raldir/taxi

TAXI: a Taxonomy Induction Method based on Lexico-Syntactic Patterns, Substrings and Focused Crawling

TAXI: a Taxonomy Induction Method based on Lexico-Syntactic Patterns, Substrings and Focused Crawling

This page contains implementation of a method for taxonomy induction that reached the first place in the SemEval 2016 challenge on taxonomy extraction evaluation. The method builds a taxonomy from a domain vocabulary. It extracts hypernyms from substrings and large domain-specific corpora bootstrapped from the input vocabulary. Multiple evaluations based on the SemEval taxonomy extraction datasets of four languages and three domains show state-of-the-art performance of our approach. This page contains implementations of the method including all resources needed to reproduce experiment described in the following paper presented in San Diego at SemEval co-located with the NAACL'2016:

@inproceedings{panchenko2016taxi,
  title={TAXI at SemEval-2016 Task 13: a Taxonomy Induction Method based on Lexico-Syntactic Patterns,  Substrings and Focused Crawling},
  author={Panchenko, Alexander and Faralli, Stefano and  Ruppert, Eugen and Remus, Steffen and  Naets, Hubert and  Fairon, Cedrick and Ponzetto, Simone Paolo and Biemann, Chris},
  booktitle={Proceedings of the 10th International Workshop on Semantic Evaluation},
  year={2016},
  address={San Diego, CA, USA},
  organization={Association for Computational Linguistics}
}

If you would like to refer to the system please use this citation. More information about the approach can be found at the TAXI web site.

System Requirements

The system was tested on Debian/Ubuntu Linux and Mac OS X. To load all resources in memory you need about 64 Gb of RAM.

Installation

1. Clone repository:

git clone https://github.com/tudarmstadt-lt/taxi.git

2. Download resources into the repository (4.4G compressed by gzip):

cd taxi && wget http://panchenko.me/data/joint/taxi/res/resources.tgz && tar xzf resources.tgz

3. Install dependencies for using pygraphviz:

$ sudo apt-get install python-dev graphviz libgraphviz-dev pkg-config

4. Install project dependencies:

pip install -r requirements.txt

5. Setup spaCy. Download the language models for English, Dutch, French and Italian

$ python -m spacy download en
$ python -m spacy download nl
$ python -m spacy download fr
$ python -m spacy download it

6. Setup NLTK

$ python -m nltk.downloader stopwords
$ python -m nltk.downloader wordnet

Induction of SemEval Taxonomies

Run the semeval.py to reproduce experimental results, e.g.:

For a test run (few resources loaded, quick):

python semeval.py vocabularies/science_en.csv en simple --test

For a normal run (all resources are loaded, requires 64Gb of RAM):

python semeval.py vocabularies/science_en.csv en simple

Afterwards a noisy graph is being created. Clean the output by executing(this example uses the inputfile science_en.csv-relations.csv-taxo-knn1.csv):

./run.sh taxi_output/simple_full/science_en.csv-relations.csv-taxo-knn1.csv

The vocabularies directory contains input terms for different domains and languages. The script lets you reproduce results in the SemEval 2016 Task 13 Taxonomy Extraction Evaluation described in the our paper. This script load hypernyms from the downloaded resources and constructs a taxonomy for every input vocabulary of the SemEval datasets, e.g. English Food domain. Generally, the TAXI approach takes as input a vocabulary and outputs a taxonomy for a linked subset of the terms from this vocabulary. Currently the main purpose of this repository is to ensure reproducibility of the SemEval results. The results taxonomies will be generated next to the corresponding input vocabulary file. If you need to adapt the script for your needs and require help do not hesitate to contact us.

Distributional Semantics

1. Download the required embeddings:

• $ wget http://ltdata1.informatik.uni-hamburg.de/taxi/embeddings/embeddings_poincare_wordnet
• $ wget http://ltdata1.informatik.uni-hamburg.de/taxi/embeddings/own_embeddings_w2v
• $ wget http://ltdata1.informatik.uni-hamburg.de/taxi/embeddings/own_embeddings_w2v.trainables.syn1neg.npy
• $ wget http://ltdata1.informatik.uni-hamburg.de/taxi/embeddings/own_embeddings_w2v.wv.vectors.npy

2. Set the directory path in line 45 of distributional_semantics.py to the directory containing the embeddings download above.

3. To apply distributional semantics to the generated taxonomy, use the script distributional_semantics.py or the notebook distributional_semantics.py.ipynb

The script can be used with following options:

Option	Alternate	Description	Default	Choices
--taxonomy	-t	Input file containing the taxonomy	-	-
--mode	-m	Mode of the algorithm	ds	ds, root, remove
--domain	-d	Domain of the taxonomy	science	science, science_wordnet, food, environment_eurovoc
--exparent	-ep	Exculde parent while calculating cluster similarity	False	-
--exfamily	-ef	Exculde family while calculating cluster similarity	False	-

Example:
$ python distributional_semantics.py -t taxi_output/simple_full/science_en.csv-relations.csv-taxo-knn1.csv -d food -ep

Visualizing taxonomies

To visualize the taxonomy structures in a .csv file, you must have Networkx and Pygraphviz setup in your environment.

To construct a hierarchical taxonomy structure:
$ python visualize_taxonomy.py --file <csv filename>

The images generated will be very large, so alternatively, the graph can be constructed inside the notebook networkx_graph.ipynb

taxi's Issues

Print the nearest neighbors

Print the nearest neighbors for these words and send by email (sort by confidence): apple, mango, BMW, USA, Berlin, IBM, jaguar. The candidates are ALL words in the vocabulary of the pre-trained hypenet model (e.g. ~100000 words).

HypeNet: use pre-trained embeddings containing MWE terms

Make the prediction for all domain of the taxonomy induction

Make the prediction (in the similar fashion as #2 ) but for all domain of the taxonomy induction challenge. The candidates are from the vocabulary of the domain-specific words (e.g. food, science, etc. ~500–1000 words).

Compute the taxonomy based on the HypeNet

Instead of using the pre-extracted database of hypernyms, use HypeNet and evaluate on the shared task. Add the results to the (a) Google Spreadsheet.