Dialog Act Classifier in Tensorflow

Tensorflow implementation of "Sequential Short-Text Classification with Recurrent and Convolutional Neural Networks"

Requirements

• Python 3.x
• Tensorflow > 0.12
• Numpy
• Pandas

Pre-execution instructions

Datasets to download

Download following files in the program root dirctory (*.../Act-Classification-with-CNN)

• DSTC2 : download
• GloVe : download

csv generation

CSV files are generated from original DSTC2 dataset and used by the program.

Generation process

For each dataset [train | dev | test]

1. Generate a integraed json file from multiple json files in DSTC2 dataset.
2. generate a csv file using the integrated json file.

After csv files are generated, original DSTC2 dataset files and json files generated midway can be deleted.

Basic Guideline

In the root directory:

1. Unzip downloaded DSTC2 dataset.

   $ unzip dstc2-clean-transcriptions.zip
   

2. Unzip downloaded pretrained embeddings into data directory.

   $ unzip glove.txt.zip -d glove
   

3. Run data_generator.py.

   $ python3 data_generator.py -a
   

4. (Optional) Delete unnecessary files.

   $ rm -r dstc2-simpleacts flist _json
   

Directory structure

After following the guideline, the directory structure will look like:

Act-Classification-with-CNN
├── csvs
│   ├── act.csv (generated csv file)
│   ├── train.csv (generated csv file)
│   ├── dev.csv (generated csv file)
│   └── test.csv (generated csv file)
├── glove
│   └── glove.txt (pretrained embedding)
├── data_generator.py
├── DataHelper.py
├── model.py
├── train.py
├── eval.py
└── utils.py

Training

Print parameters:

$ python3 train.py --help

  optional arguments:

  -h, --help            show this help message and exit
  --train_data_file TRAIN_DATA_FILE
 	                    Data source for the train data.
  --dev_data_file DEV_DATA_FILE
                        Data source for the dev data.
  --act_data_file ACT_DATA_FILE
                        Data source for the act data.
  --embed_size EMBED_SIZE
                        Dimensionality of character embedding (default: 300)
   --filter_size FILTER_SIZE
                        Filter sizes (default: 3)
   --num_filters NUM_FILTERS
                        Number of filters per filter size (default: 500)
   --dropout_keep_prob DROPOUT_KEEP_PROB
                        Dropout keep probability (default: 0.5)
   --history_size1 HISTORY_SIZE1
                        History size of first FNN layer (default: 2)
   --history_size2 HISTORY_SIZE2
                        History size of second FNN layer (default: 1)
   --allow_soft_placement [ALLOW_SOFT_PLACEMENT]
                        Allow device soft device placement
   --noallow_soft_placement
   --log_device_placement [LOG_DEVICE_PLACEMENT]
                        Log placement of ops on devices
   --nolog_device_placement
   --num_epochs NUM_EPOCHS
                        Number of training epochs (default: 150)
   --evaluate_every EVALUATE_EVERY
                        Evaluate model on dev set after this many steps (default: 233550)
   --checkpoint_every CHECKPOINT_EVERY
                        Save model after this many steps (default: 233550)
   --num_checkpoints NUM_CHECKPOINTS
                        Number of checkpoints to store (default: 10)
   --num_classes NUM_CLASSES
                        Number of output act classes (default:23)
   --max_sentence_len MAX_SENTENCE_LEN
                        Max sentence length (default:30)
   --vocab_size VOCAB_SIZE
                        Vocabulary size (default:400000)
   --glove GLOVE        Using glove (default:True)

Train:

$ python3 train.py

Evaluating

$ python3 eval.py --checkpoint_dir="./runs/150437919/checkpoints/" 

Output CSV files

• act_accuracy.csv : This file contains [ number of labels | number of predictions | correct prediction | recall | precision | f1score | remark ] for each act.
• prediction.csv : This file contains [ model's predictions | actual answers ] for each sentence. If you want to see only the wrong samples, set the answer column to the filter column and check the zero label sample.

References

1. Ji Young Lee et al., Sequential Short-Text Classification with Reccurent and Convolutional Neural Networks, 2016
2. DSTC2_handbook http://camdial.org/~mh521/dstc/downloads/handbook.pdf
3. GloVe https://nlp.stanford.edu/pubs/glove.pdf