented using weights from a model trained in one domain. This approach aimed to enhance performance in a different but related domain. The process aimed to determine whether transferring knowledge from one domain could enhance chatbot responses in another, especially in domains with limited data. The experiments included comparisons between classical and transfer learning to evaluate performance improvements.

The implementation extended to robotic platforms, specifically the Temi personal assistant robot and the Pepper semi-humanoid robot by Softbank Robotics. These robots were chosen for their relevance in customer service applications. The chatbots were embedded within robot-compatible wrappers, with Pepper requiring bridging software due to its reliance on older ML libraries. While supporting modern libraries, Temi faced challenges with speech synchronization, which were addressed using Harvard sentences to measure speech quality and command timing.

Inference speed was benchmarked using a consumer-level Nvidia RTX 2080Ti graphics processing unit (GPU). Three hundred sixty-one chatbots were trained, including 19 trained classically and 342 through transfer learning, with training taking approximately ten days per model. The study concluded by deploying these chatbots on robotic platforms, enhancing their ability to interact with customers in real time.

Results Overview

The network topology was initially explored by tuning 16 configurations, focusing on the number of attention heads and dense layer neurons. After 10 epochs on all data, the differences in validation results were marginal, with the lowest loss observed using 8 attention heads and 256 neurons. Researchers selected this configuration for its simplicity and consistent performance, and they are using it for subsequent chatbot experiments.

The smaller transformers used in this study are considerably less complex than state-of-the-art models like GPT-2 and GPT-3. These advanced models are significantly larger and more intricate by comparison. Despite this, the chosen vocabulary size of 30,000 words was deemed sufficient for covering multiple domains while staying within computational limits.

Transfer learning proved effective, with 13 of the 19 domains showing lower loss and 15 showing higher accuracy than classical learning. Specific domains like PlayStation and Xbox benefited significantly from knowledge transfer due to their similarities. However, the largest datasets saw less improvement, suggesting that transfer learning is particularly valuable in data-scarce situations. Statistical analysis with the Wilcoxon signed-rank test and Cohen's d-effect size confirmed that transfer learning generally resulted in significant improvements. This approach was particularly effective in reducing sparse categorical cross-entropy loss.

Finally, the chatbots were implemented on consumer robots Temi and Pepper to test their feasibility in real-world applications. Challenges like speech synchronization on Temi and outdated libraries on Pepper were encountered and addressed during implementation. Temi's speech speed was measured to ensure accurate text-to-speech transitions, while Pepper required a Python version bridge to interface with modern ML libraries. These solutions enabled the smooth operation of the chatbots on both robotic platforms.

Conclusion

To sum up, this study demonstrated that knowledge transfer between chatbots across different domains significantly improved their performance in most cases, with statistical tests confirming the significance of these improvements. Despite challenges like limitations in Python compatibility, effective solutions were implemented to enable chatbot deployment on physical robots. The experiments also revealed that the chatbots could engage in more natural and empathetic communication, adapting their responses to the tone and content of user inputs.