The article presents a new method for answering tourism-related questions that seek Point-of-Interest (POI) recommendations. Traditional methods of encoding each pair of question and POI become inefficient with an increase in the number of candidates. To tackle this, the authors propose treating the task as a dense vector retrieval problem. Questions and POIs are encoded separately and the most relevant POIs for a question are retrieved based on embedding space similarity. The approach uses pretrained language models to encode textual information and a location encoder to capture spatial information of POIs. The method has been tested on a real-world tourism QA dataset and has shown to outperform previous methods across all metrics.
Publication date: 5 Jan 2024
Project Page: https://github.com/haonan-li/LAMB
Paper: https://arxiv.org/pdf/2401.02187
