Model Card of lmqg/mbart-large-cc25-frquad-qg-ae

This model is fine-tuned version of facebook/mbart-large-cc25 for question generation and answer extraction jointly on the lmqg/qg_frquad (dataset_name: default) via lmqg.

Overview

Usage

from lmqg import TransformersQG

# initialize model
model = TransformersQG(language="fr", model="lmqg/mbart-large-cc25-frquad-qg-ae")

# model prediction
question_answer_pairs = model.generate_qa("Créateur » (Maker), lui aussi au singulier, « le Suprême Berger » (The Great Shepherd) ; de l'autre, des réminiscences de la théologie de l'Antiquité : le tonnerre, voix de Jupiter, « Et souvent ta voix gronde en un tonnerre terrifiant », etc.")
  • With transformers
from transformers import pipeline

pipe = pipeline("text2text-generation", "lmqg/mbart-large-cc25-frquad-qg-ae")

# answer extraction
answer = pipe("generate question: Créateur » (Maker), lui aussi au singulier, « <hl> le Suprême Berger <hl> » (The Great Shepherd) ; de l'autre, des réminiscences de la théologie de l'Antiquité : le tonnerre, voix de Jupiter, « Et souvent ta voix gronde en un tonnerre terrifiant », etc.")

# question generation
question = pipe("extract answers: Pourtant, la strophe spensérienne, utilisée cinq fois avant que ne commence le chœur, constitue en soi un vecteur dont les répétitions structurelles, selon Ricks, relèvent du pur lyrisme tout en constituant une menace potentielle. Après les huit sages pentamètres iambiques, l'alexandrin final <hl> permet une pause <hl>, « véritable illusion d'optique » qu'accentuent les nombreuses expressions archaïsantes telles que did swoon, did seem, did go, did receive, did make, qui doublent le prétérit en un temps composé et paraissent à la fois « très précautionneuses et très peu pressées ».")

Evaluation

Score Type Dataset
BERTScore 72.56 default lmqg/qg_frquad
Bleu_1 16.16 default lmqg/qg_frquad
Bleu_2 4.88 default lmqg/qg_frquad
Bleu_3 1.85 default lmqg/qg_frquad
Bleu_4 0.91 default lmqg/qg_frquad
METEOR 8.56 default lmqg/qg_frquad
MoverScore 50.46 default lmqg/qg_frquad
ROUGE_L 18.54 default lmqg/qg_frquad
Score Type Dataset
QAAlignedF1Score (BERTScore) 77.72 default lmqg/qg_frquad
QAAlignedF1Score (MoverScore) 51.65 default lmqg/qg_frquad
QAAlignedPrecision (BERTScore) 76.9 default lmqg/qg_frquad
QAAlignedPrecision (MoverScore) 51.15 default lmqg/qg_frquad
QAAlignedRecall (BERTScore) 78.58 default lmqg/qg_frquad
QAAlignedRecall (MoverScore) 52.16 default lmqg/qg_frquad
Score Type Dataset
AnswerExactMatch 0 default lmqg/qg_frquad
AnswerF1Score 3.66 default lmqg/qg_frquad
BERTScore 58.41 default lmqg/qg_frquad
Bleu_1 2.56 default lmqg/qg_frquad
Bleu_2 0.76 default lmqg/qg_frquad
Bleu_3 0 default lmqg/qg_frquad
Bleu_4 0 default lmqg/qg_frquad
METEOR 3.24 default lmqg/qg_frquad
MoverScore 45.72 default lmqg/qg_frquad
ROUGE_L 3.48 default lmqg/qg_frquad

Training hyperparameters

The following hyperparameters were used during fine-tuning:

  • dataset_path: lmqg/qg_frquad
  • dataset_name: default
  • input_types: ['paragraph_answer', 'paragraph_sentence']
  • output_types: ['question', 'answer']
  • prefix_types: ['qg', 'ae']
  • model: facebook/mbart-large-cc25
  • max_length: 512
  • max_length_output: 32
  • epoch: 5
  • batch: 2
  • lr: 0.0001
  • fp16: False
  • random_seed: 1
  • gradient_accumulation_steps: 32
  • 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",
}
Downloads last month
7
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Evaluation results

  • BLEU4 (Question Generation) on lmqg/qg_frquad
    self-reported
    0.910
  • ROUGE-L (Question Generation) on lmqg/qg_frquad
    self-reported
    18.540
  • METEOR (Question Generation) on lmqg/qg_frquad
    self-reported
    8.560
  • BERTScore (Question Generation) on lmqg/qg_frquad
    self-reported
    72.560
  • MoverScore (Question Generation) on lmqg/qg_frquad
    self-reported
    50.460
  • QAAlignedF1Score-BERTScore (Question & Answer Generation (with Gold Answer)) on lmqg/qg_frquad
    self-reported
    77.720
  • QAAlignedRecall-BERTScore (Question & Answer Generation (with Gold Answer)) on lmqg/qg_frquad
    self-reported
    78.580
  • QAAlignedPrecision-BERTScore (Question & Answer Generation (with Gold Answer)) on lmqg/qg_frquad
    self-reported
    76.900
  • QAAlignedF1Score-MoverScore (Question & Answer Generation (with Gold Answer)) on lmqg/qg_frquad
    self-reported
    51.650
  • QAAlignedRecall-MoverScore (Question & Answer Generation (with Gold Answer)) on lmqg/qg_frquad
    self-reported
    52.160