Style transfer in NLP: a framework and multilingualanalysis with Friends TV series

This is the repository of "Style transfer in NLP: a framework and multilingual analysis with Friends TV series" paper.

Style transfer is an important and a rapidly developing of Natural Language Processing. This days more and more methods and models are proposed which allow us to generate text in predefined style. In this paper we propose a framework for style transfer of "Friends" TV series. The trained models are able to mimic one of 6 main characters of this famous TV-series in English and Russian. We also present a dialogue dataset of "Friends" subtitles in English and its Russian automatic translation. In addition to that we perform a multilingual comparison of "Friends" style transfer in the two considered languages.

Content

bot

This folder contains data for Telegram-bot:

• data - DB for storing state of each chat, rating given to each message; paths to models and log
• models - Folder template holding the pre-trained models
• ui - Utilities for enhancing UI
• utils - Database control, Model uploader and Rating
• main.py - The main file to start bot itself

data

Folder folder contains all output datasets we have:

• bigram_pics - pictures of frineds without background
• data_for_tone_analysis - statistics of tone analysis from positive and negative words
• generated - phrases generated by GPT3-Large
• questions - quections in English and Russian for mannual assessment of generated phrases
• scripts - all scripts with speakers' annotation and phrases of all friends in English and Russian
• train_data - train data for two step finetuning of GPT3-Large models split in 9 to 1 ratio (monologues and cleaned replics) in English and Russian

utils

The folder folder contains all Jupiter notebooks:

• bigrams_trigrams - a notebook to create bigrams and trigrams for each friend
• binary_classifier - notebooks for Bianry Classifiers (Training + Evaluation)
• multilabel_classifier - a notebook for Multilabel Classifiers
• Other files:
  • Parser.ipynb - parses website with series' scripts
  • Data_preparation.ipynb - cleans parsed scripts from irrelevant symbols and words
  • Statistics.ipynb - gets statistics of most frequently used words and visualizes it
  • Phrases_Preprocessing.ipynb - gets phrases that are common for friends and hard to detect by a classifier in English
  • Ru_Phrases_Preprocessing.ipynb - gets phrases that are common for friends and hard to detect by a classifier in Russian
  • Text_Analysis.ipynb - brief analysis of most frequently used words
  • Metrics.ipynb - preprocessing and furhter tone anaylis

The checkpoints of the trained models stored here.