You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@ctakes.apache.org by "Masanz, James J." <Ma...@mayo.edu> on 2015/01/09 21:53:16 UTC
dictionary lookup config for best F1 measure [was RE: cTakes
Annotation Comparison
Sean (or others),
Of the various configuration options described below, which values/choices would you recommend for best F1 measure for something like the shared clef 2013 task?
https://sites.google.com/site/shareclefehealth/
I'm looking for something that doesn't have to be the best speed-wise, but that is the recommended for optimizing F1 measure.
Regards,
James
-----Original Message-----
From: Finan, Sean [mailto:Sean.Finan@childrens.harvard.edu]
Sent: Friday, December 19, 2014 11:55 AM
To: dev@ctakes.apache.org; kim.ebert@imatsolutions.com
Subject: RE: cTakes Annotation Comparison
Well, I guess that it is time for me to speak up …
I must say that I’m happy that people are showing interest in the fast lookup. I am also happy (sort of) that some concerns are being raised – and that there is now community participation in my little toy. I have some concerns about what people are reporting. This does not coincide with what I have seen at all. Yesterday I started (without knowing this thread existed) testing a bare-minimum pipeline for CUI extraction. It is just the stripped-down Aggregate with only: segment, tokens, sentences, POS, and the fast lookup. The people at Children’s wanted to know how fast we could get. 1,196 notes in under 90 seconds on my laptop with over 210,000 annotations, which is 175/note. After reading the thread I decided to run the fast lookup with several configurations. I also ran the default for 10.5 hours. I am comparing the annotations from each system against the human annotations that we have, and I will let everybody know what I find – for better or worse.
The fast lookup does not (out-of-box) do the exact same thing as the default. Some things can be configured to make it more closely approximate the default dictionary.
1. Set the minimum annotation span length to 2 (default is 3). This is in desc/[ae]/UmlsLookupAnnotator.xml : line #78. The annotator should then pick up text like “CT” and improve recall, but it will hurt precision.
2. Set the Lookup Window to LookupWindowAnnotation. This is in desc/[ae]/UmlsLookupAnnotator.xml: lines #65 & #93. The LookupWindowAnnotator will need to be added to the aggregate pipeline AggregatePlaintextFastUMLSProcesor.xml lines #50 & #172. This will narrow the lookup window and may increase precision, but (in my experience) reduces recall.
3. Allow the –rough- identification of Overlapping spans. The default dictionary will often identify text like “metastatic colorectal carcinoma” when that text actually does not exist anywhere in umls. It basically ignores “colorectal” and gives the whole span the CUI for “metastatic carcinoma”. In this case it is arguably a good thing. In many others it is arguably not so much. There is a Class ... lookup2.ae.OverlapJCasTermAnnotator.java that will do the same thing. You can create a new desc/[ae]/*Annotator.xml or just change the <annotatorImplementationName> in desc/[ae]/UmlsLookupAnnotator.xml line #25. I will check in a new desc xml (sorry; thought I had) because there are 2 parameters unique to OverlapJCasTermAnnotator
4. You can play with the OverlapJCasTermAnnotator parameters “consecutiveSkips” and “totalTokenSkips”. These control just how lenient you want the overlap tagging to be.
5. Create a new dictionary database. There is a (bit messy) DictionaryTool in sandbox that will let you dump whatever you do or do not want from UMLS into a database. It will also help you clean up or –select- stored entries as well. There is a lot of garbage in the default dictionary database: repeated terms with caps/no caps (“Cancer”,”cancer”), text with metadata (“cancer [finding]”) and text that just clutters (“PhenX: entry for cancer”, “1”, “2”). The fast lookup database should have most of the Snomed and RxNorm terms (and synonyms) of interest, but you could always make a new database that is much more inclusive.
The main key to the speed of the fast dictionary lookup is actually … the key. It is the way that the database is indexed and the lookup by “rare” word instead of “first” word. Everything else can be changed around it and it should still be a faster version.
As for the false positives like “Today”, that will always be a problem until we have disambiguation. The lookup is basically a glorified grep.
Sean
From: Chen, Pei [mailto:Pei.Chen@childrens.harvard.edu]
Sent: Friday, December 19, 2014 10:43 AM
To: dev@ctakes.apache.org; kim.ebert@imatsolutions.com
Subject: RE: cTakes Annotation Comparison
Also check out stats that Sean ran before releasing the new component on:
http://svn.apache.org/repos/asf/ctakes/trunk/ctakes-dictionary-lookup-fast/doc/DictionaryLookupStats.docx
From the evaluation and experience, the new lookup algorithm should be a huge improvement in terms of both speed and accuracy.
This is very different than what Bruce mentioned… I’m sure Sean will chime here.
(The old dictionary lookup is essentially obsolete now- plagued with bugs/issues as you mentioned.)
--Pei
From: Kim Ebert [mailto:kim.ebert@perfectsearchcorp.com]
Sent: Friday, December 19, 2014 10:25 AM
To: dev@ctakes.apache.org<ma...@ctakes.apache.org>
Subject: Re: cTakes Annotation Comparison
Guergana,
I'm curious to the number of records that are in your gold standard sets, or if your gold standard set was run through a long running cTAKES process. I know at some point we fixed a bug in the old dictionary lookup that caused the permutations to become corrupted over time. Typically this isn't seen in the first few records, but over time as patterns are used the permutations would become corrupted. This caused documents that were fed through cTAKES more than once to have less codes returned than the first time.
For example, if a permutation of 4,2,3,1 was found, the permutation would be corrupted to be 1,2,3,4. It would no longer be possible to detect permutations of 4,2,3,1 until cTAKES was restarted. We got the fix in after the cTAKES 3.2.0 release. https://issues.apache.org/jira/browse/CTAKES-310 Depending upon the corpus size, I could see the permutation engine eventually only have a single permutation of 1,2,3,4.
Typically though, this isn't very easily detected in the first 100 or so documents.
We discovered this issue when we made cTAKES have consistent output of codes in our system.
[IMAT Solutions]<http://imatsolutions.com>
Kim Ebert
Software Engineer
[Office:]801.669.7342
kim.ebert@imatsolutions.com<ma...@imatsolutions.com>
On 12/19/2014 07:05 AM, Savova, Guergana wrote:
We are doing a similar kind of evaluation and will report the results.
Before we released the Fast lookup, we did a systematic evaluation across three gold standard sets. We did not see the trend that Bruce reported below. The P, R and F1 results from the old dictionary look up and the fast one were similar.
Thank you everyone!
--Guergana
-----Original Message-----
From: David Kincaid [mailto:kincaid.dave@gmail.com]
Sent: Friday, December 19, 2014 9:02 AM
To: dev@ctakes.apache.org<ma...@ctakes.apache.org>
Subject: Re: cTakes Annotation Comparison
Thanks for this, Bruce! Very interesting work. It confirms what I've seen in my small tests that I've done in a non-systematic way. Did you happen to capture the number of false positives yet (annotations made by cTAKES that are not in the human adjudicated standard)? I've seen a lot of dictionary hits that are not actually entity mentions, but I haven't had a chance to do a systematic analysis (we're working on our annotated gold standard now). One great example is the antibiotic "Today". Every time the word today appears in any text it is annotated as a medication mention when it almost never is being used in that sense.
These results by themselves are quite disappointing to me. Both the UMLSProcessor and especially the FastUMLSProcessor seem to have pretty poor recall. It seems like the trade off for more speed is a ten-fold (or more) decrease in entity recognition.
Thanks again for sharing your results with us. I think they are very useful to the project.
- Dave
On Thu, Dec 18, 2014 at 5:06 PM, Bruce Tietjen < bruce.tietjen@perfectsearchcorp.com<ma...@perfectsearchcorp.com>> wrote:
Actually, we are working on a similar tool to compare it to the human
adjudicated standard for the set we tested against. I didn't mention
it before because the tool isn't complete yet, but initial results for
the set (excluding those marked as "CUI-less") was as follows:
Human adjudicated annotations: 4591 (excluding CUI-less)
Annotations found matching the human adjudicated standard
UMLSProcessor 2245
FastUMLSProcessor 215
[image: IMAT Solutions] <http://imatsolutions.com><http://imatsolutions.com> Bruce Tietjen
Senior Software Engineer
[image: Mobile:] 801.634.1547
bruce.tietjen@imatsolutions.com<ma...@imatsolutions.com>
On Thu, Dec 18, 2014 at 3:37 PM, Chen, Pei
<Pe...@childrens.harvard.edu>
wrote:
Bruce,
Thanks for this-- very useful.
Perhaps Sean Finan comment more-
but it's also probably worth it to compare to an adjudicated human
annotated gold standard.
--Pei
-----Original Message-----
From: Bruce Tietjen [mailto:bruce.tietjen@perfectsearchcorp.com]
Sent: Thursday, December 18, 2014 1:45 PM
To: dev@ctakes.apache.org<ma...@ctakes.apache.org>
Subject: cTakes Annotation Comparison
With the recent release of cTakes 3.2.1, we were very interested in
checking for any differences in annotations between using the
AggregatePlaintextUMLSProcessor pipeline and the
AggregatePlanetextFastUMLSProcessor pipeline within this release of
cTakes
with its associated set of UMLS resources.
We chose to use the SHARE 14-a-b Training data that consists of 199
documents (Discharge 61, ECG 54, Echo 42 and Radiology 42) as the
basis for the comparison.
We decided to share a summary of the results with the development
community.
Documents Processed: 199
Processing Time:
UMLSProcessor 2,439 seconds
FastUMLSProcessor 1,837 seconds
Total Annotations Reported:
UMLSProcessor 20,365 annotations
FastUMLSProcessor 8,284 annotations
Annotation Comparisons:
Annotations common to both sets: 3,940
Annotations reported only by the UMLSProcessor: 16,425
Annotations reported only by the FastUMLSProcessor: 4,344
If anyone is interested, following was our test procedure:
We used the UIMA CPE to process the document set twice, once using
the AggregatePlaintextUMLSProcessor pipeline and once using the
AggregatePlaintextFastUMLSProcessor pipeline. We used the
WriteCAStoFile CAS consumer to write the results to output files.
We used a tool we recently developed to analyze and compare the
annotations generated by the two pipelines. The tool compares the
two outputs for each file and reports any differences in the
annotations (MedicationMention, SignSymptomMention,
ProcedureMention, AnatomicalSiteMention, and
DiseaseDisorderMention) between the two output sets. The tool
reports the number of 'matches' and 'misses' between each annotation set. A 'match'
is
defined as the presence of an identified source text interval with
its associated CUI appearing in both annotation sets. A 'miss' is
defined as the presence of an identified source text interval and
its associated CUI in one annotation set, but no matching identified
source text interval
and
CUI in the other. The tool also reports the total number of
annotations (source text intervals with associated CUIs) reported in
each annotation set. The compare tool is in our GitHub repository at
https://github.com/perfectsearch/cTAKES-compare
RE: dictionary lookup config for best F1 measure [was RE: cTakes
Annotation Comparison
Posted by "Finan, Sean" <Se...@childrens.harvard.edu>.
Hi James,
Great question. In truth, you may need to run a few times to find out. Doing that with a full pipeline would be tedious, but there is a descriptor in clinical-pipeline named CuisOnlyPlaintextUMLSProcessor.xml that will only obtain Umls cuis. It runs ~50,000 notes per hour on my laptop as-is, so I suggest that you test with that ae. It has lvg commented out by default (for speed). Adding lvg will increase the runtime, but it also will (as you know) find a few additional terms. You can try a few configurations without it and then the best option with it. If you want to test the default dictionary lookup then you can certainly swap the referenced lookup xmls.
Changes to the fast dictionary configuration are made in two places:
1. The main descriptor ...-fast/desc/analysis_engine/UmlsLookupAnnotator.xml
2. The resource (dictionary) configuration file resources/.../fast/cTakesHsql..xml
A few suggestions, in order of impact:
1. I am guessing that the annotations in clef are human annotated with longest-length spans only. In other words, "colon cancer" instead of "colon cancer" and "cancer". To best approximate this style of annotation, edit the cTakesHsql.xml in the section <rareWordConsumer> and change the selected implementation. By default it is DefaultTermConsumer (go figure), but you will want to use the commented-out PrecisionTermConsumer. As the above cTakesHsql comment indicates " DefaultTermConsumer will persist all spans.
PrecisionTermConsumer will only persist only the longest overlapping span of any semantic group." Doing this should increase precision, and depending upon how "good" the annotations are it should not greatly change recall.
2. Just for kicks, try using SemanticCleanupTermConsumer. It may slightly increase precision, but it also may decrease recall. Hopefully it doesn't do much at all (PrecisionTermConsumer and proper semantic typing in the dictionary should suffice without this term consumer).
3. Especially for task 2 (acronyms & abbreviations), you should try a run with <name>minimumSpan</name> in UmlsLookupAnnotator.xml set to 2. This changes the minimum allowable span of a term. The default is 3 to increase precision on acronyms & abbreviations, but decreasing to 2 may improve recall on the same. The dictionary is not built with anything below 2 characters.
4. On that note (character length), if task 1 does not include acronyms & abbreviations, then you can try increasing the minimum span length above 3 and see if there is a good increase in precision without a significant decrease in recall.
5. Try a few runs with overlapping spans in addition to exact matches. To do this use the OverlapJCasTermAnnotator instead of the DefaultJCasTermAnnotator annotator implementation. DefaultJCasTermAnnotator is specified in UmlsLookupAnnotator.xml but I will check in a descriptor for overlap matching. There are additional parameters for that option, but I'll email them after I checkin.
6. By default the new lookup uses Sentence as the lookup window. I did this for two reasons: 1. Not all terms are within Noun Phrases, 2. Some Noun Phrases overlapped, causing repeated lookups (in my 3.0 candidate trials), and 3. Not all cTakes Noun Phrases are accurate. Because the lookup is fast, using a full Sentence for lookup doesn't seem to hurt much. However, you can always switch it back to see if precision is increased enough to warrant the decrease in recall. This is changed in UmlsLookupAnnotator.xml
I have run my own tests with the various setups, but I don't want to adversely influence what you run just in case the trends with the share/clef annotations differ.
Sean
-----Original Message-----
From: Masanz, James J. [mailto:Masanz.James@mayo.edu]
Sent: Friday, January 09, 2015 3:57 PM
To: 'dev@ctakes.apache.org'
Subject: dictionary lookup config for best F1 measure [was RE: cTakes Annotation Comparison
Sean (or others),
Of the various configuration options described below, which values/choices would you recommend for best F1 measure for something like the shared clef 2013 task?
https://sites.google.com/site/shareclefehealth/
I'm looking for something that doesn't have to be the best speed-wise, but that is the recommended for optimizing F1 measure.
Regards,
James
-----Original Message-----
From: Finan, Sean [mailto:Sean.Finan@childrens.harvard.edu]
Sent: Friday, December 19, 2014 11:55 AM
To: dev@ctakes.apache.org; kim.ebert@imatsolutions.com
Subject: RE: cTakes Annotation Comparison
Well, I guess that it is time for me to speak up …
I must say that I’m happy that people are showing interest in the fast lookup. I am also happy (sort of) that some concerns are being raised – and that there is now community participation in my little toy. I have some concerns about what people are reporting. This does not coincide with what I have seen at all. Yesterday I started (without knowing this thread existed) testing a bare-minimum pipeline for CUI extraction. It is just the stripped-down Aggregate with only: segment, tokens, sentences, POS, and the fast lookup. The people at Children’s wanted to know how fast we could get. 1,196 notes in under 90 seconds on my laptop with over 210,000 annotations, which is 175/note. After reading the thread I decided to run the fast lookup with several configurations. I also ran the default for 10.5 hours. I am comparing the annotations from each system against the human annotations that we have, and I will let everybody know what I find – for better or worse.
The fast lookup does not (out-of-box) do the exact same thing as the default. Some things can be configured to make it more closely approximate the default dictionary.
1. Set the minimum annotation span length to 2 (default is 3). This is in desc/[ae]/UmlsLookupAnnotator.xml : line #78. The annotator should then pick up text like “CT” and improve recall, but it will hurt precision.
2. Set the Lookup Window to LookupWindowAnnotation. This is in desc/[ae]/UmlsLookupAnnotator.xml: lines #65 & #93. The LookupWindowAnnotator will need to be added to the aggregate pipeline AggregatePlaintextFastUMLSProcesor.xml lines #50 & #172. This will narrow the lookup window and may increase precision, but (in my experience) reduces recall.
3. Allow the –rough- identification of Overlapping spans. The default dictionary will often identify text like “metastatic colorectal carcinoma” when that text actually does not exist anywhere in umls. It basically ignores “colorectal” and gives the whole span the CUI for “metastatic carcinoma”. In this case it is arguably a good thing. In many others it is arguably not so much. There is a Class ... lookup2.ae.OverlapJCasTermAnnotator.java that will do the same thing. You can create a new desc/[ae]/*Annotator.xml or just change the <annotatorImplementationName> in desc/[ae]/UmlsLookupAnnotator.xml line #25. I will check in a new desc xml (sorry; thought I had) because there are 2 parameters unique to OverlapJCasTermAnnotator
4. You can play with the OverlapJCasTermAnnotator parameters “consecutiveSkips” and “totalTokenSkips”. These control just how lenient you want the overlap tagging to be.
5. Create a new dictionary database. There is a (bit messy) DictionaryTool in sandbox that will let you dump whatever you do or do not want from UMLS into a database. It will also help you clean up or –select- stored entries as well. There is a lot of garbage in the default dictionary database: repeated terms with caps/no caps (“Cancer”,”cancer”), text with metadata (“cancer [finding]”) and text that just clutters (“PhenX: entry for cancer”, “1”, “2”). The fast lookup database should have most of the Snomed and RxNorm terms (and synonyms) of interest, but you could always make a new database that is much more inclusive.
The main key to the speed of the fast dictionary lookup is actually … the key. It is the way that the database is indexed and the lookup by “rare” word instead of “first” word. Everything else can be changed around it and it should still be a faster version.
As for the false positives like “Today”, that will always be a problem until we have disambiguation. The lookup is basically a glorified grep.
Sean
From: Chen, Pei [mailto:Pei.Chen@childrens.harvard.edu]
Sent: Friday, December 19, 2014 10:43 AM
To: dev@ctakes.apache.org; kim.ebert@imatsolutions.com
Subject: RE: cTakes Annotation Comparison
Also check out stats that Sean ran before releasing the new component on:
http://svn.apache.org/repos/asf/ctakes/trunk/ctakes-dictionary-lookup-fast/doc/DictionaryLookupStats.docx
From the evaluation and experience, the new lookup algorithm should be a huge improvement in terms of both speed and accuracy.
This is very different than what Bruce mentioned… I’m sure Sean will chime here.
(The old dictionary lookup is essentially obsolete now- plagued with bugs/issues as you mentioned.) --Pei
From: Kim Ebert [mailto:kim.ebert@perfectsearchcorp.com]
Sent: Friday, December 19, 2014 10:25 AM
To: dev@ctakes.apache.org<ma...@ctakes.apache.org>
Subject: Re: cTakes Annotation Comparison
Guergana,
I'm curious to the number of records that are in your gold standard sets, or if your gold standard set was run through a long running cTAKES process. I know at some point we fixed a bug in the old dictionary lookup that caused the permutations to become corrupted over time. Typically this isn't seen in the first few records, but over time as patterns are used the permutations would become corrupted. This caused documents that were fed through cTAKES more than once to have less codes returned than the first time.
For example, if a permutation of 4,2,3,1 was found, the permutation would be corrupted to be 1,2,3,4. It would no longer be possible to detect permutations of 4,2,3,1 until cTAKES was restarted. We got the fix in after the cTAKES 3.2.0 release. https://issues.apache.org/jira/browse/CTAKES-310 Depending upon the corpus size, I could see the permutation engine eventually only have a single permutation of 1,2,3,4.
Typically though, this isn't very easily detected in the first 100 or so documents.
We discovered this issue when we made cTAKES have consistent output of codes in our system.
[IMAT Solutions]<http://imatsolutions.com>
Kim Ebert
Software Engineer
[Office:]801.669.7342
kim.ebert@imatsolutions.com<ma...@imatsolutions.com>
On 12/19/2014 07:05 AM, Savova, Guergana wrote:
We are doing a similar kind of evaluation and will report the results.
Before we released the Fast lookup, we did a systematic evaluation across three gold standard sets. We did not see the trend that Bruce reported below. The P, R and F1 results from the old dictionary look up and the fast one were similar.
Thank you everyone!
--Guergana
-----Original Message-----
From: David Kincaid [mailto:kincaid.dave@gmail.com]
Sent: Friday, December 19, 2014 9:02 AM
To: dev@ctakes.apache.org<ma...@ctakes.apache.org>
Subject: Re: cTakes Annotation Comparison
Thanks for this, Bruce! Very interesting work. It confirms what I've seen in my small tests that I've done in a non-systematic way. Did you happen to capture the number of false positives yet (annotations made by cTAKES that are not in the human adjudicated standard)? I've seen a lot of dictionary hits that are not actually entity mentions, but I haven't had a chance to do a systematic analysis (we're working on our annotated gold standard now). One great example is the antibiotic "Today". Every time the word today appears in any text it is annotated as a medication mention when it almost never is being used in that sense.
These results by themselves are quite disappointing to me. Both the UMLSProcessor and especially the FastUMLSProcessor seem to have pretty poor recall. It seems like the trade off for more speed is a ten-fold (or more) decrease in entity recognition.
Thanks again for sharing your results with us. I think they are very useful to the project.
- Dave
On Thu, Dec 18, 2014 at 5:06 PM, Bruce Tietjen < bruce.tietjen@perfectsearchcorp.com<ma...@perfectsearchcorp.com>> wrote:
Actually, we are working on a similar tool to compare it to the human
adjudicated standard for the set we tested against. I didn't mention
it before because the tool isn't complete yet, but initial results for
the set (excluding those marked as "CUI-less") was as follows:
Human adjudicated annotations: 4591 (excluding CUI-less)
Annotations found matching the human adjudicated standard
UMLSProcessor 2245
FastUMLSProcessor 215
[image: IMAT Solutions] <http://imatsolutions.com><http://imatsolutions.com> Bruce Tietjen
Senior Software Engineer
[image: Mobile:] 801.634.1547
bruce.tietjen@imatsolutions.com<ma...@imatsolutions.com>
On Thu, Dec 18, 2014 at 3:37 PM, Chen, Pei
<Pe...@childrens.harvard.edu>
wrote:
Bruce,
Thanks for this-- very useful.
Perhaps Sean Finan comment more-
but it's also probably worth it to compare to an adjudicated human
annotated gold standard.
--Pei
-----Original Message-----
From: Bruce Tietjen [mailto:bruce.tietjen@perfectsearchcorp.com]
Sent: Thursday, December 18, 2014 1:45 PM
To: dev@ctakes.apache.org<ma...@ctakes.apache.org>
Subject: cTakes Annotation Comparison
With the recent release of cTakes 3.2.1, we were very interested in
checking for any differences in annotations between using the
AggregatePlaintextUMLSProcessor pipeline and the
AggregatePlanetextFastUMLSProcessor pipeline within this release of
cTakes
with its associated set of UMLS resources.
We chose to use the SHARE 14-a-b Training data that consists of 199
documents (Discharge 61, ECG 54, Echo 42 and Radiology 42) as the
basis for the comparison.
We decided to share a summary of the results with the development
community.
Documents Processed: 199
Processing Time:
UMLSProcessor 2,439 seconds
FastUMLSProcessor 1,837 seconds
Total Annotations Reported:
UMLSProcessor 20,365 annotations
FastUMLSProcessor 8,284 annotations
Annotation Comparisons:
Annotations common to both sets: 3,940
Annotations reported only by the UMLSProcessor: 16,425
Annotations reported only by the FastUMLSProcessor: 4,344
If anyone is interested, following was our test procedure:
We used the UIMA CPE to process the document set twice, once using
the AggregatePlaintextUMLSProcessor pipeline and once using the
AggregatePlaintextFastUMLSProcessor pipeline. We used the
WriteCAStoFile CAS consumer to write the results to output files.
We used a tool we recently developed to analyze and compare the
annotations generated by the two pipelines. The tool compares the
two outputs for each file and reports any differences in the
annotations (MedicationMention, SignSymptomMention,
ProcedureMention, AnatomicalSiteMention, and
DiseaseDisorderMention) between the two output sets. The tool
reports the number of 'matches' and 'misses' between each annotation set. A 'match'
is
defined as the presence of an identified source text interval with
its associated CUI appearing in both annotation sets. A 'miss' is
defined as the presence of an identified source text interval and
its associated CUI in one annotation set, but no matching identified
source text interval
and
CUI in the other. The tool also reports the total number of
annotations (source text intervals with associated CUIs) reported in
each annotation set. The compare tool is in our GitHub repository at
https://github.com/perfectsearch/cTAKES-compare