bert_finetuned_qa

Fine-Tuned BERT Models for Thermoelectric Materials Question Answering

Introduction

This repository contains three BERT models fine-tuned for question-answering (QA) tasks related to thermoelectric materials. The models are trained on different datasets to evaluate their performance on specialised QA tasks in the field of materials science.

We present a method for auto-generating a large question-answering dataset about thermoelectric materials for language model applications. The method was used to generate a dataset with sentence-wide contexts from a database of thermoelectric material records. The dataset was contrasted with SQuAD-v2, as well as the mixed combination of the two datasets. Hyperparameter optimisation was employed to fine-tune BERT models on each dataset, and the three best-performing models were then compared on a manually annotated test set of thermoelectric material paragraph contexts with questions spanning material names, five different properties, and temperatures during recording. The best BERT model fine-tuned on the mixed dataset outperforms the other two models when evaluated on the test dataset, indicating that mixing datasets with different semantic and syntactic scopes might be a beneficial approach to improving performance on specialised question-answering tasks.

Models Included

1. squad-v2_best

   Description: Fine-tuned on the SQuAD-v2 dataset, which is a widely used benchmark for QA tasks.
   Dataset: SQuAD-v2
   Location: squad-v2_best/

2. te-cde_best

   Description: Fine-tuned on a thermoelectric materials-specific dataset generated using our auto-generation method.
   Dataset: Thermoelectric Materials QA Dataset (TE-CDE)
   Location: te-cde_best/

3. mixed_best

   Description: Fine-tuned on a mixed dataset combining SQuAD-v2 and the thermoelectric materials dataset to enhance performance on specialised QA tasks.
   Dataset: Combination of SQuAD-v2 and TE-CDE
   Location: mixed_best/

Dataset Details

SQuAD-v2

A reading comprehension dataset consisting of questions posed by crowdworkers on a set of Wikipedia articles.
Some questions are unanswerable, adding complexity to the QA task.

Thermoelectric Materials QA Dataset (TE-CDE)

Auto-generated dataset containing QA pairs about thermoelectric materials.
Contexts are sentence-wide excerpts from a database of thermoelectric material records.
Questions cover:
    Material names
    Five different properties
    Temperatures during recording

Mixed Dataset

A combination of SQuAD-v2 and TE-CDE datasets.
Aims to leverage the strengths of both general-purpose and domain-specific data.

Training Details

Base Model: BERT Base Uncased
Hyperparameter Optimisation: Employed to find the best-performing models for each dataset.
Training Parameters:
    Epochs: Adjusted per dataset based on validation loss.
    Batch Size: Optimized during training.
    Learning Rate: Tuned using grid search.

Evaluation Metrics

Evaluation Dataset: Manually annotated test set of thermoelectric material paragraph contexts.
Metrics Used:
    Exact Match (EM): Measures the percentage of predictions that match any one of the ground truth answers exactly.
    F1 Score: Harmonic mean of precision and recall, considering overlap between the prediction and ground truth answers.

Performance Comparison

Model Exact Match (EM) F1 Score squad-v2_best 57.60% 61.82% te-cde_best 65.39% 69.78% mixed_best 67.92% 72.29%

Usage Instructions

Installing Dependencies

pip install transformers

Loading a Model

Replace model_name with one of the following:

"odysie/bert-finetuned-qa-datasets/squad-v2_best"
"odysie/bert-finetuned-qa-datasets/te-cde_best"
"odysie/bert-finetuned-qa-datasets/mixed_best"

from transformers import BertForQuestionAnswering, BertTokenizer

model_name = "odysie/bert-finetuned-qa-datasets/mixed_best"

tokenizer = BertTokenizer.from_pretrained(model_name)
model = BertForQuestionAnswering.from_pretrained(model_name)

# Example question and context
question = "What is the chemical formula for water?"
context = "Water is a molecule composed of two hydrogen atoms and one oxygen atom, with the chemical formula H2O."

# Tokenize input
inputs = tokenizer.encode_plus(question, context, return_tensors="pt")

# Get model predictions
outputs = model(**inputs)
start_scores = outputs.start_logits
end_scores = outputs.end_logits

# Get the most likely beginning and end of answer with the argmax of the score
start_index = start_scores.argmax()
end_index = end_scores.argmax()

# Convert tokens to answer
tokens = inputs["input_ids"][0][start_index : end_index + 1]
answer = tokenizer.decode(tokens)

print(f"Answer: {answer}")

License

This project is licensed under Apache 2.0

Citation

If you use these models in your research or application, please cite our work:

bibtex

(PENDING)

@article{ ... }

Acknowledgments

We thank the contributors of the SQuAD-v2 dataset and the developers of the Hugging Face Transformers library for providing valuable resources that made this work possible.