Solr Spellchecker internals (now with tests!)

Let’s talk about spellcheckers. A spellchecker, as you may know, is that device that tells you whether you misspelled or not a word, and makes you some suggestions. One of the first spellcheckers that i remember seeing is the MS Word spellchecker. One could say that MS Word defined the “standard interface” of word processors spellcheckers: you misspelled a word, it was underlined with a zig zag style line in red, and if you right clicked on it, a list of suggested words appeared. I have seen this interface in many different software, for example, google docs.

Another more modern example is the famous “Did you mean” from Google. You type some words like “britny spears” and google would suggest “Britney Spears”. It appears that a lot of people have issues spelling Britney Spears. But Google is different. As usual, they use artificial intelligence algorithms to suggest misspelled words. Google algorithms are the closest you’ll get to magic in computer engineering.
But today I’m going to talk about Solr SpellChecker. In contrast with from google, Solr spellcheker isn’t much more than a pattern similarity algorithm. You give it a word and it will find similar words. But what is interpreted as “similar” by Solr? The words are interpreted just as an array of characters, so, two words are similar if they have many coincidences in their character sequences. That may sound obvious, but in natural languages the bytes (letters) have little meaning. It is the entire word that has a meaning. So, Solr algorithms won’t even know that you are giving them words. Those byte sequences could be sequences of numbers, or sequences of colors. Solr will find the sequences of numbers that have small differences with the input, or the sequences of colors, etc. By the way, this is not the approach that Google follows. Google knows the frequent words, the frequent misspelled words, and the frequent way humans make mistakes. It is my intention to talk about these interesting topics in a next post, but now let’s study how solr spellchecker works in detail, and then make some tests.

Solr spellchecker follows the same strategy as many other spellcheckers. It has a dictionary of correct spelled terms (correct by definition, if there is a misspelled word in the dictionary that will pass as a correct word). When somebody asks for suggestions to a word, Solr Spellchecker first obtains a list of candidate words and then ranks those candidates according to some criteria. The first step is accomplished by ngrams. A ngram is a substring in a string. For example if you have the word ‚house‘, some ngrams would be ‚hou‘, ‚ous‘, ’se‘ (there are many other ngrams of differents lenghts, i’ve shown you only three of them). Two similar words will have many matching ngrams: ‚mouse‘ also has ‚ous‘ and ’se‘ but not ‚hou‘. What Solr does is create a Lucene index of the words in the dictionary and filter them with ngram filters. So when you ask for suggestions for “house”, Solr searches ‚ho‘ OR ‚ou‘ OR ‚us‘ OR ’se‘ OR ‚hou‘ OR ‚ous‘ OR ‚use‘ OR ‚hous‘ OR ‚ouse‘ and, because Solr ranks boolean querys by the document with more coincidences, what you’ll get is a list of some similar words from our dictionary.

How does Solr rank the words afterwards? There is something called edit distance that tells us how many operations we have to perform in order to transform a word into another. By operations we understand insertions, deletions, or modifications of single characters. There are many algorithms to find the edit distance, one is Levenshtein (that is the default algorithm used in Solr). These algorithms are computationally complex, and that’s the reason why Solr doesn’t use them as the first choice in selecting the suggestions from among all the words in the dictionary. The dictionary is reduced and then this “difficult” ranking process is performed.
Perhaps, now you understand what I meant by “solr spellchecker only find similar byte arrays”. You never introduce information about our natural language into the algorithm and the only thing you provide is “a set of byte secuences” (ie a dictionary)

So far, so well. Does this approach work? Yes. Could it work better? Of course! And we have a lot of things that we can do to improve the algorithm. But first, let’s try to make this look scientific (if you remember, that was the idea of internet in the first place…) We need tests to see where we are standing. Something that I find boring is moving from the theoretical side to the experimental side. But that is a must in this that we call research. So, next I present a series of different tests that I performed on a Solr instance (that we have for experimental purposes) of the wikipedia (I recommend reading this post about how we indexed wikipedia, for all of you trying to index huge amounts of text)
I created a dictionary using the words from wikipedia and then tested a lot of different misspelled words taken from different origins.