Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time


Automatic Evaluation Metric described in the paper UNION: An UNreferenced MetrIc for Evaluating Open-eNded Story Generation (EMNLP 2020). Please refer to the Paper List for more information about Open-eNded Language Generation (ONLG) tasks. Hopefully the paper list will help you know more about this field.



The code is written in TensorFlow library. To use the program the following prerequisites need to be installed.

  • Python 3.7.0
  • tensorflow-gpu 1.14.0
  • numpy 1.18.1
  • regex 2020.2.20
  • nltk 3.4.5

Computing Infrastructure

We train UNION based on the platform:

  • OS: Ubuntu 16.04.3 LTS (GNU/Linux 4.4.0-98-generic x86_64)

Quick Start

1. Constructing Negative Samples

Execute the following command:

cd ./Data
python3 ./ your_mode
python3 ./ your_mode
  • your_mode is roc for ROCStories corpus or wp for WritingPrompts dataset. Then the summary of vocabulary and the corresponding frequency and pos-tagging will be found under ROCStories/ini_data/entitiy_vocab.txt or WritingPrompts/ini_data/entity_vocab.txt.
  • Negative samples and human-written stories will be constructed based on the original training set. The training set will be found under ROCStories/train_data or WritingPrompts/train_data.
  • Note: currently only 10 samples of the full original data and training data are provided. The full data can be downloaded from THUcloud or GoogleDrive.

2. Training of UNION

Execute the following command:

python3 ./ --data_dir your_data_dir \
    --output_dir ./model/union \
    --task_name train \
    --init_checkpoint ./model/uncased_L-12_H-768_A-12/bert_model.ckpt
  • your_data_dir is ./Data/ROCStories or ./Data/WritingPrompts.
  • The initial checkpoint of BERT can be downloaded from bert. We use the uncased base version of BERT (about 110M parameters). We train the model for 40000 steps at most. The training process will task about 1~2 days.

3. Prediction with UNION

Execute the following command:

python3 ./ --data_dir your_data_dir \
    --output_dir ./model/output \
    --task_name pred \
    --init_checkpoint your_model_name
  • your_data_dir is ./Data/ROCStories or ./Data/WritingPrompts. If you want to evaluate your custom texts, you only need tp change your file format into ours.

  • your_model_name is ./model/union_roc/union_roc or ./model/union_wp/union_wp. The fine-tuned checkpoint can be downloaded from the following link:

Dataset Fine-tuned Model
ROCStories THUcloud; GoogleDrive
WritingPrompts THUcloud; GoogleDrive
  • The union score of the stories under your_data_dir/ant_data can be found under the output_dir ./model/output.

4. Correlation Calculation

Execute the following command:

python3 ./ your_mode

Then the correlation between the human judgements under your_data_dir/ant_data and the scores of metrics under your_data_dir/metric_output will be output. The figures under "./figure" show the score graph between metric scores and human judgments for ROCStories corpus.

Data Instruction for files under ./Data

├── Data
   └── `negation.txt`             # manually constructed negation word vocabulary.
   └── `conceptnet_antonym.txt`   # triples with antonym relations extracted from ConceptNet.
   └── `conceptnet_entity.csv`    # entities acquired from ConceptNet.
   └── `ROCStories`
       ├── `ant_data`        # sampled stories and corresponding human annotation.
              └── `ant_data.txt`        # include only binary annotation for reasonable(1) or unreasonable(0)
              └── `ant_data_all.txt`    # include the annotation for specific error types: reasonable(0), repeated plots(1), bad coherence(2), conflicting logic(3), chaotic scenes(4), and others(5). 
              └── `reference.txt`       # human-written stories with the same leading context with annotated stories.
              └── `reference_ipt.txt`
              └── `reference_opt.txt`
       ├── `ini_data`        # original dataset for training/validation/testing.
              └── `train.txt`
              └── `dev.txt`
              └── `test.txt`
              └── `entity_vocab.txt`    # generated by ``, consisting of all the entities and the corresponding tagged POS followed by the mention frequency in the dataset.
       ├── `train_data`      # negative samples and corresponding human-written stories for training, which are constructed by ``.
              └── `train_human.txt`
              └── `train_negative.txt`
              └── `dev_human.txt`
              └── `dev_negative.txt`
              └── `test_human.txt`
              └── `test_negative.txt`
       ├── `metric_output`   # the scores of different metrics, which can be used to replicate the correlation in Table 5 of the paper. 
              └── `bleu.txt`
              └── `bleurt.txt`
              └── `ppl.txt`             # the sign of the result of Perplexity needs to be changed to get the result for *minus* Perplexity.
              └── `union.txt`
              └── `union_recon.txt`     # the ablated model without the reconstruction task
              └── ...
   └── `WritingPrompts`
       ├── ...
  • The annotated data file ant_data.txt and ant_data_all.txt are formatted as Story ID ||| Story ||| Seven Annotated Scores.
  • ant_data_all.txt is only available for ROCStories corpus. ant_data_all.txt is the same with ant_data.txt for WrintingPrompts dataset.


Please kindly cite our paper if this paper and the code are helpful.

    title={UNION: An Unreferenced Metric for Evaluating Open-ended Story Generation},
    author={Jian Guan and Minlie Huang},


UNION: An Unreferenced Metric for Evaluating Open-ended Story Generation







No releases published


No packages published