Metadoc is a post-truth era news article metadata retrieval service. It does social media activity lookup, source authenticity rating, checksum creation, json-ld and metatag parsing as well as information extraction for named entities, pullquotes, fulltext and other useful things based off of arbitrary article URLs. Also, Metadoc is built to be relatively fast.
You just throw it any news article URL, and Metadoc will yield.
from metadoc import Metadoc
url = "https://theintercept.com/2016/11/17/iphones-secretly-send-call-history-to-apple-security-firm-says"
metadoc = Metadoc(url=url).query_all()
metadoc.return_ball()
=>
{
"title": "iPhones Secretly Send Call History to Apple, Security Firm Says",
"url": "https://theintercept.com/2016/11/17/iphones-secretly-send-call-history-to-apple-security-firm-says/",
"domain": {
"date_registered": "2009-10-01T00:00:00",
"credibility": {
"is_blacklisted": false,
"fake_confidence": "0.00"
},
"name": "theintercept.com",
"favicon": "https://logo.clearbit.com/theintercept.com?size=200"
},
"text": {
"contenthash": "517b04163d95c264dfffb9797b824026",
"fulltext": "Apple emerged as a guardian of user privacy this year [...]",
"pullquotes": [
"Anyone else who might be able to obtain the user’s iCloud credentials, like hackers [...]",
],
"reading_time": 442
},
"entities": {
"keywords": {
"apple": 1.020775623268698,
"data": 1.0173130193905817,
"logs": 1.0235457063711912,
"icloud": 1.0242382271468145
},
"names": [
"San Bernardino",
"Syed Rizwan",
"Vladimir Katalov",
"Apple CallKit",
"Chris Soghoian",
"Jonathan Zdziarski"
]
},
"image": "https://prod01-cdn04.cdn.firstlook.org/wp-uploads[...]",
"authors": [
"Kim Zetter"
],
"language": "en",
"social": [{
"metrics": [
{
"label": "sharecount",
"count": 5994
}
],
"provider": "facebook"
}]
}
Since it's painful watching humans be all aggrevated about those fake news making the rounds, Metadoc does a rather crude background check on article sources. This means compiling a simple blacklist-lookup and whois
data on the domain. Blacklists taken into account include the widely discredited PropOrNot. Thus, only if a domain is found on every blacklist do we spit out a fake_confidence
of 1. The resulting metadata should be taken with lots of salt. To put it in the words of Don DeLillo – "lists are a form of cultural hysteria."
For speed and simplicity, I decided against nltk
and instead rely on the Averaged Perceptron as imagined by Matthew Honnibal @explosion. The pip install comes pre-trained with a CoNLL 2000 training set which works reasonably well to detect proper nouns. Since training is non-deterministic, unwanted stopwords might slip through. If you want to try out other datasets, simply replace metadoc/extract/data/training_set.txt
with your own and run metadoc.extract.pos.do_train
.
This humble library is used in the context of a news-related software undertaking. We're synthesizing what we dub "audience-evaluated content" with automated metadata. Machine-in-the-loop, if you will. Sounds like bullshit, but it's totally not. If you're intrigued and might want to work with us, feel free to drop a line to p at psolbach dot com.
Requires python 3.5.
pip install metadoc
curl -L git.io/v1Muh | python3
brew install python3 libxml2 libxslt libtiff libjpeg webp little-cms2
apt-get install -y python3 libxml2-dev libxslt-dev libtiff-dev libjpeg-dev webp whois
dnf install libxml2-devel libxslt-devel libtiff-devel libjpeg-devel libjpeg-turbo-devel libwebp whois
pip3 install -r requirements-dev.txt
python metadoc/__install__.py
python serve.py => serving @ 6060
- Page concatenation is needed in order to properly calculate wordcount and reading time.
- Authenticity heuristic with sharecount deviance detection (requires state).
Perf: Worst offender is nltk's pos tagger. Roll own w/ Average Perceptron.Newspaper's summarize produces pullquotes, fulltext takes a while. Move to libextract?
Metadoc stems from a pedigree of nice libraries like libextract, langdetect and nltk.
Metadoc leans on this perceptron implementation inspired by Matthew Honnibal.
Metadoc is work-in-progress and maintained by @___paul