chaii - hindi & tamil question answering

Last update: Dec 18, 2022

Overview

chaii - hindi & tamil question answering

This is the solution for rank 5th in Kaggle competition: chaii - Hindi and Tamil Question Answering. The competition can be found here: https://www.kaggle.com/c/chaii-hindi-and-tamil-question-answering

Datasets required

Download squadv2 data from https://rajpurkar.github.io/SQuAD-explorer/

$ mkdir input && cd input
$ wget https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v2.0.json
$ wget https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v2.0.json

Download tydiqa data in the input folder:

$ wget https://storage.googleapis.com/tydiqa/v1.1/tydiqa-goldp-v1.1-train.json
$ wget https://storage.googleapis.com/tydiqa/v1.1/tydiqa-goldp-v1.1-dev.json

Download data from https://www.kaggle.com/tkm2261/google-translated-squad20-to-hindi-and-tamil to input folder

Download original competition dataset to input folder: https://www.kaggle.com/c/chaii-hindi-and-tamil-question-answering/data

Download outputs of this kernel: https://www.kaggle.com/rhtsingh/external-data-mlqa-xquad-preprocessing/ to input folder

Now, you have all the data needed to train the model. We will first create folds and munge the data a bit.

To create folds, please use the following command:

$ cd src
$ python create_folds.py

To munge the datasets and prepare for training, please run the following command:

$ cd src
$ python munge_data.py

Training

There are two GPU models and one model needs TPUs.

GPU models: XLM-Roberta & Rembert TPU model: Muril-Large

XLM-Roberta:

$ cd src
$ TOKENIZERS_PARALLELISM=false python xlm_roberta.py --fold 0
$ TOKENIZERS_PARALLELISM=false python xlm_roberta.py --fold 1
$ TOKENIZERS_PARALLELISM=false python xlm_roberta.py --fold 2
$ TOKENIZERS_PARALLELISM=false python xlm_roberta.py --fold 3
$ TOKENIZERS_PARALLELISM=false python xlm_roberta.py --fold 4

Rembert:

$ cd src
$ TOKENIZERS_PARALLELISM=false python rembert.py --fold 0
$ TOKENIZERS_PARALLELISM=false python rembert.py --fold 1
$ TOKENIZERS_PARALLELISM=false python rembert.py --fold 2
$ TOKENIZERS_PARALLELISM=false python rembert.py --fold 3
$ TOKENIZERS_PARALLELISM=false python rembert.py --fold 4

Muril-Large

** please note that training this model needs TPUs **

$ cd src
$ TOKENIZERS_PARALLELISM=false python muril_large.py --fold 0
$ TOKENIZERS_PARALLELISM=false python muril_large.py --fold 1
$ TOKENIZERS_PARALLELISM=false python muril_large.py --fold 2
$ TOKENIZERS_PARALLELISM=false python muril_large.py --fold 3
$ TOKENIZERS_PARALLELISM=false python muril_large.py --fold 4

Inference

After training all the models, the outputs were pushed to Kaggle Datasets.

The final model datasets can be found here:

- https://www.kaggle.com/abhishek/xlmrobertalargewithsquadv2tydiqasqdtrans384f
- https://www.kaggle.com/ubamba98/modelsrembertwithsquadv2tydiqa384
- https://www.kaggle.com/ubamba98/murillargecasedchaii

And the final inference kernel can be found here: https://www.kaggle.com/abhishek/chaii-xlm-roberta-x-muril-x-rembert-score-based

Solution writeup: https://www.kaggle.com/c/chaii-hindi-and-tamil-question-answering/discussion/288049

chaii - hindi & tamil question answering

Related tags

Overview

chaii - hindi & tamil question answering

Datasets required

Training

XLM-Roberta:

Rembert:

Muril-Large

Inference

Owner

abhishek thakur

The official implementation of "BERT is to NLP what AlexNet is to CV: Can Pre-Trained Language Models Identify Analogies?, ACL 2021 main conference"

Nystromformer: A Nystrom-based Algorithm for Approximating Self-Attention

Simple Text-Generator with OpenAI gpt-2 Pytorch Implementation

Rich Prosody Diversity Modelling with Phone-level Mixture Density Network

Guide: Finetune GPT2-XL (1.5 Billion Parameters) and GPT-NEO (2.7 B) on a single 16 GB VRAM V100 Google Cloud instance with Huggingface Transformers using DeepSpeed

Every Google, Azure & IBM text to speech voice for free

Chinese segmentation library

leaking paid token generator that was a shit lmao for 100$ haha

Russian words synonyms and antonyms

This repo contains simple to use, pretrained/training-less models for speaker diarization.

easySpeech is an open-source Python wrapper for google speech to text API that doesn't require PyAudio(So you especially windows user don't have to deal with the errors while installing PyAudio) and also works with hugging face transformers

Official codebase for Can Wikipedia Help Offline Reinforcement Learning?

NVDA, the free and open source Screen Reader for Microsoft Windows

Th2En & Th2Zh: The large-scale datasets for Thai text cross-lingual summarization

PyTorch impelementations of BERT-based Spelling Error Correction Models.

A tool helps build a talk preview image by combining the given background image and talk event description

SGMC: Spectral Graph Matrix Completion

An open collection of annotated voices in Japanese language

EdiTTS: Score-based Editing for Controllable Text-to-Speech

To classify the News into Real/Fake using Features from the Text Content of the article