Drew, Tom and I are feverishly working away on finishing up Taming Text.  We are currently in the process of addressing the feedback we got from our final review and should have updates up soon.  I have also posted all of the book’s source code up on Github under the Taming Text user.  The source includes, amongst other things, a simple Question Answering system using Solr and OpenNLP, as well as analyzers for Lucene that use OpenNLP for sentence detection, part of speech tagging and Named Entity Recognition.  As with most books, these examples are meant to be just that, examples.

First off, I was doing a fair amount of work calculating document similarities across whole collections using, at first, the RowSimilarityJob and later a map-side simplification I wrote that uses the distributed cache called the VectorDistanceSimilarityJob.  Both of these come in handy when one wants to calculate pairwise-similarity between all (or most) items in a collection.  The original Mahout implementation was focused on providing recommendations, but as outlined in the Elsayed, Lin and Oard paper, it is quite useful for text as well in cases where one wants to precompute “more like this” for all documents.  As for the need for two similar approaches, see the discussion at http://www.lucidimagination.com/search/document/40c4f124795c6b5/rowsimilarity_s#42ab816c27c6a9e7.  In essence, it boils down to I didn’t need a fully generic implementation that was a bit slower on larger matrices since I mainly wanted to compare all my vectors in HDFS against a subset of “core” vectors that fit into memory.  That being said, Sebastian is already hard at work on making the more generic version perform better when certain distance measures are used while still offering the full suite of capabilities of the existing RowSimilarityJob.  See MAHOUT-767 for more info on that work.

Now, I’m looking into some more pruning techniques via MAHOUT-688.  After that quick patch, I think I’m going to dig in a bit more to recommendations as well as run some tests on the ASF mail archives I posted a while back (see below for an update).

Also, I’ve switched to using Git and Github for managing my Mahout changes (as well as other work), so if you want to see what I’m up to, check out my Github account.

It’s not complete yet, but the ASF Public Mail archive I put up last September on Amazon AWS is getting a fresh new version.  The interim solution is available at https://s3.amazonaws.com/asf-mail-archives-7-18-2011/index.html, but look for it to be a Public Data Set hosted by Amazon soon.  The September version of this data contained roughly 6.7M emails sent to the public mailing lists at the Apache Software Foundation, so I suspect this version has somewhere in the 7M+ item range, but I haven’t counted them.  At any rate, I hope it is useful to people.

Finally, on a personal note, I’m back at Lucid Imagination after a brief move elsewhere, this time in a new role as Chief Scientist.  Lucid is a company I co-founded and helped build up for the past 4 years.  I’m looking forward to be back working closely with Lucene and Solr again and a top notch technical team.  I’m also looking forward to working on Mahout more, as well as other technologies like Hadoop, Pig, HBase and the like, especially as they relate to search and recommendations.

I really enjoyed the meetup.  Lots of good questions and insights into machine learning.  For those at the meeting who were asking about references, check out Mahout’s references page, especially the Background Material section.

Deploying a massively scalable recommender system with Apache Mahout | “I for one welcome our new computer overlords”.

Lucid Imagination » Apache Lucene 3.1.0 and Apache Solr 3.1.0.

Thanks to everyone for all their hard work in making these releases happen.

Regardless of the hype, I had a good time.  Obviously I’m biased, but I think the talk RC Johnson and I gave on using Solr in NoSQL situations went well, was well attended and well received.  RC and I split duties during the talk.  I covered the big picture around Solr as a NoSQL solution and RC covered how Bazaarvoice uses Solr to serve billions of searches/lookups per month.  A good chunk of the audience was already familiar with NoSQL and Lucene/Solr, so it was a pretty technical audience, which is exactly what we hoped for.   We also had some good questions during the talk as well as afterward.  The slides are up on SlideRocket.