ACL Anthology Network derivatives

aanrelease2013.tar.gz is a mess.

1. Encodings are all over the place.
2. Some data is obviously missing, or present but corrupted.
3. It's unclear which files are the raw data and which are statistics derived from those data.
4. Files that look like they should be the outputs of scripts are actually the help messages of those scripts.
5. Many of the papers_text/ files (plain text extracted from the PDF) have issues:
   • Some are gibberish due to non-compliant PDF encoding (e.g., P00-1032, W06-3709)
   • Others are gibberish to varying degrees, due to bad OCR in the original PDF (from problematic, e.g., T75-2033, to unusable, e.g., J79-1013)
   • Others are not in English (e.g., C73-2029)
   • Others contain the text of a completely different paper (e.g., L08-1302)

Makefile declaratively provides some documentation of the issues and the cleanup work involved.

This repository does not contain any of the original data, only a programmatic description of how to fix it.

To run, call make in the root directory.

Summary statistics

The University of Michigan CLAIR Group's ACL Anthology Network interface reports the following statistics:

Measure	Value
Number of papers	21,212
Number of authors	17,792
Number of venues	342
Number of paper citations	110,975
Number of author collaborations	142,450
Citation network diameter	22
Collaboration network diameter	15

Some of these are inaccurate, or describe only one of the data sources. Different sources in the dataset contain different subsets of the data; for example, citations are reported for some papers that do not have a corresponding papers_text/ file (e.g., L08-1098).

aan/release/2013/acl.txt

Measure	Value
citing→cited relationships	110,930
unique citing papers	16,554
avg. cited per citing	6.7011
unique cited papers	12,840
avg. citing per cited	8.6394
unique papers	18,160
unique papers that both cite and are cited	11,234

Top 10 most-cited papers	# of papers citing	authors	title
J93-2004	928	Mitchell et al.	Building A Large Annotated Corpus Of English: The Penn Treebank Computational Linguistics
P02-1040	891	Papineni et al.	Bleu: A Method For Automatic Evaluation Of Machine Translation
J93-2003	729	Brown et al.	The Mathematics Of Statistical Machine Translation: Parameter Estimation
P03-1021	667	Och & Josef	Minimum Error Rate Training In Statistical Machine Translation
J03-1002	656	Och & Josef	A Systematic Comparison Of Various Statistical Alignment Models
P07-2045	591	Koehn et al.	Moses: Open Source Toolkit for Statistical Machine Translation
N03-1017	556	Koehn et al.	Statistical Phrase-Based Translation
P03-1054	394	Klein & Manning	Accurate Unlexicalized Parsing
J96-1002	376	Berger et al.	A Maximum Entropy Approach To Natural Language Processing
A00-2018	371	Charniak	A Maximum-Entropy-Inspired Parser

Top 10 most-citing papers	# of papers cited
P10-1142	88
J10-3003	80
W13-4917	71
W13-2201	65
J12-1006	62
J98-1001	59
J13-2003	59
J07-4004	57
J11-2002	52
D11-1108	52

aan/release/2013/acl-metadata.txt

The formatting of this file is, frankly, befuddling. The general structure is BibTeX-esque, but no BibTeX parser could possibly handle it. Worse, the mixture of encodings is insane! If ftfy was ever looking for a great real-world case study, this would be it.

• There are 20,989 papers.
• There is 1 missing author, W10-4238, and 16,308 unique author sequences (author lists all authors for that paper).
• There are around 16,786 unique author names (approximate due to inconsistencies in both the character encoding and the metadata's representation of the names).
• There are 12 missing venues, and 494 unique venue values.
• The year field ranges from 1965 to 2013, with 41 unique values (there are some gaps prior to 1978).
• There are 18,152 papers (all but 8) in the citation network that have an entry in this metadata file
• Of the 110,930 total edges in the citation network, 110,889 have entries in this metadata file.

aan/papers_text/???-????.txt

There are a lot of other files in this directory; some of the papers are segmented into body and references sections; there are some files that seem like they were intended to go in aan/release/2013/; and many of the files that match this pattern are empty.

• There are 20,194 files that match this pattern and are non-empty.
• There are 17,593 papers in the citation network that have a corresponding file in papers_text/.
• Of the 110,930 total edges in the citation network, 107,873 have corresponding files in papers_text/.

Credits

Despite these flaws, the ACL Anthology Network is a great resource; many thanks to the many contributors.

Dragomir R. Radev, Pradeep Muthukrishnan, Vahed Qazvinian, Amjad Abu-Jbara. 2013. The ACL Anthology Network Corpus. Language Resources and Evaluation 47 (4), pp. 919–944. 10.1007/s10579-012-9211-2.

License

Copyright 2016–2018 Christopher Brown. MIT Licensed.