GraphWOZ

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #typescript #llm #go #golang #android #machine-learning #rust #deep-learning #linux

★ 9 Open GitHub ↗

No description available.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# GraphWOZ

This repository contains code for:
1) Transforming raw JSON formatted GraphWOZ data into a PyTorch Geometric graph Dataset with BERT encodings
2) Entity-linking by cosine similarity
3) Training and evaluating a GNN for entity-linking with GraphWOZ Data.

# Preprocessing

homogeneous_graph_transform.py transforms the raw json data to PyTorch Geometric Data objects using BERT to encode strings as node and edge attributes.
Please note: 
1) This is slow, because it does not currently use CUDA 
2) Each dialogue turn is saved as a separate graph in the "dialogues" folder

# Entity-Linking

Two main files handle this: cosine_compare_entities.py contains the code for computing Precision@1 and MRR metrics for entity mention-entity links in the graph.
The method for calculating this will also output a file for the dataset containing triples in the form of (Mention, Predicted Entity, True Entity)

The file gnn_link_prediction.py imports from the previous file to first calculate these over the (train, validation, test) splits.
Subsequently, the GNN then trains on the train dataset and outputs a file with predicted links and precision@1 for the validation set.
The final step is evaluation on the test set, which outputs a similar file as before, but also including MRR calculated over the output.

# Audio Files

Link to download audio files: https://files.nr.no/samba/GraphDial/

GraphWOZ

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README