Skip to content

Latest commit

 

History

History
113 lines (73 loc) · 4.75 KB

README.md

File metadata and controls

113 lines (73 loc) · 4.75 KB
Raw

Repository of the TPDL 2019 paper: A Hierarchical Label Network for Multi-Label EuroVoc Classification of Legislative Contents

Article published at: 23rd International Conference on Theory and Practice of Digital Libraries, TPDL 2019, Oslo, Norway, September 9-12, 2019, Proceedings.

  • BibTeX:
@InProceedings{HLAN2019,
author="Caled, Danielle and Won, Miguel  and Martins, Bruno and Silva, M{\'a}rio J.",
title="A Hierarchical Label Network for Multi-label EuroVoc Classification of Legislative Contents",
booktitle="Digital Libraries for Open Knowledge",
year="2019",
pages="238--252",
isbn="978-3-030-30760-8"
}

Overview

This repository contains EUR-Lex PT, a large scale and multi-label dataset with more than 220k documents, labeled under the three EuroVoc hierarchical levels. The dataset is available with the original division in training (64%), validation (16%), and test (20%) sub-sets. The training, validation and test sets have 140883, 35189 and 44254 documents, respectively, and the entire corpus has a vocabulary with 42556 words.

We also added the source code of our hierarchical deep learning model (HLAN) to address the classification of legal documents according to the EuroVoc thesaurus. Instead of training a classifier for each level, our model allows the simultaneous prediction of the three levels of the EuroVoc thesaurus.

Classifier

  • This model was implemented with Python 3.6.
  • For installation, we recommend the usage of virtualenv.

Installation

  1. Please, install the packages listed on the requirements.txt file.

    1.1 Create a virtual environment: virtualenv -p python3 envname

    1.2 Run envname/bin/pip install -r requirements.txt

  2. Download CBOW 300 dimensions (cbow_s300.zip) from NILC-Embeddings and unzip it to data/embeddings/ directory.

  3. Download the pre-processed EUR-Lex-PT dataset clean_txt_withallwords.json to data/ directory.

Running the model

  • model input: A json file containing documents with identifiers and textual data.
  • model output: Predictions on EuroVoc labels for domain, microthesauri and descriptor levels.

  1. Run classifier/hierarchical.py.

  2. The predicted EuroVoc labels are saved to the results/ directory, according to each EuroVoc level. The correspondence of the EuroVoc labels and terms can be obtained from the EuroVoc Thesaurus at data/eurovoc.json.

Attention: You can provide your own input at a desired format by modifying the load_eurlex_pt function accordingly.

Data

EUR-Lex-PT

The pre-processed EUR-Lex-PT is formated with a JSON representation as follows:

{   
    "celex": "eur-lex_identifier", 
    "url": "eur-lex_url",
    "eurovoc_ids": ["descriptor_index_1", "descriptor_index_2", ..., "descriptor_index_n"], 
    "eurovoc_mts": ["micro-thesaurus_index_1", "micro-thesaurus_index_2", ..., "micro-thesaurus_index_n"],
    "eurovoc_dom": ["domain_index_1", "domain_index_2", ..., "domain_index_n"], 
    "txt": "document_text"
}

Note: The documents' sentences ("txt" field) are separated by a __SENT__ delimiter.

EUR-Lex-PT LIBSVM format

The EUR-Lex-PT corpus is also available under the LIBSVM format at EUR-Lex-PT LIBSVM:

  • Each LIBSVM file contains a header in the format:

<#instances> <#features> <#classes>

  • The other lines of the LIBSVM files are in the format:

<label> <index1>:<value1> <index2>:<value2> ...

EuroVoc Thesaurus

The EuroVoc thesaurus can be found at the data/eurovoc.json file. Each descriptor is organized according to the structure presented bellow. As the descriptors may contain equivalent terms (preferred and non-preferred), we list all these terms under the terms list.

"descriptor_index": {
        "id": "descriptor_index",
        "terms": [
            "descriptor_label_1",
            "descriptor_label_2",
        ],
        "preferred_term": "descriptor_label",
        "micro-thesauri": {
            "micro-thesaurus_index_1": "micro-thesaurus_label_1",
            "micro-thesaurus_index_2": "micro-thesaurus_label_2"
        },
        "domains": {
            "domain_index_1": "domain_index_1",
            "domain_index_2": "domain_index_2",
        }
    }

Contact

I am happy to answer questions and to give additional information via email: dcaled at gmail dot com