Model Card of lmqg/mt5-base-dequad-qg

This model is fine-tuned version of google/mt5-base for question generation task on the lmqg/qg_dequad (dataset_name: default) via lmqg.

Overview

• Language model: google/mt5-base
• Language: de
• Training data: lmqg/qg_dequad (default)
• Online Demo: https://autoqg.net/
• Repository: https://github.com/asahi417/lm-question-generation
• Paper: https://arxiv.org/abs/2210.03992

Usage

• With lmqg

from lmqg import TransformersQG

# initialize model
model = TransformersQG(language="de", model="lmqg/mt5-base-dequad-qg")

# model prediction
questions = model.generate_q(list_context="das erste weltweit errichtete Hermann Brehmer 1855 im niederschlesischen ''Görbersdorf'' (heute Sokołowsko, Polen).", list_answer="1855")

• With transformers

from transformers import pipeline

pipe = pipeline("text2text-generation", "lmqg/mt5-base-dequad-qg")
output = pipe("Empfangs- und Sendeantenne sollen in ihrer Polarisation übereinstimmen, andernfalls <hl> wird die Signalübertragung stark gedämpft. <hl>")

Evaluation

• Metric (Question Generation): raw metric file

	Score	Type	Dataset
BERTScore	80.39	default	lmqg/qg_dequad
Bleu_1	10.85	default	lmqg/qg_dequad
Bleu_2	4.61	default	lmqg/qg_dequad
Bleu_3	2.06	default	lmqg/qg_dequad
Bleu_4	0.87	default	lmqg/qg_dequad
METEOR	13.65	default	lmqg/qg_dequad
MoverScore	55.73	default	lmqg/qg_dequad
ROUGE_L	11.1	default	lmqg/qg_dequad

• Metric (Question & Answer Generation, Reference Answer): Each question is generated from the gold answer. raw metric file

	Score	Type	Dataset
QAAlignedF1Score (BERTScore)	90.63	default	lmqg/qg_dequad
QAAlignedF1Score (MoverScore)	65.32	default	lmqg/qg_dequad
QAAlignedPrecision (BERTScore)	90.65	default	lmqg/qg_dequad
QAAlignedPrecision (MoverScore)	65.34	default	lmqg/qg_dequad
QAAlignedRecall (BERTScore)	90.61	default	lmqg/qg_dequad
QAAlignedRecall (MoverScore)	65.3	default	lmqg/qg_dequad

• Metric (Question & Answer Generation, Pipeline Approach): Each question is generated on the answer generated by lmqg/mt5-base-dequad-ae. raw metric file

	Score	Type	Dataset
QAAlignedF1Score (BERTScore)	76.86	default	lmqg/qg_dequad
QAAlignedF1Score (MoverScore)	52.96	default	lmqg/qg_dequad
QAAlignedPrecision (BERTScore)	76.28	default	lmqg/qg_dequad
QAAlignedPrecision (MoverScore)	52.93	default	lmqg/qg_dequad
QAAlignedRecall (BERTScore)	77.55	default	lmqg/qg_dequad
QAAlignedRecall (MoverScore)	53.06	default	lmqg/qg_dequad

Training hyperparameters

The following hyperparameters were used during fine-tuning:

• dataset_path: lmqg/qg_dequad
• dataset_name: default
• input_types: ['paragraph_answer']
• output_types: ['question']
• prefix_types: None
• model: google/mt5-base
• max_length: 512
• max_length_output: 32
• epoch: 17
• batch: 4
• lr: 0.0005
• fp16: False
• random_seed: 1
• gradient_accumulation_steps: 16
• label_smoothing: 0.15

The full configuration can be found at fine-tuning config file.

Citation

@inproceedings{ushio-etal-2022-generative,
    title = "{G}enerative {L}anguage {M}odels for {P}aragraph-{L}evel {Q}uestion {G}eneration",
    author = "Ushio, Asahi  and
        Alva-Manchego, Fernando  and
        Camacho-Collados, Jose",
    booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, U.A.E.",
    publisher = "Association for Computational Linguistics",
}