Location Aware Modular Biencoder for Tourism Question Answering

• DOI: 10.48550/arxiv.2401.02187
• Publication Date: 2024-01-01
• Authors: Haonan Li; Martin Tomko; Timothy Baldwin
• Keywords: ENCODE; Point of interest; Similarity (geometry); Code (set theory); Baseline (sea)

Answering real-world tourism questions that seek Point-of-Interest (POI) recommendations is challenging, as it requires both spatial and non-spatial reasoning, over a large candidate pool. The traditional method of encoding each pair question POI becomes inefficient when the number candidates increases, making infeasible for applications. To overcome this, we propose treating QA task dense vector retrieval problem, where encode POIs separately retrieve most relevant by utilizing embedding space similarity. We use pretrained language models (PLMs) to textual information, train location encoder capture information POIs. Experiments on dataset demonstrate our approach effective, efficient, outperforms previous methods across all metrics. Enabled architecture, further build global evaluation baseline, expanding search 20 times compared work. also explore several factors impact model's performance through follow-up experiments. Our code model are publicly available at https://github.com/haonan-li/LAMB.