计算机科学
稳健性(进化)
价值(数学)
钥匙(锁)
数据挖掘
变化(天文学)
分类学(生物学)
水准点(测量)
人工智能
价值体系
作者
Apostolos Koukouvinis,George Katsogiannis-Meimarakis,Georgia Koutrika
出处
期刊:
[Association for Computing Machinery]
日期:2026-05-18
卷期号:4 (3): 1-26
被引量:2
摘要
Value linking in text-to-SQL systems is the problem of associating values mentioned or implied within a NL query with their database entries. Despite existing efforts, their performance and comparative effectiveness remain unexplored. This work provides the first systematic study focused exclusively on value linking. First, we present a taxonomy that encapsulates the key steps and design choices for value linking. To enable a thorough analysis, we introduce VLD-Bench, a new, challenging benchmark that incorporates 18 categories of lexical and semantic discrepancies between values referenced in the queries and database entries. Our evaluation shows that state-of-the-art text-to-SQL systems suffer up to a 64% drop in execution accuracy due to value linking errors. A fine-grained analysis reveals how design choices impact robustness to linguistic variation and system latency, offering key insights and directions for future research.
科研通智能强力驱动
Strongly Powered by AbleSci AI