Skip to content

Latest commit

 

History

History
137 lines (86 loc) · 5.38 KB

README.md

File metadata and controls

137 lines (86 loc) · 5.38 KB

HR-MultiWOZ: A Task Oriented Dialogue (TOD) Dataset for HR LLM Agent

This repository hosts the data generation recipe and benchmarking of HR-MultiWOZ: A Task Oriented Dialogue (TOD) Dataset for HR LLM Agent. This paper is accepted by EACL NLP4HR workshop as presentations.

Introduction

HR-Multiwoz is a fully-labeled dataset of 550 conversations spanning 10 HR domains to evaluate LLM Agent. It is the first labeled open-sourced conversation dataset in the HR domain for NLP research. In this repo we provides a detailed recipe for the data generation procedure described in the paper along with data analysis and human evaluations. The data generation pipeline is transferable and can be easily adapted for labeled conversation data generation in other domains. The proposed data-collection pipeline is mostly based on LLMs with minimal human involvement for annotation, which is time and cost-efficient.

Requirements

Install all required python dependencies:

pip install -r requirements.txt

Citation

@misc{xu2024hrmultiwoz,
      title={HR-MultiWOZ: A Task Oriented Dialogue (TOD) Dataset for HR LLM Agent}, 
      author={Weijie Xu and Zicheng Huang and Wenxiang Hu and Xi Fang and Rajesh Kumar Cherukuri and Naumaan Nayyar and Lorenzo Malandri and Srinivasan H. Sengamedu},
      year={2024},
      eprint={2402.01018},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Data

!Please note that the dataset is an independent reimplementation of our methods. It is not release by Amazon.

All dataset is under Data folder.

sgd_dataset.pkl: This is the schema guided dataset. https://huggingface.co/datasets/xwjzds/hr_multiwoz_tod_sgd

template.pkl: This is the place where you can access the question for each entity.

qa_dataset.pkl: This is the place where you can get question, context and extractive answers. https://huggingface.co/datasets/xwjzds/extractive_qa_question_answering_hr

Flow

The workflow of the paper

🏆 TOD Benchmarks

JAG slot accuracy Method
18.89 55.61 TransferQA with Deberta[11]
8.65 26.62 TransferQA with Bert[11]

Since most of multiwoz dataset is from other domain and is hard to do transfer learning, we can only implement our own baseline. The baseline is inspired by [11] and we select 2 best performing models from below as our base language model. Check leadership/tod_benchmark.py for our implementation.

As you can see, the performance is still really bad. This means that existed SGD method may not be able to transfer to do transfer learning in our use case.

You can follow leaderboard/tod_benchmark.py to benchmark you own methods.

You can also benchmark any question answering model from huggingface using the following code:

python3 leaderboard/qa_benchmark_evaluation.py --name [adding model name] --threshold [adding threshold between 0 and 1 to define the confidence]

🏆 Extractive QA Benchmarks

F1 Exact Match BLEU Method
0.786 0.598 0.174 bert-large-fintuned-squad [10]
0.721 0.519 0.168 distilbert-base [9]
0.710 0.000 0.217 deberta-v3-large [7]
0.642 0.000 0.140 roberta-base-squad2 [3]
0.588 0.000 0.134 mdeberta-v3-base [8]
0.045 0.000 0.011 bert-base-uncased [1]
0.047 0.000 0.010 distilbert-base-uncased [2]
0.050 0.000 0.011 albert [4]
0.050 0.001 0.011 electra-small-discriminator [5]
0.072 0.000 0.020 xlnet-base [6]

bert-large-uncased is finetuned through whole word masking on Squad dataset. This method achieves the best performance. https://huggingface.co/google-bert/bert-large-uncased-whole-word-masking-finetuned-squad

The code to benchmark your extractive QA method

import pickle
import pandas as pd
from leaderboard.metric import calculate_f1_score, calculate_exact_match, calculate_bleu, calculate_rouge, calculate_meteor


#loading qa_dataset
with open('qa_dataset.pkl', 'rb') as f:
    data = pickle.load(f)
#putting your predicted answer in data
data['predicted_answer'] = ...

#benchmark
F1 = calculate_f1_score(evaluations['answer'], evaluations[method])
Exact_Match = calculate_exact_match(evaluations['answer'].tolist(), evaluations[method].tolist())
Bleu = calculate_bleu(evaluations['answer'], evaluations[method])

Reference

[1] https://arxiv.org/abs/1810.04805

[2] https://arxiv.org/abs/1910.01108

[3] https://huggingface.co/deepset/roberta-base-squad2

[4] https://arxiv.org/abs/1909.11942

[5] https://openreview.net/pdf?id=r1xMH1BtvB

[6] https://proceedings.neurips.cc/paper_files/paper/2019/file/dc6a7e655d7e5840e66733e9ee67cc69-Paper.pdf

[7] https://arxiv.org/abs/2111.09543

[8] https://arxiv.org/abs/2111.09543

[9] https://arxiv.org/abs/1910.01108

[10] https://arxiv.org/abs/1810.04805

[11] https://arxiv.org/pdf/2109.04655.pdf

Security

See CONTRIBUTING for more information.

License

This project is licensed under the Apache-2.0 License.