Python or Java for text processing (text mining, information retrieval, natural language processing)

Question

I'm soon to start on a new project where I am going to do lots of text processing tasks like searching, categorization/classifying, clustering, and so on.

There's going to be a huge amount of documents that need to be processed; probably millions of documents. After the initial processing, it also has to be able to be updated daily with multiple new documents.

Can I use Python to do this, or is Python too slow? Is it best to use Java?

If possible, I would prefer Python since that's what I have been using lately. Plus, I would finish the coding part much faster. But it all depends on Python's speed. I have used Python for some small scale text processing tasks with only a couple of thousand documents, but I am not sure how well it scales up.

Solution 1

Both are good. Java has a lot of steam going into text processing. Stanford's text processing system, OpenNLP, UIMA, and GATE seem to be the big players (I know I am missing some). You can literally run the StanfordNLP module on a large corpus after a few minutes of playing with it. But, it has major memory requirements (3 GB or so when I was using it).

NLTK, Gensim, Pattern, and many other Python modules are very good at text processing. Their memory usage and performance are very reasonable.

Python scales up because text processing is a very easily scalable problem. You can use multiprocessing very easily when parsing/tagging/chunking/extracting documents. Once your get your text into any sort of feature vector, then you can use numpy arrays, and we all know how great numpy is...

I learned with NLTK, and Python has helped me greatly in reducing development time, so I opine that you give that a shot first. They have a very helpful mailing list as well, which I suggest you join.

If you have custom scripts, you might want to check out how well they perform with PyPy.

Solution 2

It's very difficult to answer questions like this without trying. So why don't you

1. Figure out what would be a difficult operation
2. Implement that (and I mean the simplest, quickest hack that you can make work)
3. Run it with a lot of data, and see how long it takes
4. Figure out if it's too slow

I've done this in the past and it's really the way to see if something performs well enough for something.

Solution 3

Just write it, the biggest flaw in programming people have is premature optimization. Work on a project, write it out and get it working. Then go back and fix the bugs and ensure that its optimized. There are going to be a number of people harping on about speed of x vs y and y is better than x but at the end of a day its just a language. Its not what a language is but how it does it.

Solution 4

it's not language you have to evaluate, but frameworks and app servers for clustering, data storage/retrieval etc available for the language.

you can use jython and use all the java enterprise technologies for high load system and do text parsing with python.

View more solutions

Comments

[kga]

I have never used Jython. I read that it is slower than Python. But I guess it might be made faster than Python by converting critical parts of code into java? is that correct?

[Denis Tulskiy]

@user757256: yes, jython is slightly slower and is much more memory-hungry. Hack some prototype, benchmark with CPython and Jython, then see what you can optimize. And again, make your choice only after evaluating which libraries you can use.

[Thien]

I think this is a good overview of packages used for the NLP side of a project. Another thing to consider is the machine learning side. Though I am only familiar with the Java libraries - WEKA, MALLET, Apache Mahout.