Keras implementation for our paper:

• Zheng-Xin Yong, Tiago Timponi Torrent. (2020). Semi-supervised Deep Embedded Clustering with Anomaly Detection for Semantic Frame Induction. In: Proceedings of the Twelfth International Conference on Language Resources and Evaluation (LREC 2020), Marseille, France.

The dependencies are

• bcubed==1.5
• nltk==3.4.3
• matplotlib==3.2.1
• numpy==1.18.2
• Keras==2.2.5
• scikit_learn==0.23.0

Or simply, run pip3 install -r requirements.txt to install all the dependencies.

The data used in our research are as follows:

1. Berkeley FrameNet 1.7
2. FrameNet+
3. Curated anomalous lexical units (from WordNet). Can be accessed through the LRE Map repository.

We use the Python flair library to generate the embeddings for the lexical units using the exemplar sentences and their definitions.

The data/ folder contains the embeddings of the lexical units.

1. Create a folder trained_SDEC_AD/ for saving the trained weights.
2. Run the Python script python3 semantic_frame_induction_tr.py to train the SDEC-AD model.
3. Run the Python script semantic_frame_induction_pred.py to predict and evaluate the clusters of LUs. Remember to update the parameter SDEC_trained_weights in the script to the trained weight that has the largest Bcubed F1-score (which is indicated in the name of the saved trained weights such as "SDEC_AD_bcubed_fscore_0.788.h5").

1. Follow the instructions in the previous section "Semantic Frame Induction" to generate the trained weights.
2. Update the parameter SDEC_trained_weights in the Python script anomaly_detection.py to the trained weight that has the largest Bcubed F1-score. Then, run the Python script anomaly_detection.py to train the decoder and detect anomalous lexical units.