Overview of NE Contexts (probable/possible, family history of, or history of) and (negated)
As of cTAKES 2.5, NE Contexts is no longer a part of the default pipeline. Instead the newer assertion annotator is used to identify general attributes of EntityMentions (but not attributes specific to a single type of named entity such as the dosage attribute of a medication)
The context annotator provides a mechanism for examining the context of existing annotations, finding events of interest in the context, and acting on those events in some way. The negation and status annotators both take advantage of this infrastructure by examining the context of named entities (e.g. disorders and findings) to see if they should be considered as negated (e.g. "no chest pain") or if their status should be modified (e.g. "myocardial infarction" should have status "history of").
In fact, the "negation annotator" is really just the context annotator configured to deal with negations. Similarly, the "status annotator" is the context annotator configured to identify the status of named entities.
To better understand the context annotator code you should start by reading the Javadocs for the class edu.mayo.bmi.uima.context.ContextAnnotator.java. It provides a conceptual overview of how the code works.
The negation detection annotator is a pattern-based (no MaxEnt models required/used) approach that uses finite state machines and is roughly based on the popular NegEX algorithm introduced by Wendy Chapman (University of Pittsburgh)
What follows is an explanation of how negation is performed using the context annotator.
We will start by examining the descriptor file desc/NegationAnnotator.xml. It calls edu.mayo.bmi.uima.context.ContextAnnotator, instead of a "NegationAnnotator". In fact, there is no "negation annotator" analysis engine. We simply configure the ContextAnnotator for the task. Next we will discuss each of the parameter settings in turn:
- MaxLeftScopeSize = 10 - The maximum number of annotations that will make up the left-hand side context is ten. Increase or decrease this parameter setting to increase or decrease the left hand side context.
- MaxRightScopeSize = 10 - The maximum number of annotations that will make up the right-hand side context is ten. Increase or decrease this parameter setting to increase or decrease the right-hand side context.
- ScopeOrder = LEFT, RIGHT - The context annotator will look for signs of negation on the left-hand side of a named entity first and then the right-hand side.
- ContextAnalyzerClass = edu.mayo.bmi.uima.context.negation.NegationContextAnalyzer - The context analyzer looks at the context (e.g. the 10 words on the left or right of the named entity) and determines if the named entity should be negated. If it should, then the negation context analyzer will generate a context hit to be consumed by the context hit consumer (see ContextHitConsumerClass.)
- ContextHitConsumerClass = edu.mayo.bmi.uima.context.negation.NegationContextHitConsumer - The context hit consumer handles context hits generated by the context analyzer. In this case, the negation context hit consumer simply sets the certainty of a named entity to -1 which indicates that it has been negated.
- WindowAnnotationClass = edu.mayo.bmi.uima.core.type.textspan.Sentence - When the context annotator collects the context annotations for a named entity, it will not look beyond the boundaries of the sentence that the named entity is found in.
- FocusAnnotationClass = edu.mayo.bmi.uima.core.type.textsem.IdentifiedAnnotation - The negation annotator is concerned with negating named entities and thus the focus annotation type (the annotations for which a context is generated and examined) specifies the parent of EventMention and of EntityMention classes.
- ContextAnnotationClass = edu.mayo.bmi.uima.common.type.BaseToken - The context of the named entities is a list of tokens. This is what the context analyzer is going to examine.
So, the work of negating a named entity is done by:
- finding negations by the NegationContextAnalyzer
- updating the status of NamedEntities by the NegationContextHitConsumer
The former is a pretty lightweight wrapper around another class which has all of the negation pattern finding logic, edu.mayo.bmi.fsm.machine.NegationFSM. If you want to update the pattern matching of negation detection, then you would do it in that class.
Updating Negex patterns
Updating the negation detection patterns will involve either:
- trial and error experimentation. or
- understanding how the NegationFSM works
The rules, patterns, words that identify negation are hard-coded into the class edu.mayo.bmi.fsm.machine.NegationFSM which is found in the core project. We would suggest starting off with the trial-and-error approach. For example, if you wanted to add "impossible" to the lexicon of negation words, then you could try adding it to the _negAdjectivesSet and test the behavior.
The way the status annotator works mirrors very closely how the negation annotator works. You are encouraged to read the above section, "Negation annotator", examine the parameter settings given for desc/StatusAnnotator.xml, and look at edu.mayo.bmi.fsm.machine.StatusIndicatorFSM.