Generative-Discriminative-Persona-Classification-Model

Quick Start

For Persona Classification task, we design Traditional Machine Learning Model as baseline. Three main models have been implemented in our project based on IBM/pytorch-seq2seq, which are:

• Discriminative Model
• Generative Model
• Hierarchical Model

You can enter one of these folders, and type python main.py in command line, default parameters are set carefully by us.

Also, you can use python main.py -h to check out more details:

  -h, --help                                show this help message and exit
  --path PATH                               Input data path.
  --num_sentence NUM_SENTENCE               Number of sentences in every dialog
  --max_len MAX_LEN                         Choose max_len of Encoder
  --hidden_size HIDDEN_SIZE                 Choose the hidden size
  --num_spk NUM_SPK                         Choose the speaker size
  --spk_embed_size SPK_EMBED_SIZE           Define the speaker embedding size
  --cuda CUDA                               Use cuda or not
  --lr LR                                   Choose the learning rate
  --batch_size BATCH_SIZE                   Choose the batch_size for train, dev and test

Models

Discriminative/Generative/Hierarchical Model

• main.py: to parse arguments and launch experiments
• load_model.py: resume checkpoints for continuing training or parameters tuning
• seq2mlp & seq2seq: folders for model implementation
  • dataset/: scripts for reading and parsing the corpus
  • models/: networks via pytorch, such as EncoderRNN, DecoderRNN and Seq2Seq
  • loss/: overwrite nn.Loss, nn.NLLLoss in pytorch, for bidirectional loss computing
  • optim/: optimaters for networks
  • evaluator/: evaluator and predictor for networks
  • util/: checkpoint saver

Traditional Machine Learning Model

• tfidf.py: traditional models for tfidf + LogisticRegression / RandomForest / Naive Bayes
• metric.py: confusion metrix generation for traditional models

Corpus

We use the text in Big Theory Season 1 - 9 remaining 6 main characters and regarding other characters as 'Others' category. We bring two data formats:

• .tsv: {context} \t {speaker} \t {response}
• .csv: {speaker}, {context}, {speaker}, {context}, …, {speaker}, {response}

They are in data/ folder, and we also provide data pre-processor to convert your own raw data into these two formats, which are in data/utils/ folder:

• data/utils/
  • split.py: split the raw text as 8:1:1 for every scene
  • parse_raw.py: convert raw text to csv format data with tokenization
  • csv2tsv.py: convert csv format data into tsv format
  • tsv2csv.py: convert tsv format data into csv format
  • augmentation.py: use slide window to generate dialogs like 1234, 2345, 3456 instead of 1234, 5678
  • clip.ipynb: clip the too short and too long dialogs

Visualization

• visual_speaker.py: speaker embedding visualization
• plot_confusion_matrix.py

Environment

• Python 2.7
• Pytorch 0.3.0+ with CUDA 8.0 / CPU
• Pytorchtext 0.2.1

Or create a virtual env and use pip install -r requirements.txt to install them.

Partners

• AuCson
• RshCaroline