noonebug / multi_task_concept_invention

Repository which contains complementary resources about the paper: "Mapping Lexical Knowledge to Distributed Representations for Concept Invention"

INSTALLATION

You need to have installed jupyter notebook.

I advice you to create an enviroment (virtualenv or conda)

Setup your enviroment with pip install -r req.txt

The notebooks are divided basing on their purpose, to obtain a proper and correct model which can replicate the experiments of the paper, the notebooks are to be consulted (and surely, runned) in the order showed below:

DATA PREPARATION

• (0.1) wikipedia_Abstracts.ipynb: download and preprocess a bit the corpus
  • requires the corpus, you can download it like showed in the notebook or from here
• (0.2) import_elmo_embeddings.ipynb: align ELMo's vectors with words in corpus, create the datasets
  • requires the elmo vectors, you can generate yourself with an ELMo's implementation or you can download mine from here
  • Vectors can be downloaded in bulk (54.5 GB) or 50+ zip can be downloaded, if you download the 50+ zip you have to recompose with cat splitted* > elmo_vectors.zip
  • (0.2.1) composite_words.ipynb: retrieving of the word phrases
    • Requires the corpus and the elmo vectors
  • (0.2.2) minimal_type.ipynb: solving of polytipe words (words which are retrieved from more than one class) the polytiping have to be erased to avoid a multilabel problem

BUILD THE NETWORK, TRAIN, EVALUATE

• (1) network.ipynb:
  • requires the vectors of concepts (1.1, 1.2) and the datasets (1.3):
    • (1.1) vectors of type2vec that you can find here,
    • (1.2) vectors of HyperE that you can generate with HyperE or you can find here
    • (1.3) datasets are available here, you can generate with (0.2) or you can download it