To generate a word2vec model, but using multi-word keywords instead of single words.

Finding similar keywords for "obesity"

The idea started in the Epistemonikos database (epistemonikos.org), a database of scientific articles in health. Because normally in health/medicine the language used is complex. You can easily find keywords like:

• asthma
• heart failure
• medial compartment knee osteoarthritis
• preserved left ventricular systolic function
• non-selective non-steroidal anti-inflammatory drugs

We tried some approaches to find those keywords, like ngrams, ngrams + tf-idf, identify entities, among others. But we didn't get really good results.

We found that tokenizing using stopwords + non word characters was really useful for "finding" the keywords. An example:

• input: "Timing of replacement therapy for acute renal failure after cardiac surgery"
• output: [ "timing", "replacement therapy", "acute renal failure", "cardiac surgery" ]

So we basically split the text when we find:

• a stopword
• a non word character(/,!?. etc) (except from - and ')

That's it.

But as there were some problem with some keywords that cointain stopwords, like:

• Vitamin A
• Hepatitis A
• Web of Science

So we decided to add another method (nltk with some grammar definition) to cover most of the cases.

Seem to be an old idea (2004):

Mihalcea, Rada, and Paul Tarau. "Textrank: Bringing order into text." Proceedings of the 2004 conference on empirical methods in natural language processing. 2004.

Reading an implementation of textrank, I realize they used stopwords to separate and create the graph. Then I though in using it as tokenizer for word2vec

As pointed by @deliprao in this twitter thread. It's also used by Rake (2010):

Rose, Stuart & Engel, Dave & Cramer, Nick & Cowley, Wendy. (2010). Automatic Keyword Extraction from Individual Documents. 10.1002/9780470689646.ch1.

As noted by @astent in the Twitter thread, this concept is called chinking (chunking by exclusion) https://www.nltk.org/book/ch07.html#Chinking

We worked in an implementation, that could be used in multiple languages. Of course not all languages are sutable for using this approach. We have tried with good results in English, Spanish and Portuguese (you need the stopwords for each language)

For more detail, take a look to the tokenizer here: keywords_tokenizer.py

You can try it here (takes time to load, lowercase only, doesn't work in mobile yet) MPV :)

These embedding were created using 827,341 title/abstract from @epistemonikos database. With keywords that repeat at least 10 times. The total vocab is 349,080 keywords (really manageable number)

(Need update after including nltk grammar method)

One of the main benefit of this method, is the size of the vocabulary. For example, using keywords that repeat at least 10 times, for the Epistemonikos dataset (827,341 title/abstract), we got the following vocab size:

More information regarding the comparison, take a look to the folder analyze.

mkdir -p data/inputs
wget "http://s3.amazonaws.com/episte-labs/episte_title_abstract.tsv.gz" -O data/inputs/episte_title_abstract.tsv.gz

pip install -r requirements.txt
python3 -m nltk.downloader averaged_perceptron_tagger
python3 -m nltk.downloader punkt

Let's first use only 30,000 references to train the embeddings:

mkdir -p /tmp/example
python keywords2vec.py -i data/inputs/episte_title_abstract.tsv.gz -o /tmp/example --sample=30000

Step1: Tokenizing

Step2: Reading keywords

Step3: Calculate frequency

Step4: Generate word2vec model

Step5: Save vectors

Step6: Save counter

python try_keywords2vec.py -v /tmp/example/word2vec.vec -c /tmp/example/keywords_counter.tsv.gz -p"obesity" -t 25

Finding similar keywords for "obesity"

You can start a server to play around with the data:

FLASK_APP=server/server.py VECTORS_PATH=/tmp/example/ flask run