A Transformer-based Sequence-labeling Approach to the Slovenian Cross-domain Automatic Term Extraction

1. Description

In this repo, we experiment XLM-RoBERTa on RSDO5 dataset to evaluate the performance of multilingual pretrained language models in the cross-domain sequence-labeling setting of automatic term extraction (ATE) and compare our models' performance towards the benchmarks.

2. Requirements

Please install all the necessary libraries noted in requirements.txt using this command:

pip install -r requirements.txt

3. Data

The experiments were conducted on RSDO5 dataset, which is a collection of 12 documents with altogether about 250,000 words from the Slovenian language. The dataset is available at RSDO5 1.1 and contains the following domains:

	RSDO5 dataset
Languages	Slovenian
Domains	Biomechanics, Chemistry, Veterinary, Linguistics

4. Implementation

As the orginal dataset does not follow IOB format, we preprocess the data to sequentially map each token with it regarding label. An example of IOB format is demontrated below.

For ACTER dataset, run the following command to preprocess the data:

preprocess.py [-corpus_path] [-term_path] [-output_csv_path] [-language]

where -corpus_path is the path to the directory containing the corpus files, -term_path is the path to the directory containing the term files, -output_csv_path is the path to the output csv file, and -language is the language of the corpus.

For RSDO5 dataset, the dataset is already in conll format. Please use read_conll() function in preprocess.py to get the mapping.

Run the following command to train the model with all the scenarios in RSDSO5 datasets:

run.sh

where run.sh covers the 12 scenarios regarding 12 combination of train-valid-test splits with two domains of training, one domains of validation and one domains of testing.

Feel free to hyper-parameter tune the model. The current settings are:

    num_train_epochs=20,             # total # of training epochs
    per_device_train_batch_size=32,  # batch size per device during training
    per_device_eval_batch_size=32,   # batch size for evaluation 
    learning_rate=2e-5,              # learning rate
    eval_steps = 500,
    load_best_model_at_end=True,     # load the best model at the end of training
    metric_for_best_model="f1",
    greater_is_better=True,

5. Results

5.1 XLM-RoBERTa in mono-lingual settings

Test = ling	Test = vet
Validation Test F1-score vet ling 66.69 bim ling 71.51 kem ling 69.15	Validation Test F1-score ling vet 68.82 kem vet 68.94 bim vet 68.68

Test = kem	Test = bim
Validation Test F1-score ling kem 61.16 bim kem 64.83 vet kem 64.27	Validation Test F1-score vet bim 65.11 ling bim 63.69 kem bim 63.16

5.2 Comparison with benchmark

Test = ling	Test = vet
Methods F1-score vet 34.10 bim 71.51	Methods F1-score vet 29.90 bim 68.94

Test = kem	Test = bim
Methods F1-score vet 37.80 bim 64.83	Methods F1-score vet 33.90 bim 65.11

Contributors:

• 🐮 TRAN Thi Hong Hanh 🐮
• Matej MARTINC
• Andraz Repar
• Prof. Senja POLLAK
• Prof. Antoine DOUCET