Permalink
80743f6 Dec 23, 2014
1 contributor

Users who have contributed to this file

69 lines (56 sloc) 3.42 KB

Brmson-Specific UIMA Intro

We use UIMA (Unstructured Information Management Architecture) as a generic framework for building the NLP pipeline for processing the question and possible answers.

The CAS Data Structure

UIMA works with artifacts (sentence, document, named entity...) that are coupled with featuresets that carry various metadata (features) pertaining the artifacts. Most often, these featuresets are annotations of the artifacts, i.e. covering a specific subset (start - end offset span) of the artifact. All of the above is stored in a container called CAS (Common Analysis Structure), which has a Java interface class JCas.

To recapitulate, CAS is a container that holds the artifact and its various annotations (e.g. token segmentation, dependency parse) and other featuresets (like LAT, answer feature records). The artifact can be present in the CAS in various views or forms of the artifact - at least that was the original idea, but the way we use it, we use the views to carry different artifacts. For example, we have views that hold the Question, SearchResult, Passages, PickedPassages, CandidateAnswer and AnswerHitlist. We use these different views to carry and refer to multiple artifacts at once, e.g. it is quite helpful to be able to refer to Question LATs in CandidateAnswer processing annotators. Each view has its own JCas Java interface.

The various annotations and other feature sets are defined in so-called "type systems". These are ugly XML files that you can find in the src/main/typesystem/ directory. A magic jcasgen tool autogenerates Java classes for them.

The Pipeline Flow

A simple NLP processing pipeline in UIMA is composed from

  • Collection Reader, which generates a CAS for each artifact it loads from somewhere (e.g. the standard input or a TSV file).
  • Analysis Engines (AE) which take the CAS as a paremeter and when processing it turn in turn, they annotate it with various featuresets. Some AEs just invoke a sequence of other AEs: these are Aggregate AEs, while we call those that do the grunt work annotators.
  • CAS Consumer which does something with the annotated CAS, like print it to the output.

In our situation, things are more complex since we work with multiple CASes; the collection reader generates a CAS for the question, but we will then want to create a CAS for each search result, and even each candidate answer. The idea here is that analysis engines will pass each information to each other just within the CASes, and when we juggle multiple CASes, we can use that for scale-out to multiple cluster machines. That's the point of even having CASes.

Therefore, some of our AEs are CAS Multipliers which, for each input CAS (e.g. a question), generate a bunch of output CASes (e.g. search results). They generally still carry the input artifact in a separate view of the output CASes. Then, near the end, the flow reaches a CAS Merger which will combine input CASes in a single output CAS (e.g. AnswerHitlist). Therefore, the pipeline is essentially a directed acyclic graph, with CASes passing on the edges.

The complete flow (except input and output, which can be different for various YodaQA users and interfaces) is encapsulated in a single large aggregate AE which is constructed in:

src/main/java/cz/brmlab/yodaqa/pipeline/YodaQA.java