abisee / story-generation-eval Goto Github PK

This is the code for the CoNLL 2019 paper Do Massively Pretrained Language Models Make Better Storytellers?.

This repo contains code to download the generated text data, compute the automatic metrics on it, and produce the plots in the paper.

Download and unzip data.zip (1.1GB) to make a directory called data:

cd story-generation-eval
wget https://nlp.stanford.edu/data/gpt_analysis/data.zip
unzip data.zip

In data/stories_unannotated, each json file contains a dictionary mapping from the sample id (an integer) to a story (which can be generated by Fusion Model/GPT2, or a human sample from WritingPrompts dataset).

Each story is represented by a dictionary containing the following:

• prompt_text, story_text: string. The text of the prompt/story.
• prompt_tokens, story_tokens: list of strings. The model-level tokens in the prompt/story. This is word-level for Fusion and subword-level for GPT2.
• gen_probs_orig: list of floats, same length as story_tokens. The model's probability for each of the generated tokens. Not included for the human-written stories.
• vocab_entropy_orig: list of floats, same length as story_tokens. The entropy of the model's vocabulary distribution for each step. Not included for the human-written stories.

If you want to generate the unannotated story files yourself:

• The code we used to generate text from the Fusion Model can be found in fairseq.
• The code we used to finetune and generate text from GPT2-117 can be found in Huggingface transformers.

In data/teacherforcing are two files, one for the Fusion Model and one for GPT2-117. These are the result of running the models on the WritingPrompts test data using teacher forcing (i.e. not generating) in order to measure the per-step probabilities and vocabulary entropies.

In data/stories_metric_annotated, each pickle file corresponds to a file in data/stories_unannotated. It contains a dictionary mapping from the sample id (an integer) to an annotated story.

Each annotated story is represented by a Sample (see samples.py for the class) which contains the information from stories_unannotated file, plus the metric annotations.

If you want to run the annotation code yourself, here is a guide to explain how data/stories_metric_annotated was produced:

Several of our automatic metrics use Spacy annotations, so we precompute those before computing our automatic metrics. Run

python spacy_annotate.py

This will take each story in data/stories_unannotated and save a Spacy encoding of its prompt and story in data/stories_spacy_annotated.

The Spacy annotation can take a long time. If you want to download the precomputed ones for our generated stories, you should download stories_spacy_annotated.zip (14 GB). Unzip it into the data dir to create a new directory called stories_spacy_annotated:

cd data
wget https://nlp.stanford.edu/data/gpt_analysis/stories_spacy_annotated.zip
unzip stories_spacy_annotated.zip

Note that our annotation script additionally uses these three files:

• data/arora_sentence_embedder.pkl: This is needed to compute the Arora sentence embeddings for the story-prompt similarity metric. To see how it was produced, follow the instructions in make_arora_sent_embedder.py.
• data/unigram_probdist.json: This is needed to compute the mean log unigram probability metric. It was computed by iterating through the WritingPrompts training set and counting the frequency of each unigram.
• data/concreteness.csv: This is needed to compute the mean noun and verb concreteness metrics. It is from the paper Concreteness ratings for 40 thousand generally known English word lemmas.

Run

python metrics_annotate.py

This will take each story in data/stories_unannotated, load its Spacy encoding in data/stories_spacy_annotated, and save a metric-annotated version in data/stories_metric_annotated.

Note that annotating the first file takes the longest because there are various packages and data that need to be loaded on the first turn.

Run

jupyter notebook

to open analyze.ipynb. This notebook loads the data in data/stories_metric_annotated and data/teacherforcing, and produces the plots in the paper. The notebook also allows you to browse the generated stories. You can see an in-browser preview here.

Our code to compute prompt ranking accuracy on the Fusion Model is visible in this pull request. The prompt ranking accuracy code for GPT2-117 is similar.

If you use this code, please cite our paper:

@inproceedings{see2019massively,
  title={Do Massively Pretrained Language Models Make Better Storytellers?},
  author={See, Abigail and Pappu, Aneesh and Saxena, Rohun and Yerukola, Akhila and Manning, Christopher D},
  booktitle={Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL)},
  pages={843--861},
  year={2019}
}