Introduction & Motivation
The adoption of artificial intelligence in various sectors, including the human resource (HR) domain, is on the rise. With traditional question-answering chatbots needing manual design for user intents, a shift towards generative question-answering (GQA) chatbots, capable of generating natural responses and comprehending complex queries, is being observed. However, handling lengthy input texts, typical of user queries and contextual information, remains a challenge due to increased computational complexity. This study aims to address this challenge by leveraging the LongT5 model, designed for longer sequences, for the development of an HR-specific GQA chatbot.
Research Questions
The primary research questions we aim to answer are:
1. Can the use of the LongT5 model efficiently manage lengthy input texts for an HR-specific GQA chatbot?
2. How does the quality of the training dataset impact the chatbot's performance in real-wolrd senarios, and how can it be optimized for more effective results?
3. What is the comparative performance of LongT5 and T5 models under varying data quality conditions?
Methods
To answer these research questions, we have constructed three distinct datasets consisting of structured questions, real user questions, and manually corrected context by domain experts. Both LongT5 and T5 models are trained on these datasets to evaluate their performance under varying data quality conditions. The model performance will be evaluated and compared to determine the most suitable model for an HR chatbot application.
| Name | Type | Size | Last Modification | Last Editor |
|---|---|---|---|---|
| Tao Xiang GR-Kickoff.pptx | 2,50 MB | 29.01.2024 | ||
| Tao Xiang GR_final_pres.pdf | 1,35 MB | 29.01.2024 | ||
| Tao Xiang GR_Report.pdf | 2,33 MB | 29.01.2024 |