Extracting CS Citation Graph From Semantic Scholar Corpus

Load Data From Semscholar

1. !aws s3 cp --no-sign-request --recursive s3://ai2-s2-research-public/open-corpus/2020-05-27/ corpus_data

2. Upload this Data To your Bucket which is specified in ~/.metaflowconfig/config.json to a path like `$S3_PATH/datasets/semantics_scholar_corpus_data/

   from metaflow import S3
   import os
   
   data_path = '/'.join(S3.get_root_from_config([]).split('/')[:-1])
   s3_root = os.path.join(data_path,'datasets')
   
   def sync_folder_to_s3(root_path,base_path='',s3_root=s3_root):
       sync_path = os.path.join(s3_root,base_path)
       
       file_paths = [(os.path.normpath(os.path.join(r,file)),os.path.join(r,file)) for r,d,f in os.walk(root_path) for file in f]
       # return file_paths
       for normpth,pth in file_paths:
       with S3(s3root=s3_root) as s3:
           file_paths = s3.put_files([(normpth,pth)])
           print(f"Finished Writing Path : {pth}")
   
       sync_path = os.path.join(
           sync_path,os.path.normpath(root_path
       ))
       return sync_path,file_paths
   sync_path,file_paths = sync_folder_to_s3('corpus_data',base_path='semantics_scholar_corpus_data')

Flow Descriptions

The flow are follow the following pattern:

1. Create clean data from large mound of semantic scholar cropus. Cleaning can be removing all uncited material.
2. Use the cleaned data to extract the CS based
3. Use the CS based info to create Graph and classify ontology
4. Use graph to calc page rank and collate that with ontology results.
5. Dump this final processed information to elasticsearch

1. SemScholarCorpusFlow

• citation_harvest_flow.py contains SemScholarCorpusFlow which will extract and remove all uncited material and saves dataframe for Each chunk under the Bucket $CONFIG_S3_ROOT/processed_data/SemScholarCorpusFlow/s2-corpus-{i}/usefull_citations.csv

2. CSDataExtractorFlow

• CSDataExtractorFlow -----Requires-->SemScholarCorpusFlow
• cs_citation_extractor.py contains CSDataExtractorFlow which will use the Chunks set by SemScholarCorpusFlow and extract Computer Science related Articles. It stores each chunk under the Bucket $CONFIG_S3_ROOT/processed_data/CSDataExtractorFlow/s2-corpus-{i}/category_citations.csv

3. CSDataConcatFlow

• CSDataConcatFlow ----Requires--> CSDataExtractorFlow
• concat_cs_flow.py contains CSDataConcatFlow which concatenates the data into a single dataframe. This doesn't turn out that useful as csv becomes 22GB in size. The concat csv is stored at $CONFIG_S3_ROOT/processed_data/CSDataConcatFlow/cs-concat-data.csv.

4. CSOntologyClassificationFlow

• CSOntologyClassificationFlow ----Requires--> CSDataExtractorFlow
• ontology_classify_flow.py contains CSOntologyClassificationFlow. It will use the cso_classifier. TODO: Push submodule which works as a relative import and Link it to repo. This module will classify the data in the chunked csvs according to ontology described here. The chunked dataframes are stored at $CONFIG_S3_ROOT/processed_data/CSOntologyClassificationFlow/s2-corpus-{i}/ontology_processed.csv

5. CSGraphBuilderFlow

• CSGraphBuilderFlow ----Requires--> CSDataExtractorFlow
• ontology_classify_flow.py contains CSGraphBuilderFlow which will use the inCitations and outCitations information from the chunked csv stored from CSDataExtractorFlow and store the graph to S3. The graph will be stored under $CONFIG_S3_ROOT/processed_data/CSGraphBuilderFlow/citation_network_graph.json

6. CSPageRankFinder

• CSPageRankFinder ----Requires--> CSGraphBuilderFlow
• page_rank_flow.py will use the graph stored via CSGraphBuilderFlow and performs page rank based on parameters. Finally stores the rank dictionary to $CONFIG_S3_ROOT/processed_data/CSPageRankFinder/page-rank-{run-id}.json

7. PageRankCollateFlow

• PageRankCollateFlow ----Requires--> CSPageRankFinder

• PageRankCollateFlow ----Requires--> CSOntologyClassificationFlow

• collate_page_rank_flow.py contains the flow that will collate Page-Rank results from CSPageRankFinder flow and ontology classification results from CSOntologyClassificationFlow. This will finally store chunked csvs to $CONFIG_S3_ROOT/processed_data/PageRankCollateFlow/s2-corpus-{i}/ontology_processed.csv

Utility Files

to_elastic.py

• Extracts then csv chunks created by the PageRankCollateFlow(which have ontology and page rank collated results) and throws the data to elasticsearch.

mf_utils.py

• Utility module for data-syncing/s3 looksup and Dataroot path fixing when creating the flows.

TODO

• Cleanup Readmes
• Add More information about running flows properly.
• Create flow dependency graph
• Replace Memory Hogging Networkx with https://github.com/VHRanger/CSRGraph for PageRank