pig script to filter GWB CDX by SURT regexes

6 files changed, 127 insertions, 0 deletions

diff --git a/pig/filter-cdx-paper-pdfs.pig b/pig/filter-cdx-paper-pdfs.pig
new file mode 100644
index 0000000..6559066
--- /dev/null
+++ b/pig/filter-cdx-paper-pdfs.pig
@@ -0,0 +1,41 @@
+
+-- Tries to filter down a large CDX file to a subset that is likely to be
+-- journal article content, based on SURT regex patterns.
+---
+-- Author: Bryan Newbold <bnewbold@archive.org>
+-- Date: May 2018
+
+
+%default INPUT ''
+%default OUTPUT ''
+
+set mapreduce.job.queuename default
+
+cdx = LOAD '$INPUT' AS cdxline:chararray;
+cdx = FILTER cdx BY not STARTSWITH (cdxline, 'filedesc');
+cdx = FILTER cdx BY not STARTSWITH (cdxline, ' ');
+
+cdx = FOREACH cdx GENERATE STRSPLIT(cdxline,'\\s+') as cols, cdxline;
+cdx = FOREACH cdx GENERATE (chararray)cols.$0 as surt, (chararray)cols.$1 as timestamp, (chararray)cols.$3 as mimetype, (chararray)cols.$4 as httpstatus, cdxline;
+cdx = FILTER cdx BY not surt matches '-';
+cdx = FILTER cdx BY httpstatus matches '200';
+cdx = FILTER cdx BY mimetype matches '.*pdf.*';
+
+-- This is the core regex
+cdx = FILTER cdx
+        -- academic domains; personal (tilde) directories
+        BY surt matches '(edu,|..,edu|..,ac,).*\\).*\\/~.*'
+
+        -- words in URL
+        OR surt matches '(?i).+\\).*/(pubs|research|publications?|articles?|proceedings?|papers?|fulltext)/.*'
+
+        -- words in domains 
+        OR surt matches '.*(,hal|,eprint|scielo|redalyc|revues|revistas|research|journal).*\\).*'
+
+        -- DOI-like pattern in URL
+        OR surt matches '.*\\).*/10\\.\\d{3,5}/.*';
+
+cdx = ORDER cdx by surt, timestamp PARALLEL 50;
+cdx = FOREACH cdx GENERATE cdxline;
+STORE cdx INTO '$OUTPUT' USING PigStorage(' ');
+
diff --git a/pig/tests/files/papers_domain_words.cdx b/pig/tests/files/papers_domain_words.cdx
new file mode 100644
index 0000000..48e2313
--- /dev/null
+++ b/pig/tests/files/papers_domain_words.cdx
@@ -0,0 +1,11 @@
+#http://research.fit.edu/sealevelriselibrary/documents/doc_mgr/448/Florida_Keys_Low_Island_Biodiversity_&_SLR_-_Ross_et_al_2009.pdf
+#http://ijs.sgmjournals.org:80/cgi/reprint/54/6/2217.pdf
+#http://eprints.ecs.soton.ac.uk/12020/1/mind-the-semantic-gap.pdf
+#http://eprint.uq.edu.au/archive/00004120/01/R103_Forrester_pp.pdf
+
+# should match 4:
+
+edu,fit,research)/sealevelriselibrary/documents/doc_mgr/448/Florida_Keys_Low_Island_Biodiversity_&_SLR_-_Ross_et_al_2009.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+org,sgmjournals,ijs)//cgi/reprint/54/6/2217.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+uk,ac,soton,ecs,eprints)/12020/1/mind-the-semantic-gap.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+au,edu,uq,eprint)/archive/00004120/01/R103_Forrester_pp.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
diff --git a/pig/tests/files/papers_edu_tilde.cdx b/pig/tests/files/papers_edu_tilde.cdx
new file mode 100644
index 0000000..47ca069
--- /dev/null
+++ b/pig/tests/files/papers_edu_tilde.cdx
@@ -0,0 +1,15 @@
+#http://www.stanford.edu:80/~johntayl/Papers/taylor2.pdf
+#http://met.nps.edu/~mtmontgo/papers/isabel_part2.pdf
+#http://www.pitt.edu:80/~druzdzel/psfiles/ecai06.pdf
+#http://www.comp.hkbu.edu.hk/~ymc/papers/conference/ijcnn03_710.pdf
+
+# should be 6 matches:
+hk,edu,hkbu,comp)/~ymc/papers/conference/ijcnn03_710.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+edu,stanford,www)/~johntayl/Papers/taylor2.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+edu,nps,met)/~mtmontgo/papers/isabel_part2.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+edu,pitt,www)/~druzdzel/psfiles/ecai06.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+jp,ac,pitt,www)/~druzdzel/psfiles/ecai06.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+co,edu,pitt,www)/~druzdzel/psfiles/ecai06.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+
+# NOT:
+com,corp,edu,pitt,www)/~druzdzel/psfiles/ecai06.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
diff --git a/pig/tests/files/papers_url_doi.cdx b/pig/tests/files/papers_url_doi.cdx
new file mode 100644
index 0000000..1ad5792
--- /dev/null
+++ b/pig/tests/files/papers_url_doi.cdx
@@ -0,0 +1,7 @@
+#http://journals.ametsoc.org/doi/pdf/10.1175/2008BAMS2370.1
+#http://www.nejm.org:80/doi/pdf/10.1056/NEJMoa1013607
+
+# should match 2:
+
+org,ametsoc,journals)/doi/pdf/10.1175/2008BAMS2370.1 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+org,nejm,www)/doi/pdf/10.1056/NEJMoa1013607 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
diff --git a/pig/tests/files/papers_url_words.cdx b/pig/tests/files/papers_url_words.cdx
new file mode 100644
index 0000000..e9bf661
--- /dev/null
+++ b/pig/tests/files/papers_url_words.cdx
@@ -0,0 +1,27 @@
+#http://personal.ee.surrey.ac.uk/Personal/R.Bowden/publications/2012/Gilbert_ACCV_2012pp.pdf
+#http://files.eric.ed.gov/fulltext/EJ798626.pdf
+#http://www.hbs.edu/research/pdf/10-108.pdf
+#http://www.unifr.ch/biochem/assets/files/albrecht/publications/Abraham06.pdf
+#http://www.cnbc.cmu.edu/cns/papers/Kassetal2005.pdf
+#http://www.macrothink.org/journal/index.php/ijhrs/article/download/5765/4663
+#http://www.pims.math.ca:80/science/2004/fpsac/Papers/Liskovets.pdf
+#http://www.risc.uni-linz.ac.at/publications/download/risc_3287/synasc_revised.pdf
+#http://softsys.cs.uoi.gr/dbglobe/publications/wi04.pdf
+#http://lexikos.journals.ac.za/pub/article/download/1048/564
+#http://www.siam.org/proceedings/analco/2007/anl07_029ecesaratto.pdf
+#http://www.cs.bris.ac.uk/Publications/Papers/2000249.pdf
+
+# 12 matches:
+
+uk,ac,surrey,ee,personal)/Personal/R.Bowden/publications/2012/Gilbert_ACCV_2012pp.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+gov,ed,eric,files)/fulltext/EJ798626.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+edu,hbs,www)/research/pdf/10-108.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+ch,unifr,www)/biochem/assets/files/albrecht/publications/Abraham06.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+edu,cmu,cnbc,www)/cns/papers/Kassetal2005.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+org,macrothink,www)/journal/index.php/ijhrs/article/download/5765/4663 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+ca,math,pims,www)/science/2004/fpsac/Papers/Liskovets.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+at,ac,uni-linz,risc,www)/publications/download/risc_3287/synasc_revised.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+gr,uoi,cs,softsys)/dbglobe/publications/wi04.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+za,ac,journals,lexikos)/pub/article/download/1048/564 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+org,siam,www)/proceedings/analco/2007/anl07_029ecesaratto.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
+uk,ac,bris,cs,www)/Publications/Papers/2000249.pdf 20170706005950 http://mit.edu/file.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 123 456 CRAWL/CRAWL.warc.gz
diff --git a/pig/tests/test_filter_cdx_paper_pdfs.py b/pig/tests/test_filter_cdx_paper_pdfs.py
new file mode 100644
index 0000000..a8ebd9f
--- /dev/null
+++ b/pig/tests/test_filter_cdx_paper_pdfs.py
@@ -0,0 +1,26 @@
+
+import os
+import unittest
+from pighelper import PigTestHelper
+
+def count_lines(s):
+    return len([l for l in s.strip().split('\n') if len(l) > 0])
+
+class TestFilterCDXPaperPdfs(PigTestHelper):
+
+    def test_papers_domain_words(self):
+        r = self.run_pig("filter-cdx-paper-pdfs.pig", "tests/files/papers_domain_words.cdx")
+        assert count_lines(r) == 4
+
+    def test_papers_edu_tilde(self):
+        r = self.run_pig("filter-cdx-paper-pdfs.pig", "tests/files/papers_edu_tilde.cdx")
+        assert count_lines(r) == 6
+
+    def test_papers_url_doi(self):
+        r = self.run_pig("filter-cdx-paper-pdfs.pig", "tests/files/papers_url_doi.cdx")
+        assert count_lines(r) == 2
+
+    def test_papers_url_words(self):
+        r = self.run_pig("filter-cdx-paper-pdfs.pig", "tests/files/papers_url_words.cdx")
+        assert count_lines(r) == 12
+