Recognizing Semantic Differences (RSD)

Code for the EMNLP 2023 paper "Towards Unsupervised Recognition of Token-level Semantic Differences in Related Documents".

Installation

• Requires Python >= 3.7 and PyTorch
• pip install -r requirements.txt

Usage

from recognizers import DiffAlign, DiffDel, DiffMask

diff_align = DiffAlign("ZurichNLP/unsup-simcse-xlm-roberta-base")

a = "Chinese shares close higher Friday ."
b = "Chinese shares close lower Wednesday ."
result = diff_align.predict(a, b)
# DifferenceSample(
#   tokens_a=('Chinese', 'shares', 'close', 'higher', 'Friday', '.'),
#   tokens_b=('Chinese', 'shares', 'close', 'lower', 'Wednesday', '.'),
#   labels_a=(0.07324671745300293, 0.06292498111724854, 0.082577645778656, 0.1421372890472412, 0.2610551714897156, 0.1118348240852356),
#   labels_b=(0.07324671745300293, 0.06292498111724854, 0.082577645778656, 0.1421372890472412, 0.2709317207336426, 0.1118348240852356)
# )

Reproducing the Results

Table 1: Validation results

• python -m experiments.scripts.create_validation_table

Table 2: Test results

• python -m experiments.scripts.create_test_table

Table 3: Dataset statistics

• python -m experiments.scripts.create_dataset_statistics_table

Table 4: Latency measurements

• python -m experiments.scripts.create_latency_table

Table 5: Ablations for diff_del

• python -m experiments.scripts.create_validation_table_del_ablations

Figure 2: Additional results

• python -m experiments.scripts.create_negative_ratio_figure
• python -m experiments.scripts.create_document_length_figure
• python -m experiments.scripts.create_permutation_figure
• python -m experiments.scripts.create_languages_figure

Citation

@inproceedings{vamvas-sennrich-2023-rsd,
      title={Towards Unsupervised Recognition of Token-level Semantic Differences in Related Documents},
      author={Jannis Vamvas and Rico Sennrich},
      month = dec,
      year = "2023",
      booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing",
      address = "Singapore",
      publisher = "Association for Computational Linguistics",
}