Code developed by the IRIT_IRIS team to participate in the SemEval 2023 Task 6A (Legal Rhetorical Role Labeling). Details are available in the respective paper: IRIT_IRIS_A at SemEval-2023 Task 6: Legal Rhetorical Role Labeling Supported by Dynamic-Filled Contextualized Sentence Chunks.
- Cohan_and_Sharing_edges: models based on Cohan and Sharing edges chunk layouts.
- DFCSC-CLS: models based on DFCSC-CLS approach.
- DFCSC-SEP: models based on DFCSC-SEP approach.
- SingleSC: models based on single sentence classification approach.
Each folder contains a running script for each model. For example, to run the DFCSC-CLS RoBERTa model we execute the DFCSC-CLS/run_RoBERTa.py file. Running a model yields the respective report file. There is no command line parameters.
This repository does not contain the target dataset, though it is available at this link. Each model folder in this repository has a application.py file (the DFCSC-CLS folder has also the application_longformer.py file) that sets the path of the dataset. Before running a model, set this path accordingly the location of the dataset on your system.
The hyperparameters of a model can be set in the respective run_* script. In the following we describe such hyperparameters.
Generic hyperparameters (i.e., they are available in all models):
ENCODER_ID: identifier of the exploited pre-trained Transformer model from the Hugging Face repository (https://huggingface.co/models).MODEL_REFERENCE: name utilized to reference the model in the reports.MAX_SEQUENCE_LENGTH: number of tokens in a chunk (c_len).EMBEDDING_DIM: the embedding dimension of a token embedding. It is determined by the choosen pre-trained model.N_EPOCHS: the number of fine-tuning epochs.LEARNING_RATE: the initial learning rate of the fine-tuning procedure.BATCH_SIZE: batch size of the fine-tuning procedure.DROPOUT_RATE: dropout rate of the fine-tuning procedure.USE_MLP: boolean value to indicate if the classifier must comprise one (False) or two (True) dense layers.n_iterations: number of executions of the model. Each execution adopts a different random number seed value.weight_decay: weight decay value of the Adam optimizer.eps: epsilon value of the Adam optimizer.use_mock: boolean value to indicate if it should to use a mock model instead a real one. This is used as way to speed the runing time when the code is being validated.n_documents: number of documents to be used to train and evaluate a model. This is used as way to speed the runing time when the code is being validated.SAVE_BEST: ifTrue, the weights of the best epoch are saved in disk. IfFalse, nothing is saved.
SingleSC models:
freeze_layers: IfTrue, half of the first transformer layers are freezed and thus they are not updated during fine-tuning. IfFalse, all the transformer layers are updated.warmup: indicates if a learning rate warmup procedure must be adopted (True) or not (False).
DFCSC-CLS and DFCSC-SEP models:
MIN_CONTEXT_LENGTH(m): the desired minimum number of tokens in the edges of a chunk.
Cohan models:
MAX_SENTENCE_LENGTH: maximum number of tokens in a core sentence.MAX_SENTENCES_PER_BLOCK: maximum number of core sentences in a chunk.CHUNK_LAYOUT: the layout of the chunk. SetCohanto indicate a chunk without shared edges.
Sharing edges models:
MAX_SENTENCE_LENGTH: maximum number of tokens in a core sentence.MAX_SENTENCES_PER_BLOCK: maximum number of core sentences in a chunk.WINDOW_LENGTH: maximum number of shared sentences in each edge of a chunk.CHUNK_LAYOUT: the layout of the chunk. SetVanBergto indicate a chunk with shared edges.MAX_SEQUENCE_LENGTH: number of tokens in a chunk (c_len). The provided value is adjusted in order to minimize the use ofPADtokens and so the actual value may be lower than the provided one.