Skip to content

Discourse Mutual Information (DMI) - Proposed in NAACL 2022 Paper "Representation Learning for Conversational Data using Discourse Mutual Information Maximization"

Notifications You must be signed in to change notification settings

bsantraigi/2022-DMI-Mirror

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

InfoNCE-Dialog

PyTorch implementation for Contrastive Predictive Coding (https://arxiv.org/pdf/1807.03748.pdf) for Dialogue Data (text modality).

TODO:

  • Verification
    • Dataset + Dataloader
    • Model: Transformer Encoder + CLS embedding
    • Make sure train and validation loss are comparable. (proper normalization)
    • Plot train vs valid loss

Requirements

  • gdown
  • wandb
  • transformers
  • datasets
  • pytorch

Files

├── corr.py
├── create_data.py
├── create_test.py
├── data_deb.py
├── data_json.py
├── data_persona.py
├── data_processed.py
├── datautils
│   ├── data_dialog.py
│   ├── data_swda.py
│   └── __init__.py
├── dialog_eval_refactor.py
├── dialog_train_deb_refactor.py
├── environ.yml
├── filter_wandb_runs.py
├── finetune_pipeline.sh
├── get_data.sh
├── Makefile
├── models
│   ├── core.py
│   ├── downstream.py
│   ├── __init__.py
│   └── legacy.py
├── pretrain.py
├── pull_ckpt.sh
├── README.md
├── run_finetune.py
├── search.yaml
├── summarize_wandb_runs.py
├── tests.sh
└── utils
    ├── func_utils.py
    ├── generate_run_id.py
    ├── __init__.py
    └── task_to_keys.py

Downstream Tasks [PROBING]

Types of tasks:

  • Single utterance classification (MLP(concat(x)))
  • Context-Response similarity
    • Distance/Similarity between our representations (cosine(c, r)/L2...)
    • MLP(concat(c, r))
  • Dual-utterance classification?
    • MLP(concat(u, v))

GLUE

Adding GLUE tasks just because we can! :P

DD++

(R->R, Sim) (R->A, Sim) (A->A, Sim) (R+A->A, Sim) (R->R, MLP) (R->A, MLP) (A->A, MLP) (R+A->A, MLP)
RoBERTa
BERT
T5
GPT-2
DialoGPT
Blender
DEB
SMI

All

Models DD++ DD++(Adversarial) SWDA Banking77 E-Intent Mutual Mutual++ Ubuntu-DSTC7
Sim, MLP Sim, MLP Sim, MLP Sim, MLP Sim, MLP
RoBERTa
BERT
T5
GPT-2
DialoGPT
Blender
DEB
SMI

SWDA

  • Total number of utterances: 199740
  • Max utterance length: 132
  • Mean utterance length: 9.62
  • Total Number of dialogues: 1155
  • Max dialogue length: 457
  • Mean dialogue length: 172.94
  • Vocabulary size: 22301
  • Number of labels: 41
  • Number of speakers: 2

Train set

  • Number of dialogues: 1115
  • Max dialogue length: 457
  • Mean dialogue length: 172.55
  • Number of utterances: 192390

Test set

  • Number of dialogues: 19
  • Max dialogue length: 330
  • Mean dialogue length: 214.63
  • Number of utterances: 4078

Val set

  • Number of dialogues: 21
  • Max dialogue length: 299
  • Mean dialogue length: 155.81
  • Number of utterances: 3272

About

Discourse Mutual Information (DMI) - Proposed in NAACL 2022 Paper "Representation Learning for Conversational Data using Discourse Mutual Information Maximization"

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published