# HIT-SCIR at MMNLU-22: Consistency Regularization for Multilingual Spoken Language Understanding

**Authors:** Bo Zheng, Zhouyang Li, Fuxuan Wei, Qiguang Chen, Libo Qin, Wanxiang Che  
**arXiv:** [2301.02010](https://arxiv.org/abs/2301.02010)  
**Date:** 2023-01-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2301.02010v1/

## Key points

- The paper presents a method to enhance intent detection and slot filling tasks in multilingual spoken language understanding (SLU)
- Consistency regularization is proposed as a technique to enforce consistency between predicted distributions for an example and its semantically equivalent augmentation
- Hybrid data augmentation strategy is used to achieve consistency regularization
- Experiments on the MASSIVE dataset show significant improvement in performance of both intent detection and slot filling tasks
- The proposed method ranked 1st in the MMNLU-22 competition under the full-dataset setting
- Code for the method is available on GitHub for further exploration
- This research contributes to advancing multilingual SLU by improving intent detection and slot filling accuracy.

## Summary

The paper titled "HIT-SCIR at MMNLU-22: Consistency Regularization for Multilingual Spoken Language Understanding" presents a method to enhance the performance of intent detection and slot filling tasks in multilingual spoken language understanding (SLU). The authors propose using consistency regularization, which enforces the predicted distributions for an example and its semantically equivalent augmentation to be consistent. This is achieved through a hybrid data augmentation strategy. The experiments conducted on the MASSIVE dataset, under both full-dataset and zero-shot settings, demonstrate that the proposed method significantly improves the performance of both intent detection and slot filling tasks. In fact, their system ranked 1st in the MMNLU-22 competition under the full-dataset setting. The authors provide their code on GitHub for further exploration. Overall, this research contributes to advancing multilingual SLU by introducing an effective approach for enhancing intent detection and slot filling accuracy.