[
https://issues.apache.org/jira/browse/UIMA-2106?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Nicolas Hernandez updated UIMA-2106:
------------------------------------
Description:
The HMMTagger Analysis Engine class uses the
org.apache.uima.examples.tagger.Viterbi.java implementation to determine the
pos tag list of a given sentence.
In practice this implementation is partially dependant on the part of speech
tagging (likewise the remaining HMMTagger classes actually).
For exemple it makes strong assumptions on the kind of tokens it can take as
input. It assumes no restriction about the token covertext values.
It results in using some covertext probabilities for initialization or default
value when the tagger processes an unknown token...
As a consequence if the coveredText used for setting the default value is not
present in the training model an error occurs. Roughly speaking, the process
looks first for probability associated to the current token coverText, if the
coverText is not present in the model, it looks in the model for the
probability of its longest suffix, and finally if it does not found a match,
the process assigns to the unknown coverText the probability of the arbitrary
coverText : "("
The problem is that if the probability of this coverText is not available in
the model, the probability of the unknown token is not defined and a null
pointer exception occurs latter when the variable is called.
Why the probability of the "(" text would not be available in the model ? In a
large training corpus if we consider all the tokens, there is little chance not
to find at least one occurrence of "(".
Nevertheless if we use the HMM training AE to build a model for predicting
noun gender and number, or verb tense and person, or "being a part of" named
entity... these tokens won t have the "(" coverText... and consequently an
error will occurs when the tagging will be performed.
A patch has been proposed.
was:
The HMMTagger Analysis Engine class uses the
org.apache.uima.examples.tagger.Viterbi.java implementation to determine the
pos tag list of a given sentence.
In practice this implementation is partially dependant on the part of speech
tagging (likewise the remaining HMMTagger classes actually).
For exemple it makes strong assumptions on the kind of tokens it can take as
input. It assumes no restriction about the token covertext values.
It results in using some covertext probabilities for initialization or default
value when the tagger processes an unknown token...
As a consequence if the coveredText used for setting the default value is not
present in the training model an error occurs. Roughly speaking, the process
looks first for probability associated to the current token coverText, if the
coverText is not present in the model, it looks in the model for the
probability of its longest suffix, and finally if it does not found a match,
the process assigns to the unknown coverText the probability of the arbitrary
coverText : "("
The problem is that if the probability of this coverText is not available in
the model, the probability of the unknown token is not defined and a null
pointer exception occurs latter when the variable is called.
Why the probability of the "(" text would not be available in the model ? In a
large training corpus if we consider all the tokens, there is little chance not
to find at least one occurrence of "(".
Nevertheless if we use the HMM training AE to build a model for predicting
noun gender and number, or verb tense and person, or "being a part of" named
entity... these tokens won t have the "(" coverText... and consequently an
error will occurs when the tagging will be performed.
A patch has been proposed.
> Handling tokens not present in the language model (and also with no suffix
> present in the model) causes a null pointer exception in the tagger process
> ------------------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: UIMA-2106
> URL: https://issues.apache.org/jira/browse/UIMA-2106
> Project: UIMA
> Issue Type: Bug
> Components: Sandbox-Tagger
> Affects Versions: 2.3
> Environment: OS
> Linux version 2.6.32-30-generic (buildd@vernadsky) (gcc version 4.4.3 (Ubuntu
> 4.4.3-4ubuntu5) ) #59-Ubuntu SMP Tue Mar 1 21:30:21 UTC 2011
> JVM
> java version "1.6.0_17"
> Java(TM) SE Runtime Environment (build 1.6.0_17-b04)
> Java HotSpot(TM) Server VM (build 14.3-b01, mixed mode)
> Reporter: Nicolas Hernandez
> Priority: Minor
> Fix For: 2.3
>
> Attachments: TaggerHandlingTokensNotPresentInTheLanguageModel.patch
>
> Original Estimate: 5m
> Remaining Estimate: 5m
>
> The HMMTagger Analysis Engine class uses the
> org.apache.uima.examples.tagger.Viterbi.java implementation to determine the
> pos tag list of a given sentence.
> In practice this implementation is partially dependant on the part of speech
> tagging (likewise the remaining HMMTagger classes actually).
> For exemple it makes strong assumptions on the kind of tokens it can take as
> input. It assumes no restriction about the token covertext values.
> It results in using some covertext probabilities for initialization or
> default value when the tagger processes an unknown token...
> As a consequence if the coveredText used for setting the default value is not
> present in the training model an error occurs. Roughly speaking, the process
> looks first for probability associated to the current token coverText, if the
> coverText is not present in the model, it looks in the model for the
> probability of its longest suffix, and finally if it does not found a match,
> the process assigns to the unknown coverText the probability of the arbitrary
> coverText : "("
> The problem is that if the probability of this coverText is not available in
> the model, the probability of the unknown token is not defined and a null
> pointer exception occurs latter when the variable is called.
> Why the probability of the "(" text would not be available in the model ? In
> a large training corpus if we consider all the tokens, there is little chance
> not to find at least one occurrence of "(".
> Nevertheless if we use the HMM training AE to build a model for predicting
> noun gender and number, or verb tense and person, or "being a part of" named
> entity... these tokens won t have the "(" coverText... and consequently an
> error will occurs when the tagging will be performed.
> A patch has been proposed.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
