computational-stylistic-variations

Stylistic Variations in Distributional Vector Space Models

This repository contains implementations for

1. Xing Niu and Marine Carpuat. "Discovering Stylistic Variations in Distributional Vector Space Models via Lexical Paraphrases". Workshop on Stylistic Variation at EMNLP 2017.

@InProceedings{niu-carpuat:2017:StyVa,
  author    = {Niu, Xing  and  Carpuat, Marine},
  title     = {Discovering Stylistic Variations in Distributional Vector Space Models via Lexical Paraphrases},
  booktitle = {Proceedings of the Workshop on Stylistic Variation},
  year      = {2017},
  address   = {Copenhagen, Denmark},
  publisher = {Association for Computational Linguistics},
  pages     = {20--27}
}

2. Xing Niu, Marianna Martindale, and Marine Carpuat. "A Study of Style in Machine Translation: Controlling the Formality of Machine Translation Output". EMNLP 2017.

@InProceedings{niu-martindale-carpuat:2017:EMNLP2017,
  author    = {Niu, Xing  and  Martindale, Marianna  and  Carpuat, Marine},
  title     = {A Study of Style in Machine Translation: Controlling the Formality of Machine Translation Output},
  booktitle = {Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing},
  year      = {2017},
  address   = {Copenhagen, Denmark},
  publisher = {Association for Computational Linguistics},
  pages     = {2804--2809}
}

Dependencies

• Python 2.7
• gensim
• scikit-learn
• Other toolkits and data (see global.cfg)

Usage Instructions

1. Set up parameters and pointers in global.cfg.

• Get a hint of parameter settings from the Evaluation section below.
• Choose Word2vec based models (e.g. SVM-W2V-subspace) for ranking purpose.
• Choose LSA based models (e.g. PCA-LSA) for scoring purpose.

2. Initialize and test.

> bash formality/evaluate.sh

3. Calculate lexical formality for lines of text.

> bash formality/calc-formality-score.sh -i input-file -o output-file -p -s

Usage: calc-formality-score.sh -i INPUT_FILE -o OUTPUT_FILE [-s] [-l] [-p]
Optional arguments:
  -i INPUT_FILE    input file (absolute path)
  -o OUTPUT_FILE   output file (absolute path)
  -s               sort lines by formality score
  -l               only output lexical scores"
  -p               preprocess input file (tokenization and lowercasing)"

Evaluation

Method	VSM	Dimension	PCA-Data	Sub-Dim	CTRW Accuracy	BEAN Spearman's r	BEAN RMSE
SVM	W2V	10			0.776	0.566	0.424
PCA	W2V	10			0.770	0.656	0.390
SimDiff	W2V	10			0.780	0.646	0.404
SVM	W2V	300	ppdb	20	0.844	0.662	0.372
PCA	W2V	300	ppdb	20	0.829	0.660	0.389
SimDiff	W2V	300	ppdb	20	0.832	0.662	0.386
SVM	W2V	300	seed	20	0.801	0.576	0.384
PCA	W2V	300	seed	20	0.768	0.653	0.377
SimDiff	W2V	300	seed	20	0.781	0.658	0.364
SVM	LSA	10			0.737	0.661	0.361
PCA	LSA	10			0.730	0.655	0.352
SimDiff	LSA	10			0.780	0.646	0.353
SVM	LSA	300	ppdb	20	0.712	0.457	0.641
PCA	LSA	300	ppdb	20	0.671	0.498	0.545
SimDiff	LSA	300	ppdb	20	0.686	0.492	0.563
SVM	LSA	300	seed	20	0.727	0.481	0.575
PCA	LSA	300	seed	20	0.699	0.522	0.513
SimDiff	LSA	300	seed	20	0.714	0.524	0.526

• VSM: Vector Space Model
• W2V: word2vec
• LSA: Latent Semantic Analysis
• CTRW: Choose the Right Word, see paper 1
• BEAN: Blog, Email, Answers and News, see paper 2
• RMSE: Root-Mean-Square Error