license: mit
tags:
- generated_from_keras_callback
model-index:
- name: bart-large-finetuned-filtered-spotify-podcast-summ
results: []
bart-large-finetuned-filtered-spotify-podcast-summ
This model is a fine-tuned version of facebook/bart-large-cnn on on the Spotify Podcast Dataset. Take a look to the github repository of this project.
It achieves the following results on the evaluation set:
- Train Loss: 2.2967
- Validation Loss: 2.8316
- Epoch: 2
Intended uses & limitations
This model is intended to be used for automatic podcast summarisation. Given the podcast transcript in input, the objective is to provide a short text summary that a user might read when deciding whether to listen to a podcast. The summary should accurately convey the content of the podcast, be human-readable, and be short enough to be quickly read on a smartphone screen.
Training and evaluation data
In our solution, an extractive module is developed to select salient chunks from the transcript, which serve as the input to an abstractive summarizer. An extensive pre-processing on the creator-provided descriptions is performed selecting a subset of the corpus that is suitable for the training supervised model. We split the filtered dataset into train/dev sets of 69,336/7,705 episodes. The test set consists of 1,027 episodes. Only 1025 have been used because two of them did not contain an episode description.
How to use
The model can be used for the summarization as follows:
from transformers import pipeline
summarizer = pipeline("summarization", model="gmurro/bart-large-finetuned-filtered-spotify-podcast-summ", tokenizer="gmurro/bart-large-finetuned-filtered-spotify-podcast-summ")
summary = summarizer(podcast_transcript, min_length=39, max_length=250)
print(summary[0]['summary_text'])
Training hyperparameters
The following hyperparameters were used during training:
- optimizer: {'name': 'AdamWeightDecay', 'learning_rate': 2e-05, 'decay': 0.0, 'beta_1': 0.9, 'beta_2': 0.999, 'epsilon': 1e-07, 'amsgrad': False, 'weight_decay_rate': 0.01}
- training_precision: float32
Training results
| Train Loss | Validation Loss | Epoch |
|---|---|---|
| 3.0440 | 2.8733 | 0 |
| 2.6085 | 2.8549 | 1 |
| 2.2967 | 2.8316 | 2 |
Framework versions
- Transformers 4.19.4
- TensorFlow 2.9.1
- Datasets 2.3.1
- Tokenizers 0.12.1
Authors
| Name | Surname | Username | |
|---|---|---|---|
| Giuseppe | Boezio | [email protected] |
giuseppeboezio |
| Simone | Montali | [email protected] |
montali |
| Giuseppe | Murro | [email protected] |
gmurro |