grammar-synthesis-large - beta

A fine-tuned version of google/t5-v1_1-large for grammar correction on an expanded version of the JFLEG dataset.

usage in Python (after pip install transformers):

from transformers import pipeline

corrector = pipeline(
              'text2text-generation',
              'pszemraj/grammar-synthesis-large',
              )
raw_text = 'i can has cheezburger'
results = corrector(raw_text)
print(results)

give it a spin in Colab at this notebook

Model description

The intent is to create a text2text language model that successfully completes "single-shot grammar correction" on a potentially grammatically incorrect text that could have a lot of mistakes with the important qualifier of it does not semantically change text/information that IS grammatically correct.

Compare some of the heavier-error examples on other grammar correction models to see the difference :)

Other checkpoints

If trading a slight decrease in grammatical correction quality for faster inference speed makes sense for your use case, check out the base and small checkpoints fine-tuned from the relevant t5 checkpoints.

Limitations

• dataset: cc-by-nc-sa-4.0
• model: apache-2.0
• this is still a work-in-progress and while probably useful for "single-shot grammar correction" in a lot of cases, give the outputs a glance for correctness ok?

Use Cases

Obviously, this section is quite general as there are many things one can use "general single-shot grammar correction" for. Some ideas or use cases:

1. Correcting highly error-prone LM outputs. Some examples would be audio transcription (ASR) (this is literally some of the examples) or something like handwriting OCR.
   • To be investigated further, depending on what model/system is used it might be worth it to apply this after OCR on typed characters.
2. Correcting/infilling text generated by text generation models to be cohesive/remove obvious errors that break the conversation immersion. I use this on the outputs of this OPT 2.7B chatbot-esque model of myself.

   An example of this model running on CPU with beam search:

original response:
                ive heard it attributed to a bunch of different philosophical schools, including stoicism, pragmatism, existentialism and even some forms of post-structuralism. i think one of the most interesting (and most difficult) philosophical problems is trying to let dogs (or other animals) out of cages. the reason why this is a difficult problem is because it seems to go against our grain (so to
synthesizing took 306.12 seconds
Final response in 1294.857 s:
        I've heard it attributed to a bunch of different philosophical schools, including solipsism, pragmatism, existentialism and even some forms of post-structuralism. i think one of the most interesting (and most difficult) philosophical problems is trying to let dogs (or other animals) out of cages. the reason why this is a difficult problem is because it seems to go against our grain (so to speak)

Note: that I have some other logic that removes any periods at the end of the final sentence in this chatbot setting to avoid coming off as passive aggressive

3. Somewhat related to #2 above, fixing/correcting so-called tortured-phrases that are dead giveaways text was generated by a language model. Note that SOME of these are not fixed, especially as they venture into domain-specific terminology (i.e. irregular timberland instead of Random Forest).

Training and evaluation data

More information needed 😉

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 8e-05
• train_batch_size: 4
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 32
• total_train_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.02
• num_epochs: 1

Framework versions

• Transformers 4.20.1
• Pytorch 1.11.0+cu113
• Datasets 2.3.2
• Tokenizers 0.12.1