Source code for AAAI20 "Generating Persona Consistent Dialogues by Exploiting Natural Language Inference".

Overview

Generating Persona Consistent Dialogues by Exploiting Natural Language Inference

Source code for RCDG model in AAAI20 Generating Persona Consistent Dialogues by Exploiting Natural Language Inference, a natural language inference (NLI) enhanced reinforcement learning dialogue model.

Requirements:

The code is tested under the following env:

  • Python 3.6
  • Pytorch 0.3.1

Install with conda: conda install pytorch==0.3.1 torchvision cudatoolkit=7.5 -c pytorch

This released code has been tested on a Titan-XP 12G GPU.

Data

We have provided some data samples in ./data to show the format. For downloading the full datasets, please refer to the following papers:

How to Run:

For a easier way to run the code, here the NLI model is GRU+MLP, i.e. RCDG_base, and we remove the time-consuming MC search.

Here are a few steps to run this code:

0. Prepare Data

python preprocess.py -train_src data/src-train.txt -train_tgt data/tgt-train.txt -train_per data/per-train.txt -valid_src data/src-val.txt -valid_tgt data/tgt-val.txt -valid_per data/per-val.txt -train_nli data/nli-train.txt -valid_nli data/nli-valid.txt -save_data data/nli_persona -src_vocab_size 18300 -tgt_vocab_size 18300 -share_vocab

And as introduced in the paper, there are different training stages:

1. NLI model Pretrain

cd NLI_pretrain/

python train.py -data ../data/nli_persona -batch_size 32 -save_model saved_model/consistent_dialogue -rnn_size 500 -word_vec_size 300 -dropout 0.2 -epochs 5 -learning_rate_decay 1 -gpu 0

And you should see something like:

Loading train dataset from ../data/nli_persona.train.1.pt, number of examples: 1
31432
Epoch  1, nli_step     1/ 4108; nli: 0.28125
Epoch  1, nli_step    11/ 4108; nli: 0.38125
Epoch  1, nli_step    21/ 4108; nli: 0.43438
Epoch  1, nli_step    31/ 4108; nli: 0.48125
Epoch  1, nli_step    41/ 4108; nli: 0.53750
Epoch  1, nli_step    51/ 4108; nli: 0.56250
Epoch  1, nli_step    61/ 4108; nli: 0.49062
...

2. Generator G Pretrain

cd ../G_pretrain/

python train.py -data ../data/nli_persona -batch_size 32 -rnn_size 500 -word_vec_size 300  -dropout 0.2 -epochs 15 -g_optim adam -g_learning_rate 1e-3 -learning_rate_decay 1 -train_from PATH_TO_PRETRAINED_NLI -gpu 0

Here the PATH_TO_PRETRAINED_NLI should be replaced by your model path, e.g., ../NLI_pretrain/saved_model/consistent_dialogue_e3.pt.

If , you should see the ppl comes down during training, which means the dialogue model is in training:

Loading train dataset from ../data/nli_persona.train.1.pt, number of examples: 131432
Epoch  4, teacher_force     1/ 4108; acc:   0.00; ppl: 18619.76; 125 src tok/s; 162 tgt tok/s;      3 s elapsed
Epoch  4, teacher_force    11/ 4108; acc:   9.69; ppl: 2816.01; 4159 src tok/s; 5468 tgt tok/s;      3 s elapsed
Epoch  4, teacher_force    21/ 4108; acc:   9.78; ppl: 550.46; 5532 src tok/s; 6116 tgt tok/s;      4 s elapsed
Epoch  4, teacher_force    31/ 4108; acc:  11.15; ppl: 383.06; 5810 src tok/s; 6263 tgt tok/s;      5 s elapsed
...
Epoch  4, teacher_force   941/ 4108; acc:  25.40; ppl:  90.18; 5993 src tok/s; 6645 tgt tok/s;     63 s elapsed
Epoch  4, teacher_force   951/ 4108; acc:  27.49; ppl:  77.07; 5861 src tok/s; 6479 tgt tok/s;     64 s elapsed
Epoch  4, teacher_force   961/ 4108; acc:  26.24; ppl:  83.17; 5473 src tok/s; 6443 tgt tok/s;     64 s elapsed
Epoch  4, teacher_force   971/ 4108; acc:  24.33; ppl:  97.14; 5614 src tok/s; 6685 tgt tok/s;     65 s elapsed
...

3. Discriminator D Pretrain

cd ../D_pretrain/

python train.py -epochs 20 -d_optim adam -d_learning_rate 1e-4 -data ../data/nli_persona -train_from PATH_TO_PRETRAINED_G -batch_size 32 -learning_rate_decay 0.99 -gpu 0

Similarly, replace PATH_TO_PRETRAINED_G with the G Pretrain model path.

The acc of D will be displayed during training:

Loading train dataset from ../data/nli_persona.train.1.pt, number of examples: 131432
Epoch  5, d_step     1/ 4108; d: 0.49587
Epoch  5, d_step    11/ 4108; d: 0.51580
Epoch  5, d_step    21/ 4108; d: 0.49853
Epoch  5, d_step    31/ 4108; d: 0.55248
Epoch  5, d_step    41/ 4108; d: 0.55168
...

4. Reinforcement Training

cd ../reinforcement_train/

python train.py -epochs 30 -batch_size 32 -d_learning_rate 1e-4 -g_learning_rate 1e-4 -learning_rate_decay 0.9 -data ../data/nli_persona -train_from PATH_TO_PRETRAINED_D -gpu 0

Remember to replace PATH_TO_PRETRAINED_D with the D Pretrain model path.

Note that all the -epochs are global among all stages, if you want to tune this parameter. Actually, there are 30 - 20 = 10 training epochs in this Reinforcement Training stage if the D Pretrain model was trained 20 epochs in total.

Loading train dataset from ../data/nli_persona.train.1.pt, number of examples: 131432
Epoch  7, self_sample     1/ 4108; acc:   2.12; ppl:   0.28; 298 src tok/s; 234 tgt tok/s;      2 s elapsed
Epoch  7, teacher_force    11/ 4108; acc:   3.32; ppl:   0.53; 2519 src tok/s; 2772 tgt tok/s;      3 s elapsed
Epoch  7, d_step    21/ 4108; d: 0.98896
Epoch  7, d_step    31/ 4108; d: 0.99906
Epoch  7, self_sample    41/ 4108; acc:   0.00; ppl:   0.27; 1769 src tok/s; 260 tgt tok/s;      7 s elapsed
Epoch  7, teacher_force    51/ 4108; acc:   2.83; ppl:   0.43; 2368 src tok/s; 2910 tgt tok/s;      9 s elapsed
Epoch  7, d_step    61/ 4108; d: 0.75311
Epoch  7, d_step    71/ 4108; d: 0.83919
Epoch  7, self_sample    81/ 4108; acc:   6.20; ppl:   0.33; 1791 src tok/s; 232 tgt tok/s;     12 s elapsed
...

5. Testing Trained Model

Now we have a trained dialogue model, we can test by:

Still in ./reinforcement_train/

python predict.py -model TRAINED_MODEL_PATH  -src ../data/src-val.txt -tgt ../data/tgt-val.txt -replace_unk -verbose -output ./results.txt -per ../data/per-val.txt -nli nli-val.txt -gpu 0

MISC

  • Initializing Model Seems Slow?

    This is a legacy problem due to pytorch < 0.4, not brought by this project. And the training efficiency will not be affected.

  • BibTex

     @article{Song_RCDG_2020,
     	title={Generating Persona Consistent Dialogues by Exploiting Natural Language Inference},
     	volume={34},
     	DOI={10.1609/aaai.v34i05.6417},
     	number={05},
     	journal={Proceedings of the AAAI Conference on Artificial Intelligence},
     	author={Song, Haoyu and Zhang, Wei-Nan and Hu, Jingwen and Liu, Ting},
     	year={2020},
     	month={Apr.},
     	pages={8878-8885}
     	}
    
LightSpeech: Lightweight and Fast Text to Speech with Neural Architecture Search

LightSpeech UnOfficial PyTorch implementation of LightSpeech: Lightweight and Fast Text to Speech with Neural Architecture Search.

Rishikesh (ऋषिकेश) 54 Dec 03, 2022
Constituency Tree Labeling Tool

Constituency Tree Labeling Tool The purpose of this package is to solve the constituency tree labeling problem. Look from the dataset labeled by NLTK,

张宇 6 Dec 20, 2022
This repository is home to the Optimus data transformation plugins for various data processing needs.

Transformers Optimus's transformation plugins are implementations of Task and Hook interfaces that allows execution of arbitrary jobs in optimus. To i

Open Data Platform 37 Dec 14, 2022
Multi-Scale Temporal Frequency Convolutional Network With Axial Attention for Speech Enhancement

MTFAA-Net Unofficial PyTorch implementation of Baidu's MTFAA-Net: "Multi-Scale Temporal Frequency Convolutional Network With Axial Attention for Speec

Shimin Zhang 87 Dec 19, 2022
The projects lets you extract glossary words and their definitions from a given piece of text automatically using NLP techniques

Unsupervised technique to Glossary and Definition Extraction Code Files GPT2-DefinitionModel.ipynb - GPT-2 model for definition generation. Data_Gener

Prakhar Mishra 28 May 25, 2021
Bnagla hand written document digiiztion

Bnagla hand written document digiiztion This repo addresses the problem of digiizing hand written documents in Bangla. Documents have definite fields

Mushfiqur Rahman 1 Dec 10, 2021
Code for the paper "Are Sixteen Heads Really Better than One?"

Are Sixteen Heads Really Better than One? This repository contains code to reproduce the experiments in our paper Are Sixteen Heads Really Better than

Paul Michel 143 Dec 14, 2022
A repository to run gpt-j-6b on low vram machines (4.2 gb minimum vram for 2000 token context, 3.5 gb for 1000 token context). Model loading takes 12gb free ram.

Basic-UI-for-GPT-J-6B-with-low-vram A repository to run GPT-J-6B on low vram systems by using both ram, vram and pinned memory. There seem to be some

90 Dec 25, 2022
PyTorch implementation and pretrained models for XCiT models. See XCiT: Cross-Covariance Image Transformer

Cross-Covariance Image Transformer (XCiT) PyTorch implementation and pretrained models for XCiT models. See XCiT: Cross-Covariance Image Transformer L

Facebook Research 605 Jan 02, 2023
An extensive UI tool built using new data scraped from BBC News

BBC-News-Analyzer An extensive UI tool built using new data scraped from BBC New

Antoreep Jana 1 Dec 31, 2021
txtai: Build AI-powered semantic search applications in Go

txtai: Build AI-powered semantic search applications in Go txtai executes machine-learning workflows to transform data and build AI-powered semantic s

NeuML 49 Dec 06, 2022
PyTorch source code of NAACL 2019 paper "An Embarrassingly Simple Approach for Transfer Learning from Pretrained Language Models"

This repository contains source code for NAACL 2019 paper "An Embarrassingly Simple Approach for Transfer Learning from Pretrained Language Models" (P

Alexandra Chronopoulou 89 Aug 12, 2022
2021语言与智能技术竞赛:机器阅读理解任务

LICS2021 MRC 1. 项目&任务介绍 本项目基于官方给定的baseline(DuReader-Checklist-BASELINE)进行二次改造,对整个代码框架做了简单的重构,对核心网络结构添加了注释,解耦了数据读取的模块,并添加了阈值确认的功能,一些小的细节也做了改进。 本次任务为202

roar 29 Dec 05, 2022
This project converts your human voice input to its text transcript and to an automated voice too.

Human Voice to Automated Voice & Text Introduction: In this project, whenever you'll speak, it will turn your voice into a robot voice and furthermore

Hassan Shahzad 3 Oct 15, 2021
Natural Language Processing at EDHEC, 2022

Natural Language Processing Here you will find the teaching materials for the "Natural Language Processing" course at EDHEC Business School, 2022 What

1 Feb 04, 2022
Search for documents in a domain through Google. The objective is to extract metadata

MetaFinder - Metadata search through Google _____ __ ___________ .__ .___ / \

Josué Encinar 85 Dec 16, 2022
[WWW 2021 GLB] New Benchmarks for Learning on Non-Homophilous Graphs

New Benchmarks for Learning on Non-Homophilous Graphs Here are the codes and datasets accompanying the paper: New Benchmarks for Learning on Non-Homop

94 Dec 21, 2022
A Persian Image Captioning model based on Vision Encoder Decoder Models of the transformers🤗.

Persian-Image-Captioning We fine-tuning the Vision Encoder Decoder Model for the task of image captioning on the coco-flickr-farsi dataset. The implem

Hamtech-ai 15 Aug 25, 2022
Web mining module for Python, with tools for scraping, natural language processing, machine learning, network analysis and visualization.

Pattern Pattern is a web mining module for Python. It has tools for: Data Mining: web services (Google, Twitter, Wikipedia), web crawler, HTML DOM par

Computational Linguistics Research Group 8.4k Dec 30, 2022
Mysticbbs-rjam - rJAM splitscreen message reader for MysticBBS A46+

rJAM splitscreen message reader for MysticBBS A46+

Robbert Langezaal 4 Nov 22, 2022