Accelerating KBQA via Logical-Question Bidirectional Reranking

ACL ARR 2025 May Submission1781 Authors

Abstract:

Semantic parsing based on large language models aims to transform natural language questions into logical forms to support the generation of answers. Although beam search-based decoding strategies are widely adopted to ensure the golden logical form appears in candidate lists, the golden logical form often fails to rank first, which raises execution time and answer error rate. To solve this problem, we propose RankKBQA, a flexible plugin that optimizes for speed and effectiveness in KBQA via a logical-question bidirectional reranking framework. Specifically, RankKBQA
first converts generated logical forms into corresponding questions via a fine-tuned PLM-based transcriber, and then measures question similarity with the original input, which obtains the second logical form sorting list. Finally, we utilize a bidirectional reranking algorithm to merge the original sorting with the new sorting. Through the above steps,
the proposed framework raises the golden logical form ranking list, simultaneously improving execution efficiency (most +42.1 speedup) and QA accuracy (most +2.9 F1) by reducing the candidate search space. Our code is available at https://anonymous.4open.science/r/RankKBQA-F702/.

Paper Type: Long

Research Area: Question Answering

Research Area Keywords: Semantic Parsing, Logical Form Generation, Beam Search, Question Transcription, KBQA

Languages Studied: English

Submission Number: 1781