Trigger3:Refining Query Correction via Adaptive Model Selector

Authors: Kepu Zhang, Zhongxiang Sun, Xiao Zhang, Xiaoxue Zang, Kai Zheng, Yang Song, Jun Xu

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To validate the effectiveness and efficiency and of the proposed Trigger3 framework, we conduct experiments on two query correction datasets, using three small models and two LLMs. The results consistently demonstrate that Trigger3 achieves optimal performance and high efficiency.
Researcher Affiliation Collaboration Kepu Zhang,1 Zhongxiang Sun,1 Xiao Zhang,1,* Xiaoxue Zang,2 Kai Zheng,2 Yang Song,2 Jun Xu1 1 Gaoling School of Artificial Intelligence, Renmin University of China 2 Kuaishou Technology Co., Ltd. EMAIL EMAIL, EMAIL
Pseudocode Yes Algorithm 1: Process flow of Trigger3. 1 Input: Original query x and Trigger3 s models. 2 Output: Final corrected query yfinal.
Open Source Code Yes The source code, datasets, more experimental results and details can be found in the following link: Code https://github.com/ke-01/Trigger3.
Open Datasets Yes QQ is a publicly available search-related dataset, due to the lack of publicly available query correction datasets, we modify it as a query correction dataset. Following (Ye et al. 2023), we first use a language model to filter the queries, selecting those with a high probability of being correct. We then perform similar operations like Commercial dataset on these queries to construct a query correction dataset. The source code, datasets, more experimental results and details can be found in the following link: Code https://github.com/ke-01/Trigger3.
Dataset Splits Yes Table 1: Statistics of the used query correction datasets. Avg len #Query Error Rate Commercial 9.43 1,444,213 97.8% QQ 9.81 111,703 79.1% Valid Avg len #Query Error Rate Commercial 9.41 14,737 97.8% QQ 9.78 12,412 75.1% Test Avg len #Query Error Rate Commercial 9.43 14,737 97.8% QQ 9.79 13,791 74.7%
Hardware Specification Yes All experiments are performed on NVIDIA V100 32GB GPUs.
Software Dependencies No Our code implementation is based on Huggingface Transformers (Wolf et al. 2020) in Pytorch. No specific version numbers for Huggingface Transformers or Pytorch are provided.
Experiment Setup Yes We utilize the Adam (Kingma and Ba 2014) optimizer, setting the initial learning rate to 5e-5, the batch size to 16, and applying a cosine learning rate schedule for 3 epochs.