Vietnamese-Spoken-Language-Understanding

This repository is based on the SOICT 2023 contest's Spoken Language Understanding track. The goal of this track is to use audio for intent detection and slot tagging.

Data

https://drive.google.com/drive/folders/1FqCmmSjMMgkYjANXY7FD6tzqsfDwZJrY

Pre-trained Model

https://huggingface.co/nguyenvulebinh/wav2vec2-large-vi

Processed Data and Model

• Due to the limited resources for training, I use Kaggle's free accelerator. As a result, the processed data and additional models are located on the Kaggle platform.
  • https://www.kaggle.com/code/noobhocai/soict2023-slu-preprocess-and-augmentation-stage-1
  • https://www.kaggle.com/code/noobhocai/soict2023-slu-wav2vec2-training-stage-1
  • https://www.kaggle.com/code/noobhocai/soict-2023-wav2vec2-n-gram-inference-stage-1
  • https://www.kaggle.com/code/huynhtuannam/soict2023-slu-training-stage-2-v2
  • https://www.kaggle.com/code/noobhocai/soict2023-slu-inference-stage-2
• Finetuned wav2vec2 model: https://huggingface.co/foxxy-hm/wav2vec2-base-finetune-vi-v6

Solution

• First, I tried adding noise and adjusting the volume to expose the model to different types of data. This helped to improve predictions and make the model more versatile.

• Next, I fine-tuned the pre-trained wav2vec2 model to get even better accuracy.

• I also created a language model using ngrams to correct any mistakes made by the automatic speech recognition system. The accuracy of the model was significantly improved by this.

• Finally, I used A Bi-model based RNN Semantic Frame Parsing Model for intent detection and slot filling. This model uses shared states to enhance task assistance and improve input interpretation accuracy.

Conclusion

• As I mentioned earlier, my limited resources prevented the model from fully converging.
• In addition, maybe I'll use other models to correct grammatical errors for better accuracy.