bart-perspectives

Overview

The BART-perspectives model is a sequence-to-sequence transformers mode;. Built on top of Facebook's BART-large (specifically the philschmid/bart-large-cnn-samsum finetune), it is specifically designed to extract perspectives from textual data at scale. The model provides an in-depth analysis of the speaker's identity, their emotions, the object of these emotions, and the reason behind these emotions.

Usage

It is designed to be used with the perspectives library:

from perspectives import DataFrame

# Load DataFrame
df = DataFrame(texts = [list of sentences]) 

# Get perspectives
df.get_perspectives()

# Search
df.search(speaker='...', emotion='...')

You can use also this model directly with a pipeline for text generation:

from transformers import pipeline

# Load the model
generator = pipeline('text-generation', model='helliun/bart-perspectives')

# Get perspective
perspective = generator("Describe the perspective of this text: <your text>", max_length=1024, do_sample=False)
print(perspective)

You can also use it with transformers.AutoTokenizer and transformers.AutoModelForSeq2SeqLM:

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

# Load the model
tokenizer = AutoTokenizer.from_pretrained("helliun/bart-perspectives")
model = AutoModelForSeq2SeqLM.from_pretrained("helliun/bart-perspectives")

# Tokenize the sentence
inputs = tokenizer.encode("Describe the perspective for this sentence: <your text>", return_tensors='pt')

# Pass the tensor through the model
results = model.generate(inputs)

# Decode the results
decoded = tokenizer.decode(results[:,0])
print(decoded)

Training

The model was fine-tuned on a subset of the mteb/tweet-sentiment-extraction dataset with emotional analyses generated synthetically by GPT-4.

About me

I'm a recent grad of Ohio State University where I did an undergraduate thesis on Synthetic Data Augmentation using LLMs. I've worked as an NLP consultant for a couple awesome startups, and now I'm looking for a role with an inspiring company who is as interested in the untapped potential of LMs as I am! Here's my LinkedIn.

Contributing and Support

Please raise an issue here if you encounter any problems using the model. Contributions like fine-tuning on additional data or improving the model architecture are always welcome!

Buy me a coffee!

License

The model is open source and free to use under the MIT license.