ConFiguRe

This repository presents the dataset and baseline implementation for the COLING 2022 long paper (oral): "ConFiguRe: Exploring Discourse-level Chinese Figures of Speech".

Overview

ConFiguRe is a Chinese corpus for Contextualized Figure Recognition concerning 12 commonly used rhetoric figures. It involves both extracting a figurative unit from the discourse-level context and classifying this unit into the right figure type. On the basis of ConFiguRe, three tasks are devised as benchmarks, i.e. figure extraction, figure type classification and figure recognition.

ConFiguRe Dataset

ConFiguRe dataset situates within the data folder of this repository. Train, valid and test set can be find in data/train.json, data/valid.json, data/test.json, respectively. The 12 figure types in ConFiguRe are Metaphor/Simile, Personification, Metonymy, Hyperbole, Irony, Synaesthesia, Rhetorical question, Parallelism, Duality, Repetition, Antithesis, and Quote. For definition and example of each figure type, please refer to the original paper.

Data Format

Each instance of ConFiguRe comprises two parts: fragment and units. The former is piece of Chinese literary works containing figures of speech (fos). The latter is a collection of all the figurative units in the fragment.

A sample data point in the json file is as follows:

{
    "北平的四季-郁达夫_片段4": {
        "fragment": "到了下雪的时候哩，景象当然又要一变。早晨从厚棉被里张开眼来，一室的清光会使你的眼睛眩晕。在阳光照耀之下，雪也一粒一粒地放起光来了，蛰伏得很久的小鸟，在这时候会飞出来觅食振翎，谈天说地般吱吱地叫个不休。数日来的灰暗天空，愁云一扫，忽然变得澄清见底，翳障全无；于是，年轻的北方住民，就可以营屋外的生活了——溜冰，做雪人，赶冰车雪车……就在这一种日子里最有劲儿。我曾于这一种大雪时晴的傍晚，和几位朋友跨上跛驴，出西直门上骆驼庄去过过一夜。北平郊外的一片大雪地，无数枯树林，以及西山隐隐现现的不少白峰头，和时时吹来的几阵雪样的西北风，所给予人的印象实在是深刻、伟大，神秘到了不可以言语来形容。直到了十余年后的现在，我一想起当时的情景，还会打一个寒战而吐一口清气，如同在钓鱼台溪旁立着的一瞬间一样。",
        "units": [
            {
                "figurativeUnit": "在阳光照耀之下，雪也一粒一粒地放起光来了，",
                "fos": "夸张 (Hyperbole)",
                "begin": 44,
                "end": 65
            },
            {
                "figurativeUnit": "在这时候会飞出来觅食振翎，谈天说地般吱吱地叫个不休。",
                "fos": "比拟 (Personification)",
                "begin": 74,
                "end": 100
            }
        ]
    },
}

Corpus Statistics

ConFiguRe includes 4,192 fragments and 9,010 figurative units. Train, valid and test set is split according to the proportion of 7:1:2. Detailed information for each figure type is demonstrated below:

Environment

Overall Information

• transformers version: 4.18.0
• Platform: Linux-5.4.0-124-generic-x86_64-with-glibc2.17
• Python version: 3.8.13
• Huggingface_hub version: 0.5.1
• PyTorch version (GPU?): 1.11.0+cu113 (True)

Reproduction

conda env create -n configure python=3.8.13 -y
pip install -r requirements.txt

Code Layout

We provide a general overview of our code repo. For detailed annotation, please refer to the comments in each file.

/
├── configs/ # yaml style configs
│   ├── accelerate_config.yaml # sample config for huggingface accelerate module
│   ├── default.yaml # default training config
│   ├── hydra/
│   │   └── job_logging/
│   │       └── custom.yaml # handles auto logging
│   └── model_args/ # task specific config
│       ├── Classification.yaml
│       ├── ClassificationContext.yaml
│       ├── CRF.yaml
│       ├── End2end.yaml
│       └── Extraction.yaml
├── delimit_clause.py # logic for delimiting clauses, which would serve as basis for `figurative unit`
├── dataset/ # handles dataset loading
│   ├── __init__.py
│   ├── DatasetForClassification.py
│   ├── DatasetForClassificationContext.py
│   ├── DatasetForCRF.py
│   ├── DatasetForExtraction.py
│   └── DatasetForRecognition.py
├── main.py # template for training
├── metrics/ # handles metric calculation
│   ├── __init__.py
│   ├── MetricForClassification.py
│   ├── MetricForExtraction.py
│   ├── MetricForRecognition.py
│   └── MetricForRecognitionCRF.py
├── model/ # implementation for models, including forward logic
│   ├── __init__.py
│   ├── BertForFigClassification.py
│   ├── BertForFigClassificationContext.py
│   ├── BertForFigExtraction.py
│   ├── BertForFigExtractionContrast.py
│   ├── BertForFigRecognition.py
│   └── BertForFigRecognitionCRF.py
├── scripts/ # helpful scripts
│   ├── debug_hydra.sh
│   ├── eval.sh
│   └── run.sh
├── task/ # task specific logic, called in `main.py`
│   ├── __init__.py
│   ├── Classification.py
│   ├── ClassificationContext.py
│   ├── CRF.py
│   ├── End2end.py
│   ├── Extraction.py
│   ├── ExtractionContrast.py
│   └── ExtractionCRF.py
└── train/ # Trainer-like module, handles training + eval steps
    ├── __init__.py
    ├── TrainClassifier.py
    ├── TrainExtraction.py
    ├── TrainRecognition.py
    └── TrainRecognitionCRF.py

Hydra Config

We leverage the hydra module to store hyperparameters with respect to each model, to enable auto-logging and to modularize our repo. It is helpful to have a basic idea of the hydra configuration style.

Scripts

Useful scripts are under the code/scripts folder.

Navigate to code/ folder, and run the scripts, eg.

# cd code
bash scripts/run.sh

Experimental Results

Citation

If you use ConFiguRe in your work, please cite our paper:

@inproceedings{zhu-etal-2022-configure,
    title = "{C}on{F}igu{R}e: Exploring Discourse-level {C}hinese Figures of Speech",
    author = "Zhu, Dawei  and
      Zhan, Qiusi  and
      Zhou, Zhejian  and
      Song, Yifan  and
      Zhang, Jiebin  and
      Li, Sujian",
    booktitle = "Proceedings of the 29th International Conference on Computational Linguistics",
    month = oct,
    year = "2022",
    address = "Gyeongju, Republic of Korea",
    publisher = "International Committee on Computational Linguistics",
    url = "https://aclanthology.org/2022.coling-1.298",
    pages = "3374--3385",
}

Contact

If you have any questions, feel free to open an issue, or contact

• dwzhu@pku.edu.cn;

• lisujian@pku.edu.cn;

For implementation details

• zhouzhejian@pku.edu.cn;