DeepOnto.KOM

Python 3 package for relation extraction and classification.

pip install . to install the package

Requirements

Python 3 packages

• numpy
• pandas
• nltk
• textblob
• sklearn
• keras
• tqdm

For embedding creation one of the following are required

• fasttext (https://github.com/facebookresearch/fastText or https://github.com/salestock/fastText.py on Windows)
• GloVe binaries (https://github.com/stanfordnlp/GloVe)

Examples

Usage examples can be found in the examples folder

API Overview

preprocessing.py

download_preprocessing_prerequisites()
text_blob_from_file(file_path)
remove_stop_words(text_blob)
link_noun_phrases(text_blob)

embeddings.py

class WordEmbeddings
    load()
    embedding_for(word)
    words
    as_dict()
    as_key_values()

class GloVeEmbeddings(path) # For creating GloVe vectors
    train(self, corpus_path, epochs=30, embedding_size=300,
          context_size=15, min_occurrences=5, glove_path=".",
          threads=8, keep_csv=False)

class FastTextEmbeddings(path) # For FastText models
    train(self, input_file, **kwargs)

class DataFrameEmbeddings(path) # For loading csv files

def create_relation_dataset(embeddings, out_relations_path, out_labels_path,
                            relation_paths, out_false_relations_path=None, max_per_class=None,
                            unknown_word=None)

relationextraction.py

extract_concept_net(file_path, out_path, chunk_size=1000)
extract_yago(file_path, out_path, chunk_size=1000)
extract_relations_hearst(corpus_path, stop_words_path)

classification.py

class RelationClassifier
    new(self, input_dim, relation_count, one_hot=False, filters=32, max_filters=128,
        subtract_embeddings=False, dropout=False, learn_rate=0.001, optimizer="rmsprop",
        kernel_size=3, lr_decay=0)
    save(self, path)
    load(self, path)
    train(self, relations, labels, batch_size=256, validation_split=0.1, epochs=10,
          val_data=None, verbose=1)
    predict(self, relations)

visualization.py

show_embeddings_tsne(embeddings, word_count=1000, size=(100, 100), save_path=None,
                     clusters=None, **tsne_args)

clustering.py

class EmbeddingClusterer
    cluster(self, embeddings, min_clusters=5, max_clusters=100)