Yitayal-hub/Amharic-Coreference-Resolution

Amharic Coreference Resolution with Integrated NER and Multi-Head Attention It is an advanced coreference resolution system that integrates Named Entity Recognition (NER) with multi-head attention mechanisms. #Installation Prerequisites Python 3.6+ TensorFlow 2.0+ Other dependencies listed in requirements.txt #Setup Install dependencies: bash pip3 install -r requirements.txt Build custom TensorFlow kernels: bash -x -e setup_all.sh Prepare training data and download Amharic word2vec from this link sparknlp.org: Divide the dataset into 80%, 20%, and 20% for training, development and testing data respectively. Prepare Embeddings: bash -x -e setup_training.sh # Filter embeddings for your dataset python3 filter_embeddings.py cc.am.300.vec train.jsonlines dev.jsonlines test.jsonlines # Generate character vocabulary python3 get_char_vocab.py #Usage Training bash python3 train.py [--logdir ] Example: bash python3 train.py train_am_coref --logdir logs/amharic_coref Evaluation bash python3 evaluate.py [ ] Output Format The system outputs predictions in JSON Lines format with the following structure: json { "doc_key": "document_id", "clusters": [[[start, end], ...], ...], "sentences": [...], "predicted_ner": ["PER", "O", "ORG", ...], ... }