This repository contains the implementation of the research paper
A post-processing framework for detecting unknown intent of dialogue system via pre-trained deep neural network classifier, submitted to Knowledge-based Systems by Tingen Lin, Hua Xu
In this paper, we propose SofterMax and deep novelty detection (SMDN), a simple yet effective post-processing method for detecting unknown intent in dialogue systems based on pre-trained deep neural network classifiers
. We calibrate the confidence of DNN classifier via temperature scaling to make more reasonable/conservative prediction. Here is an example illustrating the difference between Softmax and SofterMax:
The architecture of the proposed SMDN method for the single-turn dialogue system.
The architecture of the proposed SMDN method for the multi-turn dialogue system.
- Install all required library
pip install -r requirements.txt
- Unzip and preprocess SwDA dataset (and download your GloVe embedding & change the path in file)
unzip data/swda.zip
python preprocessing_swda.py
- Get the pre-trained intent classifiers with different settings
python train.py
or you can get a single pre-trained classifier with
python train_BiLSTM.py <dataset> <proportion>
python train_BiLSTM-DOC.py <dataset> <proportion>
python train_HCNN.py <proportion>
python train_HCNN-DOC.py <proportion>
- Evaluate the results with different detection method (DOC, SofterMax, LOF, SMDN, ...)
python eval.py
or you can get a single result with
python eval_BiLSTM.py <dataset> <proportion>
python eval_BiLSTM-DOC.py <dataset> <proportion>
python eval_HCNN.py <proportion>
python eval_HCNN-DOC.py <proportion>
| % of known intents | 25% | 50% | 75% | 25% | 50% | 75% | 25% | 50% | 75% |
|---|---|---|---|---|---|---|---|---|---|
| Datasets | SNIPS | ATIS | SwDA | ||||||
| Softmax (t=0.5) | - | 6.15 | 8.32 | 8.14 | 15.3 | 17.2 | 19.3 | 18.4 | 8.36 |
| DOC | 72.5 | 67.9 | 63.9 | 61.6 | 63.8 | 37.7 | 25.4 | 19.6 | 7.63 |
| DOC (Softmax) | 72.8 | 65.7 | 61.8 | 63.6 | 63.3 | 39.7 | 23.6 | 18.9 | 7.67 |
| SofterMax | 78.8 | 70.5 | 67.2 | 67.2 | 65.5 | 40.7 | 28.0 | 20.7 | 7.51 |
| LOF | 76.0 | 69.4 | 65.8 | 67.3 | 61.8 | 38.9 | 21.1 | 12.7 | 4.50 |
| SMDN | 79.8 | 73.0 | 71.0 | 71.1 | 66.6 | 41.7 | 20.8 | 18.4 | 8.44 |
Confusion matrix for SMDN experiment results on three different datasets
