Exploring citation parsing using Conditional Random Fields (CRF). Heavily influenced by ParsCit and AnyStyle. My main goal here is to get something simple working as a starting point for learning more about CRF. Nothing here is state of the art, for that see, e.g.:
- Synthetic vs. Real Reference Strings for Citation Parsing, and the Importance of Re-training and Out-Of-Sample Data for Meaningful Evaluations: Experiments with GROBID, GIANT and Cora
- GIANT: The 1-Billion Annotated Synthetic Bibliographic-Reference-String Dataset for Deep Citation Parsing
- Neural-ParsCit
editor.html is a simple HTML editor inspired by MarsEdit Live Source Preview where you can edit XML and see a live preview.
data/core.xml is the training data from AnyStyle (1510 references).
dict.php uses the dictionary that comes with ParsCit.
For background see Conditional random fields. I use CRF++: Yet Another CRF toolkit, which is also used in ParsCit.
To get CRF++ to work on Heroku we need to compile the executable. For background on this see How to run an executable on Heroku from node, works locally and C++ buildpack.
I forked felkr/heroku-buildpack-cpp and added it as a buildpack to my Heroku app (under the Settings tab). I put the source code for CRF++ into the root folder of the app (which makes things messy) then when the app is deployed CRF++ is compiled. Note that I tried simply logging in to the Heroku app:
heroku run bash -a citation-parser
and compiling the code. This failed with g++ errors:
configure: error: Your compiler is not powerful enough to compile CRF++.
Turns out that g++ is only available at build time, hence to use g++ I need a buildpack.
The buildpack compiled the code, but when I logged into the shell the executable wouldn’t run:
heroku run bash -a citation-parser
./crf_learn
/app/.libs/crf_learn: error while loading shared libraries: libcrfpp.so.0: cannot open shared object file: No such file or directory
For whatever reason the executable is looking for a shared library which doesn’t exist. To fix this I edited the buildpack compile script to set the "LDFLAGS=--static" --disable-shared flags for configure. This then compiled an executable that worked.
I’ve now updated the buildpack to use the src folder so that this repo is much tidier.
CRF++ didn’t want to build using autotools, but it is available on Homebrew so we can just brew install crf++ to get a working version which is installed in /opt/homebrew/bin.
To train model we need some data that has been marked up. I follow AnyStyle’s XML, e.g.:
<?xml version="1.0" encoding="UTF-8"?>
<dataset>
<sequence>
<author>Heidegger M.,</author>
<date>1927,</date>
<title>Être et temps,</title>
<editor>Gallimard, Ed.</editor>
<date>1986,</date>
<location>Paris.</location>
</sequence>
.
.
.
</dataset>
We need to convert this into the format expected by CRF, which is one token per line, with features following, and then the tag indicating what part of the sequence this token belongs to.
php parse_train.php data/core.xml parses the training XML and outputs a .train file with the features and tags. Having converted the training data we now build the model using crf_learn in the CRF++ package:
crf_learn data/parsCit.template data/core.train core.model
.
.
.
Done!1065.81 s
Note the template file data/parsCit.template which tells CRF++ how to process the features, see Preparing feature templates.
To use the model we need to take some data and convert it into the training format. refs_to_train.php reads a text file with one reference string per line and outputs XML with each line enclosed in a <title> tag. This file can then be processed as if it were training data.
php refs_to_train.php refs.txt
php parse_train.php refs.src.xml
Now we use our model to process the data using crf_test. In this case crf_test takes the data (each reference tagged with <title>) and outputs the tags based on the model. These tags are the ones we use to extracted the structured data.
crf_test -m core.model refs.src.train > out.train
We then convert the output (trained format) to XML, and we then can convert the XML to a “native” format (e.g., RIS for bibliographic data).
php parse_results_to_xml.php out.train > out.xml
php parse_results_to_native.php out.xml
Need to think about how to post process tags, and how to handle cases like this where a date has been inserted in the title so that we with the initial model we end up with two dates and titles:
<author>Aguilar, C., K. Siu-Ting, and P. J. Venegas.</author>
<date>2007.</date>
<title>The rheophilous tadpole of Telmatobius atahualpai Wiens,</title>
<date>1993</date>
<title>(Anura: Ceratophryidae).</title>
<journal>South American Journal of Herpetology</journal>
<volume>2:</volume>
<pages>165–174.</pages>
Take a RIS file and output Anystyle XML format:
php ris_to_training.php nsp.ris > nsp.xml
Convert to training format:
php parse_train.php nsp.xml
Add the output to core.train and then rebuild model:
crf_learn data/parsCit.template core.train core.model
Do this with each new set of training data so that we build a better model (we hope).
As above we can add items that we know have failed to our training set. Just follow the steps:
php parse_train.php fail.xml
Add the output to core.train and then rebuild model:
crf_learn data/parsCit.template core.train core.model
It is a good idea to then rename the fail.xml as it is now in the model, and start a new, empty fail.xml to collect new failures. If we keep doing this iteratively the model should improve.
Take some references marked up in XML and generate training format.
php parse_train.php fail.xml
Run crf_test to get tags from model
crf_test -m core.model fail.train > f.train
Output from crf_test has original tags and ones from model, so compare those
php parse_results_to_test.php f.train
Hogg, H.R. (1896). Araneidae. In B. Spencer (ed.) Report of the Horn Expedition to Central Australia. Pt. 2. Zoology. pp. 309-356. Melville, Mullen and Slade, Melbourne.