This repo is a pipeline that takes an HTML/JSON formatted news articles and do several NLP tasks on them.

An input file will cross by all/some the components of the pipeline. The output of each component will be written directly to a copy of the given file. where the output of each component is as follows,

• Document classifier -> 'doc_label'=0/1

• Violent_classifer -> 'is_violent'=0/1

• Sentence_classifier will tokenize the sentences and get their labels for 4 different classifers->

  • 'sentences'=[]
  • 'sentence_labels'=[]
  • 'sentence_tokens'=[[]]
  • 'Trigger_Semantic_label'=[]
  • 'participant_semantic'=[]
  • 'organizer_semantic'=[]

• Token_classifers -> 'token_labels'=[[]]

• report script (bin/out_to_csv.py) will create a detailed report on sentence base and merge coreferensed setnences using [NS] token.

An example output can be found at the end of the page.

Pipeline Configuration.

• when cascaded parameter is true

It will run the pipeline as illustrated in the flowchart below. Where no filtering will be applied hence, files will cross by all component in all level.

Flowchart,

gif version

• when cascaded parameter is false It will run the pipeline as illustrated in the flowchart below. Where filtering will be applied and only postive sentences of a postive document will be passed.

gif version

Prerequisite

• Java 7 or 8
• Python3

Installation

install.sh will install/download the following models in $HOME/.pytorch_pretrained_bert folder.

• Nextflow
• Python requirments libraries (requirement.txt)
• OsmanMutlu/pytorch-pretrained-BERT
• Document Models (Protest classifier (BERT) - Violent Classifier (SVM) )
• Sentence Model (Protest classifier (BERT) - Participant Semantic Categorization (BERT), Trigger Semantic Categorization (BERT), and Organizer Semantic Categorization (BERT)- Coreference Model (ALBERT))
• Token Model(Token Classifier)
• Change "/PATH/TO/REPO" and ENV_NAME variables accordingly!

echo 'PATH="$PATH:/PATH/TO/REPO/bin"'
conda create -n ENV_NAME python=3.6
conda activate ENV_NAME
cd /PATH/TO/REPO
bash install.sh

Parameters

It can be modified in nextflow.conf file

• cascaded

  true/false

  as it's explained in 'Pipeline Configuration' section.

• classifier_first

  true/false

  if false pipeline starts with extractor.

  if true the first component of the pipeline will be the classifier, hence the input files must be JSON formatted and contain the following variables,

  {
      "id":str,
      "length":int,
      "text":str,
      "time":str,
      "title":str
  }
  

• gpu_classifier

  GPU id or group of GPU ids that is/are assgined to document proteset classifier.

  A single number or series of number joined by comma ie 1,2,3

• gpu_number_tsc

  GPU id or group of GPU ids that is/are assgined to trigger semantic classifier.

  A single number or series of number joined by comma ie 1,2,3

• gpu_number_psc

  GPU id or group of GPU ids that is/are assgined to participant semantic classifier.

  A single number or series of number joined by comma ie 1,2,3

• gpu_number_osc

  GPU id or group of GPU ids that is/are assgined to organization semantic classifier.

  A single number or series of number joined by comma ie 1,2,3

• gpu_number_protest

  GPU id or group of GPU ids that is/are assgined to sentence protest classifier.

  A single number or series of number joined by comma ie 1,2,3

• gpu_token

  GPU id or group of GPU ids that is/are assgined to token classifier.

  A single number or series of number joined by comma ie 1,2,3

• input

  input folder Path

  "<PATH>" .not ending with backslash

• output

  output folder path

  "<PATH>/jsons/"

• resume

  true/false

  nextflow's parameter https://www.nextflow.io/docs/latest/getstarted.html#modify-and-resume

• files_start_with

  file names pattern

  ie "*" ,"*json" ,"\http*"

• doc_batchsize

  Document protest classifier batch size

• token_batchsize

  token classifier batch size

• prefix

  path to pipeline scripts

  ie "<Path>/emw_pipeline_nf"

• extractor_script_path

  path to html to text script

  ie $prefix"/bin/extract/peoples_chaina.py" #

out_to_csv script's parameters

• out_output_type

  type of output record file

  csv/json

• out_name_output_file

  name of outfile record file

• out_date_key

  the key of date inside the dataset

  ie "time","date"

• filter_unprotested_sentence

  true/false

  filter protest sentences in the record file.

  if true only sentences that are postive will be inculded.

  if false all the sentences will be included irrespective of their label

• filter_unprotested_documents

  true/false

  filter protest document in the record file.

  if true only sentences that are in postive document will be inculded.

  if false all the sentences will be include irrespective of their document label

Run

Set your parameters in nextflow.conf first. start.sh script will start the FLASK API of classifiers then will format the paramters that are in nextflow.conf to a JSON file and run nextflow. After nextflow is done reporting part will be started.

bash start.sh

Output

• event_id The id of the event.
• url Using your browser, you can go this url to check the text of the document this event was extracted from.
• title Title of the document this event was extracted from.
• event_sentence_numbers All of the information of this event were extracted from these sentences. Note that our tool's sentence separation/numbering might not be perfect.
• district_name The district of the event. Categorizable. If empty, check for state_name.
• state_name State of the event. Categorizable. If empty, specific_place_name must not be empty.
• specific_place_name Specific place name returned by geopy tool. Might be empty since geopy might not have been used.
• year Publishing year.
• month Publishing month. Might not be available for some sources.
• day Publishing day. Might not be available for some sources.
• urbanrural Indicates whether the event takes place in a "urban" or "rural" setting. Decided on document level.
• violent Indicates whether the event is "violent" or "non-violent". Decided on document level.
• triggers Extracted triggers of the event.
• eventcategory Semantic category of the event.
• participants Extracted participants of the event.
• participant0category - participant3category Semantic category of the participants of the event. Goes from most common to least common, "participant0category" being the most common one.
• organizers Extracted organizers of the event.
• organizer0category - organizer3category Semantic category of the organizers of the event. Goes from most common to least common, "organizer0category" being the most common one.
• targets Extracted targets of the event.