# CASA-NLU: Context-Aware Self-Attentive Natural Language Understanding for Task-Oriented Chatbots

**Authors:** Arshit Gupta, Peng Zhang, Garima Lalwani, Mona Diab  
**arXiv:** [1909.08705](https://arxiv.org/abs/1909.08705)  
**Date:** 2019-09-18  
**Summary page:** https://www.summarizepaper.com/en/paper/1909.08705v1/

## Key points

- The paper introduces CASA-NLU, a context-aware self-attentive NLU model for task-oriented chatbots.
- CASA-NLU aims to improve the accuracy of intent classification (IC) and slot labeling (SL) tasks in dialog systems.
- Existing NLU systems focus on individual utterances, but considering contextual information is crucial for accurate predictions in conversations.
- CASA-NLU incorporates multiple signals from previous intents, slots, dialog acts, and utterances within a variable context window.
- This additional contextual information enhances the performance of IC and SL tasks.
- CASA-NLU outperforms a recurrent contextual NLU baseline on two conversational datasets, achieving up to a 7% gain in accuracy for the IC task on one dataset.
- Even a non-contextual variant of CASA-NLU achieves state-of-the-art performance for IC task on standard public datasets like Snips and ATIS.
- The proposed model effectively leverages contextual information to improve natural language understanding in task-oriented chatbots.
- The research contributes to our understanding of context-aware NLU models and their potential applications in dialog systems.

## Summary

The paper titled "CASA-NLU: Context-Aware Self-Attentive Natural Language Understanding for Task-Oriented Chatbots" introduces a context-aware self-attentive NLU (CASA-NLU) model that aims to improve the accuracy of intent classification (IC) and slot labeling (SL) tasks in dialog systems. The existing NLU systems focus on individual utterances, leaving the context management to the dialogue management component. However, considering contextual information is crucial for accurately predicting intents and slots in a conversation. Previous research on contextual NLU has been limited in terms of the types of contextual signals used and their impact on the model. To address this limitation, the authors propose CASA-NLU, which incorporates multiple signals from previous intents, slots, dialog acts, and utterances within a variable context window. This additional contextual information is combined with the current user utterance to enhance the performance of IC and SL tasks. The CASA-NLU model outperforms a recurrent contextual NLU baseline on two conversational datasets, achieving up to a 7% gain in accuracy for the IC task on one dataset. Furthermore, even a non-contextual variant of CASA-NLU achieves state-of-the-art performance for IC task on standard public datasets like Snips and ATIS. The authors' work contributes to improving natural language understanding in task-oriented chatbots by effectively leveraging contextual information. The proposed CASA-NLU model demonstrates its superiority over existing approaches by incorporating various contextual signals into the NLU process. Overall, this research enhances our understanding of context-aware NLU models and their potential applications in dialog systems.