How to change Solr standard tokenizer rules?27 Aug 2017
Although Solr comes with standard tokenizer implementation, which is well prepared to tokenize most of the texts, there are cases when it is helpless. Imagine a document with many numbers, of which many are followed by percentage sign. In a certain contexts it is expected to distinguish queries that refer to those percentages & plain numbers. How to achieve that? We need a custom tokenizer.
Another example may be connected with payloads, that is additional data that can be stored with a token. In a basic Solr convention they are served as token
|number, which will be split into two tokens, loosing connection between them. Both cases would be addressed in these post.
There may be multiple ways to define rules how text should be tokenized, but the one used in Lucene/Solr is based on Unicode Text Segmentation algorithm. As stated in StandardTokenizer docs:
Rules alone are not something you want to analyse if you don’t have to, so let’s look at the most interesting part of a documentation. It’s easy to spot that values from these table can be mapped to variables mentioned in StandardTokenizerImpl.jflex. For example Hebrew_Letter can be found as a variable HebrewLetter and as a part of a rule:
Knowing that we can start implementing changes that lead to having percentages & payloads kept in token stream next to corresponding base tokens. As we don’t want to seriously change long-lasting rules, for payloads we’ve added a new one, which can be translated as: do not split token when after a word there is a vertical bar followed by a number. It looks like this:
The second snippet is for percentages and requires a small change in existing rule:
Now the new tokenizer has to be built and tested. Exact configuration for that can be found on my Github, but in short only mvn compile is required to see how generated PayloadStandardTokenizerImpl looks like.
For the testing part I’ve added a schema field that during indexing is using newly created tokenizer, and during querying - the standard one. Of course proper factory has be defined and new classes added to Solr as a library.
For testing purposes I’ve prepared a query: The population of Gdańsk
|100 accounts for 0.012% of the population of Poland. Gdańsk has added payload and there is percentage next to a number. By the way: why would we like to add payload to Gdańsk? For example to strengthen it during search, because others would have value of 1. I’ll write about that soon, but for now let’s go back to the results of experiment:
As you can see the indexing part (the one we’ve updated) keeps payloads & percentages, while during querying (standard tokenizer used) percentage and vertical bar are removed and values are kept separately. There is nothing wrong about both behaviours - everything depends on a context of search.