by KomeijiForce
Open source · 661k downloads · 0 likes
The *bart large emojilm* model is a specialized version of BART, trained to translate sentences into a series of relevant emojis. It converts textual expressions into expressive visual combinations, such as turning "I love pizza" into "🍕😍", thereby capturing the emotion or overall meaning of the message. Its core capabilities include contextual understanding and the generation of coherent emojis, even for complex or nuanced phrases. This model is particularly useful for applications requiring immediate visual communication, such as messaging platforms, social media, or content creation tools. What sets it apart is its innovative approach that combines the power of large language models with creative and intuitive output, offering a playful and effective alternative to traditional text-based representation.
This is a BART model pre-trained on the Text2Emoji dataset to translate setences into series of emojis.
For instance, "I love pizza" will be translated into "🍕😍".
An example implementation for translation:
from transformers import BartTokenizer, BartForConditionalGeneration
def translate(sentence, **argv):
inputs = tokenizer(sentence, return_tensors="pt")
generated_ids = generator.generate(inputs["input_ids"], **argv)
decoded = tokenizer.decode(generated_ids[0], skip_special_tokens=True).replace(" ", "")
return decoded
path = "KomeijiForce/bart-large-emojilm"
tokenizer = BartTokenizer.from_pretrained(path)
generator = BartForConditionalGeneration.from_pretrained(path)
sentence = "I love the weather in Alaska!"
decoded = translate(sentence, num_beams=4, do_sample=True, max_length=100)
print(decoded)
You will probably get some output like "❄️🏔️😍".
If you find this model & dataset resource useful, please consider cite our paper:
@article{DBLP:journals/corr/abs-2311-01751,
author = {Letian Peng and
Zilong Wang and
Hang Liu and
Zihan Wang and
Jingbo Shang},
title = {EmojiLM: Modeling the New Emoji Language},
journal = {CoRR},
volume = {abs/2311.01751},
year = {2023},
url = {https://doi.org/10.48550/arXiv.2311.01751},
doi = {10.48550/ARXIV.2311.01751},
eprinttype = {arXiv},
eprint = {2311.01751},
timestamp = {Tue, 07 Nov 2023 18:17:14 +0100},
biburl = {https://dblp.org/rec/journals/corr/abs-2311-01751.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}