metadata
license: cc-by-nc-sa-4.0
language:
- en
tags:
- argument mining
datasets:
- US2016
- QT30
metrics:
- macro-f1
ALBERT-based model for Argument Relation Identification (ARI)
Argument Mining model trained with English (EN) data for the Argument Relation Identification (ARI) task using the US2016 and the QT30 corpora. This a fine-tuned albert/albert-base-v2 model, inspired by "Transformer-Based Models for Automatic Detection of Argument Relations: A Cross-Domain Evaluation" paper.
This model was trained on the full dataset: train and test merged.
Usage
from transformers import BertTokenizer,BertForSequenceClassification
classes_decoder = {
0: "Inference",
1: "Conflict",
2: "Rephrase",
3: "No-Relation"
}
model = BertForSequenceClassification.from_pretrained("yevhenkost/ArgumentMining-EN-ARI-AIF-ALBERT")
tokenizer = BertTokenizer.from_pretrained("yevhenkost/ArgumentMining-EN-ARI-AIF-ALBERT")
text_one, text_two = "The water is wet", "The sun is really hot"
model_inputs = tokenizer(text_one, text_two, return_tensors="pt")
# regular SequenceClassifierOutput
model_output = model(**model_inputs)
Metrics
precision recall f1-score support
0 0.51 0.59 0.55 833
1 0.46 0.28 0.35 200
2 0.51 0.30 0.38 156
3 0.82 0.82 0.82 2209
accuracy 0.71 3398
macro avg 0.58 0.50 0.53 3398
weighted avg 0.71 0.71 0.71 3398
Theses results for the model that was trained only on train chunk of data and tested on the test one.
Cite:
@article{ruiz2021transformer,
author = {R. Ruiz-Dolz and J. Alemany and S. Barbera and A. Garcia-Fornes},
journal = {IEEE Intelligent Systems},
title = {Transformer-Based Models for Automatic Identification of Argument Relations: A Cross-Domain Evaluation},
year = {2021},
volume = {36},
number = {06},
issn = {1941-1294},
pages = {62-70},
doi = {10.1109/MIS.2021.3073993},
publisher = {IEEE Computer Society}
}