Text Extraction using Natural Language Processing
Administrator - 18 June 2010 - 12min
Administrator - 18 June 2010 - 12min
A few months ago I was asked to look into an email processing problem. We needed to extract event related information from consumer-originated email. As a traditional programmer the first instinct was to think in terms of regular expressions and lookup tables! Experience quickly tempered that thought and I decided to look at Natural Language Processing.
There were several standard methodologies in place for natural language processing tasks and quite a few open source tools were available. The jargon was daunting: corpuses, entities, gazetteers, POS tags, transducers, and JAPE were just a few terms that I had to wade through. The thought of the alternative: debugging code with zillions of unreadable regular expressions kept me going!
I downloaded GATE and was able to quickly build a prototype parsing emails to get to our target data. GATE breaks down the task of processing text into small specialized chunks of work strung together in a “pipeline”. The tasks work by putting XML annotations in the text or enhancing/using the annotations put by a previous task. It is a simple and beautiful architecture living up to its acronym: General Architecture for Text Engineering.
Each task is called a Processing Resource (PR) in GATE. You can choose from a host of preinstalled resources, or find and install PRs from the internet or just go ahead and write your own. Let us look at a simple GATE pipeline for text processing.
The first PR in pipeline is a tokenizer: this takes the email text and converts it into a series of tokens like numbers, upper-case strings, space or punctuation, etc. The second PR splits the text into sentences based on space and punctuation tokens
We then have a Parts Of Speech (POS) tagger: it understands sentence grammar and breaks the sentence into nouns, verbs, adjectives, pronouns etc.
A gazetteer is another useful Processing Resource which marks the text which matches your lookup tables. Take a list of colleges for example. If one of these colleges appears in the text then it gets annotated as a College.
We are almost there! The last stage is the scary sounding JAPE transducer. This is nothing but a way of defining regular expressions over the GATE annotations using a rule based language. But didn’t we switch to NLP to avoid regular expressions?
JAPE is a very different beast as compared to standard regular expressions.
– It works on the annotations added by the pipeline which capture grammar and lookups instead of raw text strings.
– JAPE rules are applied in a declarative manner. Regular expressions are sequential and in many occasions the order in which they are applied affect the result.
JAPE is bit difficult to understand however the accuracy, stability and maintainability offered by the GATE pipeline are far better than using traditional programming approaches.
There are several features of NLP that make it an art rather than a science. For each type of processing task there are several different types of PRs that you can choose from. For example we found that people use a lot of abbreviations in email and regularly leave out full stops at the end of sentences. A standard sentence splitter fails in such cases. We turned to the RegEx sentence splitter where we were able to enhance the logic used by defining our own regular expressions to detect or ignore such cases.
In addition, the order of tasks in the pipeline can make a big difference to the accuracy. Moving the gazetteer up the chain and using its annotations in sentence splitting helps resolve problems where the PR might split the sentence where abbreviations like U.S.A. are used (the full stop at the end of A and the space following it causes a line break in a usage like U.S.A Today).
The Java interface to GATE is simple. Once you are happy with the pipeline, from the IDE you
– Save it as a .gapp file in the GATE IDE.
– Load the gapp file (in Java), load the documents to process into a collection (the “corpus)
– Execute the pipeline.
For each document you get an annotated XML file which you parse using a standard XML parser to look for the tags your application is interested in.
A major complexity that I have avoided discussing until now is performance. Look forward to the next post to know more!!