TextMine 2025

This repository contains the experiments conducted for the TextMine 2025 challenge.

The majority of our work focused on ATLOP-based models, namely LACE and EIDER.

Install Environment

The experiments were conducted using Python 3.9.19.

Installing Dependencies

First, install the required dependencies listed in the requirements.txt file:

pip install -r requirements.txt

Installing DGL

If running on GPUs, install the appropriate version of DGL based on your CUDA version. Replace XXX with the CUDA version you are using:

pip install dgl==1.0.0+cuXXX -f https://data.dgl.ai/wheels/cuXXX/repo.html

If not using GPUs, install the CPU version:

pip install dgl==1.0.0

Installing Apex

Install Apex by following the instructions available at the NVIDIA Apex GitHub repository. Ours experiments have been run under apex==0.1

Running the Models

There are three main approaches for training and evaluating the models:

• Train on 80% of the dataset, evaluate on 10%, and test on the remaining 10%:
  In this setup, the proportions of training, evaluation, and testing datasets are maintained for each relation type.
  Example command:

  sh ./LACE/scripts/run_texmine_xlm_roberta.sh

• Cross-validation:
  This approach uses 5-fold cross-validation to ensure robust evaluation.
  Example command:

  sh ./LACE/scripts/run_texmine_xlm_roberta_cross_validation.sh

• Train on the entire training set and predict on the test set:
  This method was used for to get predictions for the submissions.
  Example command:

  sh ./LACE/scripts/run_texmine_xlm_roberta_full_train.sh

Model Combination

To improve the robustness and accuracy of predictions, we implemented a voting-based model combination strategy. This approach aggregates the predictions from multiple models using a voting mechanism.

Two voting methods are available:

• Absolute Majority:
  A prediction is selected only if more than 50% of the models made this prediction.

• Relative Majority:
  For a given pair of entities, the prediction with the highest number of votes is selected, even if it does not exceed the 50% threshold.

Command:

sh ./combine/scripts/run_vote.sh