Skip to content

The repository contains the model used for a short EMNLP paper: "Reporting the unreported: Event Extraction for Analyzing the Local Representation of Hate Crimes"

Notifications You must be signed in to change notification settings

aidamd/HateCrime

Repository files navigation

Hate crime detection and entity extraction

This repo contains code and data for the following paper: "Reporting the unreported: Event Extraction for Analyzing the Local Representation of Hate Crimes". EMNLP 2019

The method includes a Mulit-instance Learning models for detecting hate crimes in local news articles and a RNN model for extracting the entities of each hate crime. The annotated datasets for three types of crimes (hate, kidnapping and homicide) is included in the Data folder.

Getting Started

In order to run the code you need to download the Data and embeddings directory.

Prerequisites

This project uses Python 3.6.2. The following libraries must be installed:

sklearn 0.19.1
tensorflow-gpu 1.11.0
pandas 0.23.0
nltk 3.2.5
tqdm 4.28.1
numpy 1.14.3

Parameters

All the parameters of the code are denoted in params.json. The parameters are defined as following:

  "hidden_size": 100 # hidden size of the LSTM
  "art_filter_sizes": [2, 3, 4] # filter sizes in detect model
  "art_num_filters": 10 # number of different filters in detect model
  "pretrain": true # if set to true, uses Glove embeddings
  "embedding_size": 300 # size of the embedding
  "learning_rate": 0.00001 # learning rate for the detect task
  "keep_ratio": 0.75 # keep ratio for the detect task
  "epochs": 30 # number of epochs
  "entity_keep_ratio": 0.75 # keep ratio in extract task
  "entity_learning_rate": 0.001 # learning rate in extract task
  "batch_size": 5 # size of batches, shows the number of articles in each batch

Running the detection code

In order to run the detection code, use the following script:

python3 run_detect.py --model <MODEL_NAME> --goal <GOAL> --dataset <DATASET> --params <PARAMS_FILE>

substitude the following tokens according to the task in mind:

  • <MODEL_NAME>: you can either use MICNN (the model used in the paper) or ATTN (the hierarchical attention baseline)
  • <GOAL>: the goal of the task is either train or predicts
  • <DATASET>: use one of the three datasets (hate, kidnap or homicide) to perform the detection
  • <PARAM_FILE>: is the .json file that includes all the model parameters. The model uses params.json as default.

Running the extraction code

In order to run the detection code, use the following script:

python3 run_extract.py --goal <GOAL> --params <PARAMS_FILE>

substitude the following tokens according to the task in mind:

  • <GOAL>: the goal of the task is either train or predicts
  • <PARAM_FILE>: is the .json file that includes all the model parameters. The model uses params.json as default.

About

The repository contains the model used for a short EMNLP paper: "Reporting the unreported: Event Extraction for Analyzing the Local Representation of Hate Crimes"

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages