Spatial Nominal Entity Recognition

Overview

This repository contains the source code for evaluating ML models trained for Spatial Nominal Entity Recognition as proposed in

Amine Medad, Mauro Gaio, Ludovic Moncla, Sébastien Mustière, and Yannick Le Nir. Comparing supervised learning algorithms for Spatial Nominal Entity recognition. The 23rd AGILE International Conference on Geographic Information Science. 2020

Datasets are given in the corpus directory and models in the models directory.

Installation

Install the required python librairies:

pip3 install -r requirements.txt

Then you need to download the binary file of the pretrained French FastText model (4.2 Go) and add it to the data directory:

wget -P data https://dl.fbaipublicfiles.com/fasttext/vectors-crawl/cc.fr.300.bin.gz
gzip -d data/cc.fr.300.bin.gz

TreeTagger needs also to be installed with the French parameter file before running the script:

mkdir TreeTagger
cd TreeTagger

wget https://cis.uni-muenchen.de/~schmid/tools/TreeTagger/data/tree-tagger-linux-3.2.3.tar.gz
tar -xzf tree-tagger-linux-3.2.3.tar.gz

wget https://cis.uni-muenchen.de/~schmid/tools/TreeTagger/data/tagger-scripts.tar.gz
tar -xzf tagger-scripts.tar.gz

wget https://cis.uni-muenchen.de/~schmid/tools/TreeTagger/data/install-tagger.sh
bash install-tagger.sh

cd lib/
wget https://cis.uni-muenchen.de/~schmid/tools/TreeTagger/data/french.par.gz
gunzip french.par.gz

Usage

python3 evaluate_model_snoer.py -i <input_dataset> -n <ngram_size> -alg <algorithm_name> -m <model_filepath> -ft <fasttext_model> -fr_nouns <french_nouns_filepath> -s <we_size_vec> -ti <train_dataset>

• <input_dataset>: filepath to the csv input data
• <train_dataset>: filepath to the csv training data (use for PCA fitting for the model MLP+PCA only)
• <fasttext_model>: filepath of the pretrained FastText binary model
• <french_nouns_filepath>: filepath of the file containing French nouns (use for padding ngrams)
• <algorithm_name>: name of the architecture used for training (GRU, MLP+AE, MLP+PCA, SVM, RF)
• <model_filepath>: filepath of the model to evaluate
• <ngram_size>: size of the ngram (1, 5 or 7)
• <we_size_vec>: Word Embedding dimension (default: 300)

You can also download and execute the jupyter notebook version.

Example

Run the following command to evaluate the GRU model trained with 5 grams :

python3 evaluate_model_snoer.py -i "./data/corpus_validation.csv" -n 5 -alg "GRU" -m "./models/GRU_5grams.h5" -ft "./data/cc.fr.300.bin" -fr_nouns "./data/French_nouns.txt" -ti "./data/corpus_train.csv" 

Results

Model	GRU			RF			SVM
ngram_size	1 g	5 g	7 g	1 g	5 g	7 g	1 g	5 g	7 g
Accuracy	0.67	0.76	0.79	0.71	0.73	0.74	0.69	0.75	0.72

Model	MLP + AE			MLP + PCA
ngram_size	1 g	5 g	7 g	1 g	5 g	7 g
Accuracy	0.68	0.75	0.78	0.49	0.64	0.60

Acknowledgement

This work is supported and funded in part by French National Research Agency (ANR) under the CHOUCAS project (ANR-16-CE23-0018).

The CHOUCAS project is a French interdisciplinary research project aiming to respond to a need expressed by the high mountain gendarmerie platoon to help localising victims in mountain area.