metadata

tags:
  - conversational
  - tagalog
  - filipino
language:
  - tl
inference: false

Tagalog DialoGPT

A DialoGPT model fine-tuned on Tagalog conversational data scraped from the web. This model is an output of a research on BERT-based data augmentation for low resource languages. We fine-tuned DialoGPT-medium as our base model.

Latest release: July 25, 2021

As of the moment, the model is only able to respond based on the history of 3 previous utterances before being limited. This is a result of the scarce amount of Tagalog conversations in our dataset.

Dataset and Scripts

The training data used was collected under the following categories:

Food and Drinks
Home and Garden
Style and Fashion
Travel and Leisure
Visas and Immigration
Health and Wellness
Body and Fitness
Small Talk

Pinoy Exchange (PEx) Conversational Dataset to be released soon.

Usage

Here is an example of using Beam Search as the decoding method for our model.

for step in range(2):
    # encode the new user input, add the eos_token and return a tensor in Pytorch
    new_user_input_ids = tokenizer.encode(input(">> User:") + tokenizer.eos_token, return_tensors='pt')

    # append the new user input tokens to the chat history
    bot_input_ids = torch.cat([chat_history_ids, new_user_input_ids], dim=-1) if step > 0 else new_user_input_ids

    # we limit the generation to 512 tokens, each utterance in training had a maximum of 128 tokens
    chat_history_ids = model.generate(
        bot_input_ids, max_length=512,
        pad_token_id=tokenizer.eos_token_id,
        num_beams=5, 
        no_repeat_ngram_size=3
    )
    
    # pretty print last ouput tokens from bot
    print("DialoGPT: {}".format(tokenizer.decode(chat_history_ids[:, bot_input_ids.shape[-1]:][0], skip_special_tokens=True)))