T5 for Sentence Split in English:

Sentence Split is task of dividing complex sentence in two simple sentences ex. complex sentence

Mary likes to play football in her freetime whenever she meets with her friends that are very nice people.

can be divided in

Mary likes to play football in her freetime whenever she meets with her friends.

and

Her friends are very nice people.

Goal:

To make best sentence split model available till now

Demo:

Check out the Demo

How to use in your python code:

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
tokenizer = AutoTokenizer.from_pretrained("flax-community/t5-base-wikisplit")
model = AutoModelForSeq2SeqLM.from_pretrained("flax-community/t5-base-wikisplit")

complex_sentence = "This comedy drama is produced by Tidy , the company she co-founded in 2008 with her husband David Peet , who is managing director ."
sample_tokenized = tokenizer(complex_sentence, return_tensors="pt")

answer = model.generate(sample_tokenized['input_ids'], attention_mask = sample_tokenized['attention_mask'], max_length=256, num_beams=5)
gene_sentence = tokenizer.decode(answer[0], skip_special_tokens=True)
gene_sentence

"""
Output:
This comedy drama is produced by Tidy. She co-founded Tidy in 2008 with her husband David Peet, who is managing director.
"""

Application:

• Sentence Simplification
• Data Augmentation
• Sentence Rephrase

Current Basline from paper

Our Results:

Model	Exact	SARI	BLEU
t5-base-wikisplit	17.93	67.5438	76.9
t5-v1_1-base-wikisplit	18.1207	67.4873	76.9478
byt5-base-wikisplit	11.3582	67.2685	73.1682
t5-large-wikisplit	18.6632	68.0501	77.1881

Accomplishment:

• All of our models are having better result for two metrics(Exact and SARI scores) than baseline models
• Our t5-base-wikisplit and t5-v1_1-base-wikisplit model are achieving comparative results with half model size or weights that will enable faster inference
• We added wikisplit metrics which is freely available at huggingface datasets. It will be easy to calculate relevent scores for this task from now on

Contributors:

• Bhadresh Savani
• Rahul Dev

To Do:

• t5-base training on Wiki Split
• t5-v1_1-base training on Wiki Split
• byt5-base training on Wiki Split
• t5-large training on Wiki Split
• Streamlit UI for App
• Single Websplit Evaluation Metrics Addition in Huggingface Datasets
• Challenge: Get better performance than roberta2roberta_L-24_wikisplit
• Performance improvement with Research
• Tackle Gender Biasness and fairness while text generation
• Benchmarking and Experimenting with Web Split