eqiihuu / dialog_classifier

Sentence classification for dialog management

Domain Classifier

---

Author

• Qi Hu Email: [e-mail]

---

Introduction

• This is a NLP machine learning project
• The function of this model is to classify a single sentence to a certain domain, for example "I' like some Thai food" will be classified to the RESTAURANT domain.
• Dataset is not included in this project. You can prepare your own data. The format of data should be as follow:
  1. training, testing and validating data are stored separately in 3 files. Each file is composed of multiple lines (Each line is a data sample).
  2. For each sample of data, there are 3 parts separated by TAB sign ('\t'): domain (label), sentence, word_dictionary_feature (this is a pre-processed feature, representing the high level feature of each word in the sentence)

Algorithm

• Use word embedding for each raw input word
• Use CNN model to extract feature from the sentence (after embedding layer)
• Combine the CNN feature and word_dictionary_feature together and feed into the fully-connected layer
• Use Softmax for the final classification
• I tried different feature, parameters and network structures (different position of drop layer and fully-connected layer)

---

Usage

• Train: train.py
• Test: test.py
• If you want to try different parameters, change them in the train.py/test.py scripts. If you want to tried different models (network structure), change them in corresponding script in the folder 'model'.

Files

1. model (folder): Different CNN models. Each file defines one model, and the differences lie in the features and network structure.
2. utils (folder): Some tool functions for data pre-processing, data loading, error analysis, raw data analysis
3. train.py: Train the basic model with no word_dictionary_feature
4. train_vd.py: Train the model with word_dictionary_feature
5. train_cmp2vd.py: Train the model with simplified word_dictionary_feature (choose only one most important feature for each word)