Skip to content

Latest commit

 

History

History
106 lines (87 loc) · 3.92 KB

README.md

File metadata and controls

106 lines (87 loc) · 3.92 KB
Raw

PheMT: A Phenomenon-wise Dataset for Machine Translation Robustness on User-Generated Contents

Introduction

PheMT is a phenomenon-wise dataset designed for evaluating the robustness of Japanese-English machine translation systems. The dataset is based on the MTNT dataset[1], with additional annotations of four linguistic phenomena common in UGC; Proper Noun, Abbreviated Noun, Colloquial Expression, and Variant. COLING 2020.

See the paper for more information.

New!! ready-to-use evaluation tools are now available! (Feb. 2021)

About this repository

This repository contains the following.

.
├── README.md
├── mtnt_approp_annotated.tsv # pre-filtered MTNT dataset with annotated appropriateness (See Appendix A)
├── proper
│   ├── proper.alignment # translations of targeted expressions
│   ├── proper.en # references
│   ├── proper.ja # source sentences
│   └── proper.tsv
├── abbrev
│   ├── abbrev.alignment
│   ├── abbrev.en
│   ├── abbrev.norm.ja # normalized source sentences
│   ├── abbrev.orig.ja # original source sentences
│   └── abbrev.tsv
├── colloq
│   ├── colloq.alignment
│   ├── colloq.en
│   ├── colloq.norm.ja
│   ├── colloq.orig.ja
│   └── colloq.tsv
├── variant
│   ├── variant.alignment
│   ├── variant.en
│   ├── variant.norm.ja
│   ├── variant.orig.ja
│   └── variant.tsv
└── src
    └── calc_acc.py # script for calculating translation accuracy

Please feed both original and normalized versions of source sentences to your model to get the difference of arbitrary metrics as a robustness measure. Also, we extracted translations for expressions presenting targeted phenomena. We recommend using src/calc_acc.py to measure the effect of each phenomenon more directly with the help of translation accuracy.

USAGE: python calc_acc.py system_output {proper, abbrev, colloq, variant}.alignment

Basic statistics and examples from the dataset

  • Statistics
Dataset # sent. # unique expressions (ratio) average edit distance
Proper Noun 943 747 (79.2%) (no normalized version)
Abbreviated Noun 348 234 (67.2%) 5.04
Colloquial Expression 172 153 (89.0%) 1.77
Variant 103 97 (94.2%) 3.42
  • Examples
- Abbreviated Noun

original source : 地味なアプデ (apude, meaning update) だが
normalized source : 地味なアップデート (update) だが
reference : That’s a plain update though
alignment : update

- Colloquial Expression

original source : ここまで描いて飽きた、かなちい (kanachii, meaning sad)
normalized source : ここまで描いて飽きた、かなしい (kanashii)
reference : Drawing this much then getting bored, how sad.
alignment : sad

Citation

If you use our dataset for your research, please cite the following paper:

@inproceedings{fujii-etal-2020-phemt,
    title = "{P}he{MT}: A Phenomenon-wise Dataset for Machine Translation Robustness on User-Generated Contents",
    author = "Fujii, Ryo  and
      Mita, Masato  and
      Abe, Kaori  and
      Hanawa, Kazuaki  and
      Morishita, Makoto  and
      Suzuki, Jun  and
      Inui, Kentaro",
    booktitle = "Proceedings of the 28th International Conference on Computational Linguistics",
    month = dec,
    year = "2020",
    address = "Barcelona, Spain (Online)",
    publisher = "International Committee on Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.coling-main.521",
    pages = "5929--5943",
}

Reference

[1] Michel and Neubig (2018), MTNT: A Testbed for Machine Translation of Noisy Text.