File size: 5,777 Bytes
2a01a2c e4df92c 9f6021e 2a01a2c f6e04ed e4df92c 9f6021e e4df92c 9f6021e e4df92c 9f6021e e4df92c 9f6021e e4df92c 4e21a64 e4df92c dcaefec e4df92c 9f6021e e4df92c dcaefec e4df92c 9f6021e e4df92c 9f6021e |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 |
---
datasets:
- roemmele/ablit
language:
- en
license: mit
---
# roemmele/ablit-bart-base
<!-- Provide a quick summary of what the model is/does. -->
This model is initialized from facebook/bart-base. It has been fine-tuned on the AbLit dataset, which consists of abridged versions of books aligned with their original versions at the passage level. Given a text, the model generates an abridgement of the text based on what it has observed in AbLit. See the paper cited below for more details.
## Model Details
### Model Description
<!-- Provide a longer summary of what this model is. -->
- **Developed by:** Language Weaver (Melissa Roemmele, Kyle Shaffer, Katrina Olsen, Yiyi Wang, and Steve DeNeefe)
- **Model type:** Seq2SeqLM
- **Language(s) (NLP):** English
- **License:** mit
- **Finetuned from model:** [facebook/bart-base](https://huggingface.co/facebook/bart-base)
### Model Sources
<!-- Provide the basic links for the model. -->
- **Repository:** [github.com/roemmele/AbLit](https://github.com/roemmele/AbLit)
- **Paper:** [AbLit: A Resource for Analyzing and Generating Abridged Versions of English Literature](https://arxiv.org/pdf/2302.06579.pdf)
## Uses
This model generates abridged versions of texts informed by the AbLit dataset.
## Bias, Risks, and Limitations
<!-- This section is meant to convey both technical and sociotechnical limitations. -->
This model comes from research on abridgement as an NLP task, but the dataset the model is trained on (AbLit) is derived from a small set of texts associated with a specific domain and author. In particular, AbLit consists of British English literature from the 18th and 19th centuries, abridged by a single author. Some of the linguistic properties of these original books do not generalize to other domains of English text, and therefore the model might not produce desirable abridgements for other texts.
## How to Get Started with the Model
```
In [1]: from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
...: tokenizer = AutoTokenizer.from_pretrained("roemmele/ablit-bart-base")
...: model = AutoModelForSeq2SeqLM.from_pretrained("roemmele/ablit-bart-base")
...:
...: passage = "The letter was not unproductive. It re-established peace and kindness."
...: input_ids = tokenizer(
...: passage,
...: padding='max_length',
...: return_tensors="pt").input_ids
...: output_ids = model.generate(
...: input_ids,
...: max_length=1024,
...: num_beams=5,
...: no_repeat_ngram_size=3
...: )[0]
...: abridgement = tokenizer.decode(
...: output_ids,
...: skip_special_tokens=True)
In [2]: print(abridgement)
The letter re-established peace and kindness.
```
## Training Details
### Training Data
The model is trained on [roemmele/AbLit](https://huggingface.co/datasets/roemmele/ablit), specifically the train split of the "chunks-10-sentences" subset, i.e.:
```
from datasets import load_dataset
data = load_dataset("roemmele/ablit", "chunks-10-sentences")
```
### Training Procedure
We used the training script [here](https://github.com/huggingface/transformers/blob/main/examples/pytorch/summarization/run_summarization.py).
#### Training Hyperparameters
We specified maximum length of 1024 for both the source (original passage) and target (abridged passage), and truncated all tokens beyond this limit. We evaluated each model on the AbLit development set after each epoch and concluded training when cross-entropy loss stopped decreasing. We used a batch size of 4. For all other hyperparameters we used the default values set by this script.
#### Speeds, Sizes, Times
<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
It took ≈3 hours to train each model on a g4dn.4xlarge AWS instance.
## Evaluation
### Testing Data
<!-- This should link to a Data Card if possible. -->
We evaluated on the test split of the "chunks-10-sentences" subset of [roemmele/AbLit](https://huggingface.co/datasets/roemmele/ablit)
### Results
<!-- This section describes the evaluation protocols and provides the results. -->
The model obtained a ROUGE-L score of 0.78 on the AbLit test set. See the paper for the results of other metrics.
### Conclusion
Our analysis shows that in comparison with human-authored abridgements, the model-generated abridgements tend to preserve more of the original text, suggesting it is challenging to learn what text can be removed while maintaining loyalty to the important parts of the original text.
## Citation
<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
**BibTeX:**
@inproceedings{roemmele2023ablit,
title={AbLit: A Resource for Analyzing and Generating Abridged Versions of English Literature},
author={Roemmele, Melissa and Shaffer, Kyle and Olsen, Katrina and Wang, Yiyi and DeNeefe, Steve},
booktitle = {Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume},
publisher = {Association for Computational Linguistics},
year={2023}
}
**APA:**
Roemmele, M., Shaffer, K., Olsen, K., Wang, Y., and DeNeefe, S. (2023). AbLit: A Resource for Analyzing and Generating Abridged Versions of English Literature. 17th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2023).
## Model Card Authors
Melissa Roemmele
## Model Card Contact
melissa@roemmele.io |