UPB SemEval 2020 Task 6

This repository contains the source code for the first and the second task of DeftEval 2020 competition, used by the University Politehnica of Bucharest (UPB) team to train and evaluate the models.

We fined-tuned frozen and non-frozen Transformer-based language models using the HuggingFace framework, together with a multi-task model that jointly predicts the outputs for the second and the third task.

The code for each task, with additional details on how to use it, can be found in task1 and task2 directories.

Installation

Make sure you have Python3 and PyTorch installed.

pip install -r requirements.txt

Task1 Results

Model	Valid-Prec	Valid-Rec	Valid-F1c	Test-Prec	Test-Rec	Test-F1
Frozen-BERT	76.0	75.1	75.5	-	-	-
Frozen RoBERTa	74.1	74.4	74.3	-	-	-
Frozen SciBERT	71.7	80.6	75.9	-	-	-
Frozen XLNet	66.8	70.8	68.7	-	-	-
Frozen ALBERT	77.2	69.3	73.0	-	-	-
Fine-tuned BERT	78.4	84.1	81.2	-	-	-
Fine-tuned RoBERTa	78.2	84.5	81.3	75.0	80.6	77.7
Fine-tuned SciBERT	79.4	79.7	79.6	-	-	-
Fine-tuned XLNet	75.5	85.2	80.1	-	-	-
Fine-tuned ALBERT	69.5	85.9	76.8	-	-	-

Task2 Results

Model	Valid-Prec	Valid-Rec	Valid-F1c	Test-Prec	Test-Rec	Test-F1
Frozen BERT+CRF	27.1	39.8	26.1	-	-	-
Frozen RoBERTa+CRF	32.7	27.7	22.6	-	-	-
Frozen SciBERT+CRF	29.0	37.5	26.2	-	-	-
Frozen XLNet+CRF	29.6	33.4	26.8	-	-	-
Frozen Multi-task	4.0	8.9	8.0	-	-	-
Fine-tuned BERT+CRF	47.9	51.7	45.6	-	-	-
Fine-tuned RoBERTa+CRF	41.4	66.4	46.0	39.4	55.6	43.9
Fine-tuned SciBERT+CRF	46.7	46.6	41.7	-	-	-
Fine-tuned XLNet+CRF	33.0	58.5	39.2	-	-	-
Fine-tuned Multi-task	25.7	25.2	25.5	-	-	-

Cite

If you are using this repository, please consider citing the following paper as a thank you to the authors:

@inproceedings{avram-etal-2020-upb,
    title = "{UPB} at {S}em{E}val-2020 Task 6: Pretrained Language Models for Definition Extraction",
    author = "Avram, Andrei-Marius  and
      Cercel, Dumitru-Clementin  and
      Chiru, Costin",
    booktitle = "Proceedings of the Fourteenth Workshop on Semantic Evaluation",
    month = dec,
    year = "2020",
    address = "Barcelona (online)",
    publisher = "International Committee for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.semeval-1.97",
    pages = "737--745",
    abstract = "This work presents our contribution in the context of the 6th task of SemEval-2020: Extracting Definitions from Free Text in Textbooks (DeftEval). This competition consists of three subtasks with different levels of granularity: (1) classification of sentences as definitional or non-definitional, (2) labeling of definitional sentences, and (3) relation classification. We use various pretrained language models (i.e., BERT, XLNet, RoBERTa, SciBERT, and ALBERT) to solve each of the three subtasks of the competition. Specifically, for each language model variant, we experiment by both freezing its weights and fine-tuning them. We also explore a multi-task architecture that was trained to jointly predict the outputs for the second and the third subtasks. Our best performing model evaluated on the DeftEval dataset obtains the 32nd place for the first subtask and the 37th place for the second subtask. The code is available for further research at: \url{https://github.com/avramandrei/DeftEval}",
}

avramandrei / upb-semeval-2020-task-6 Goto Github PK