davidsbatista / musico

A Minwise Hashing Method for Addressing Relationship Extraction from Text

Minwise Hashing for Relationship Extraction from Text

The MinHash-based Semantic Relationship Classifier (MuSICo) is an on-line approach for extracting of semantic relationships, based on the idea of nearest neighbor classification.

Instead of learning a statistical model, it finds the most similar relationship instances in a database and uses these similarities to make the decision of whether the sentence holds a certain relationship type. The sentence is classified according to the relationship type of the most similar relationship instances in a database.

The computation is done by leveraging min-hash and locality sensitive hashing for efficiently measuring the similarity between instances.

Usage:

First just ant to compile the source code, which should generate MuSICo.jar based on build.xml    

ant

All the external libs needed by MuSICo are in the libs/ directory. Then you can call MuSICo.jar with the following parameters, e.g.:

java -cp libs/*:MuSICo.jar bin.Main semeval true 400 50 5

parameters

MuSICo.jar bin.Main dataset true|false #min-hash-sigs #bands #kNN [train_file] [test_file]

dataset           semeval wiki aimed wikipt
true|false        generate shingles ? if false need to pass train_file/test_file
#min-hash-sigs    number of hash signatures
#bands            size of the LSH bands
#kNN              number of closest neighbors to consider

References

David S. Batista, Rui Silva, Bruno Martins, Mário J. Silva, A Minwise Hashing Method for Addressing Relationship Extraction from Text in Web Information Systems Engineering (WISE), 2013

David Soares Batista, David Forte, Rui Silva, Bruno Martins, Mário Silva, Exploring DBpedia and Wikipedia for Portuguese Semantic Relationship Extraction in Linguamática, 5(1), 2013.

David S. Batista, Ph.D. Thesis, Large-Scale Semantic Relationship Extraction for Information Discovery (Chapter 4), Instituto Superior Técnico, University of Lisbon, 2016

musico's Issues

Sample

This project looks interesting, do you have some samples I can try.. also is there a way to train the information.

Please help me on this.

Krishna