Article recommendation system for Pelican based on post similarity calculated using NLTK and scikit-learn's TFIDF vectorizer.
You will need to have numpy, scikit-learn and nltk installed. You
also need to have the necessary nltk corpus installed. Run >>>
nltk.download() to download if necessary.
When the plug-in is activated, you can use article.recommended_articles
list to get a list of recommended articles. You can then use this inside your
templates. For example:
{% if article.recommended_articles %}
<div class="related-posts">
<h1>Recommended Articles</h1>
<ul>
{% for post in article.recommended_articles %}
<li>
<a href="{{ SITEURL }}/{{ post.url }}"
rel="bookmark"
title="Permalink to {{ post.title|striptags }}">
<span itemprop="name">
{{ post.title }}
</span>
</a>
</li>
{% endfor %}
</ul>
</div>
{% endif %}A dictionary ARTICLE_RECOMMENDER_OPTIONS can be defined in Pelican's
settings file. Currently the following two keys from this dictionary are used:
max_recommendations(default = 5): the maximum number of articles recommended.min_score(default = 0.05): minimum score to be included as a recommendation. Value must be between 0 and 1.
Example:
ARTICLE_RECOMMENDER_OPTIONS = {'max_recommendations': 5,
'min_score': 0.1}The recommender will parse the content of all articles, pass the plain text (ignoring code and math blocks, for example) to an NLTK tokenizer, and then vectorizes all the articles using scikit-learn's TFIDF vectorizer. Finally, a similarity score is calculated based on cosine similarity.
Because of all the above, the plug-in will make the generation process
significantly slower. Please keep this in mind. You might want to only enable
this plug-in in publishconf.py.
This plug-in started off as modifications to the related_posts plug-in
available on Pelican's Plugin Repository, and diverged enough for me to
release it as an independent plug-in.