Skip to content

README.md

Latest commit

 

History

History
44 lines (30 loc) · 3.88 KB

README.md

File metadata and controls

44 lines (30 loc) · 3.88 KB

ABSCL (ABSA LLM-CL)

[ English | 中文 ]

1. Core Idea

ABSCL (ABSA LLM-CL) is a continual learning method designed for ABSA. It trains two types of adapters:

  1. Shared Adapter: Trained using a replay strategy, combining a subset of data from the current task and previous tasks. This adapter aims to capture general knowledge across tasks.
  2. Task-Specific Adapter: Trained using only the data from the current task.

During the training of the task-specific adapter, ABSCL applies two constraints inspired by O-LoRA:

  • Orthogonality Constraint: Encourages the task-specific adapter's weights to be orthogonal to the shared adapter's weights, promoting task-specific knowledge separation.
  • L2 Regularization: Applies L2 regularization specifically to the shared adapter's weights (loaded as a reference) while training the task-specific adapter, potentially preventing catastrophic forgetting in the shared knowledge base.

Additionally, ABSCL extracts feature statistics (mean vector and a shared covariance matrix) for each task after training its specific adapter. These statistics can be used later by abscl_selector.py to determine the most suitable adapter (task) for a given input sample based on Mahalanobis distance.

2. Specific Parameters

The following parameters are specific to the ABSCL method (refer to finetuning_args):

  • --abscl_orthogonal_lambda: (float, required) The weight for the orthogonality constraint loss between the task-specific adapter and the shared adapter.
  • --abscl_shared_l2_lambda: (float, required) The weight for the L2 regularization loss applied to the shared adapter weights during task-specific adapter training.
  • --abscl_stats_path: (str, optional) Path to save and load the feature statistics (mean vectors and covariance matrix). Defaults to adapters_save_path/abscl_stats if not provided.
  • --current_task_id: (str, required) An identifier for the current task. Used for naming the task-specific adapter and storing its statistics.
  • --adapters_save_path: (str, required) The base directory where the shared_adapter and task-specific adapters (named by current_task_id) will be saved. Feature statistics will also be stored relative to this path if abscl_stats_path is not set.

Note: Parameters like --replay_ratio, --replay_task_list, --maxsamples_list, --previous_task_dataset are used by the ABSCL workflow for the replay strategy when training the shared adapter, but might be considered general replay parameters.

3. File Descriptions

  • abscl_workflow.py: Orchestrates the main ABSCL training process. It handles the replay data preparation, trains the shared adapter, trains the task-specific adapter with ABSCL constraints, and triggers feature statistic extraction.
  • abscl_trainer.py: Defines the ABSCLTrainer, a custom Hugging Face Trainer subclass. It modifies the loss computation to include the orthogonality and shared L2 regularization terms, leveraging O-LoRA mechanisms for calculation.
  • abscl.py: Contains the ABSCLFeatureExtractor class responsible for extracting hidden state features (specifically, the last token's hidden state from the second-to-last layer) and computing/updating the per-task mean vectors and the shared covariance matrix.
  • abscl_selector.py: A script used after training multiple tasks with ABSCL. It loads the saved feature statistics and a test dataset, then assigns the most likely task adapter to each test sample based on the Mahalanobis distance of the sample's feature representation to the task means.

4. Citation

@article{ding2024boosting, title={Boosting large language models with continual learning for aspect-based sentiment analysis}, author={Ding, Xuanwen and Zhou, Jie and Dou, Liang and Chen, Qin and Wu, Yuanbin and Chen, Chengcai and He, Liang}, journal={arXiv preprint arXiv:2405.05496}, year={2024} }