Engines¶

In RITA what we call engine is a system we will compile rules to, and which will do the heavy lifting after that.

Currently there are three engines:

spaCy¶

Activated by using rita.compile(<rules_file>, use_engine="spacy")

Using this engine, all of the RITA rules will be compiled into spaCy patterns, which can be natively used by spaCy in various scenarios. Most often - to improve NER (Named Entity Recognition), by adding additional entities derived from your given rules

It requires to have spaCy package installed (pip install spacy) and to actually use it later, language model needs to be downloaded (python -m spacy download <language_code>)

Standalone¶

Activated by using rita.compile(<rules_file>, use_engine="standalone"). It compiles into pure regex and can be used with zero dependencies. By default, it uses Python re library. Since 0.5.10 version, you can give a custom regex implementation to use: eg. regex package: rita.compile(<rules_file>, use_engine="standalone", regex_impl=regex)

It is very lightweight, very fast (compared to spaCy), however lacking in some functionality which only proper language model can bring: - Patterns by entity (PERSON, ORGANIZATION, etc) - Patterns by Lemmas - Patterns by POS (Part Of Speech)

Only generic things, like WORD, NUMBER can be matched.

Rust (new in 0.6.0)¶

There's only an interface inside the code, engine itself is proprietary.

In general it's identical to standalone, but differs in one crucial part - all of the rules are compiled into actual binary code and that provides large performance boost. It is proprietary, because there are various caveats, engine itself is a bit more fragile and needs to be tinkered to be optimized to very specific case (eg. few long texts with many matches vs a lot short texts with few matches).