Integrating NELL with Expert System
Chirag Uttamsingh
Hello CMU NELL,
We are students at George Mason University and are working on a project to try to integrate NELL with a reasoning system (leaning agent expert system) that was developed at GMU. We are attempting to do this by trying to extract an ontology and use it in reasoning to solve a complex hypothesis.
We have browsed the ontology in NELL and, while we have found it to be very vast, we have had difficulty finding a suitable ontology for a particular domain.
For example, we have found thousands of different cars, but when we look at a particular car, say the Honda Civic, all we can really extract is the fact that it is a car, a vehicle and is made by Honda.
In our literature research, we have found a paper called, “Bootstrapping Biomedical Ontologies for Scientific Text using NELL” by Movshovitz-Attias and Cohen where they used NELL on a domain-specific corpus of information and were able to extract information with high precision, accuracy and recall using their ‘rank and learn’ system.
We are looking for more papers like this along with the ontology created during this research. Do you know more papers like these that deal with domain-specific applications of NELL? And would we be able to get the ontologies created by them?
Any help would be greatly appreciated.
Thank you,
Chirag Uttamsingh
Joseph Carneiro
Bedour Althukair
Bryan Kisiel
Forays into the biomedical domain were a bit of a side project. My
involvement was tangential and the student heading up the effort has
graduated, but I can at least link a few things for you as a starting
point.
Apparently there's really not much more than what you've already seen that
has been published, but these two things do contain some followup work
that might be of interest:
http://www.cs.cmu.edu/afs/cs/Web/People/dmovshov/docs/dma_phd_thesis.pdf
http://www.cs.cmu.edu/~wcohen/postscript/acl-2015-dana.pdf
And I found a copy of the ontology defintion that matches what was
described in the paper you've already seen:
http://rtw.ml.cmu.edu/bkisiel/bioOntologyRestricted_randomSeeds_ALL_AVAILABLE.tsv.gz
Each row of that TSV file defines one category. I don't know if you want
to go into all the details of what all the columns mean, but the two
columns of particular interest would be "categoryName" which gives the
name of each category and "generalizations" which gives the parent
categories of each category. "everyPromotedThing" is defined to be the
root category of the ontology. So from that you reconstruct the
hiearachical structure of the ontology.
If the BioNELL ontology isn't entirely to your liking, off the top of my
head I would suggest having a look at portions of the regular NELL
ontology to do with sports. This is a relatively highly-interconnected
and well-populated part of the KB, and we've found it to be a consistently
interesting bit to focus on when wanting to explore alogrithms that depend
on a degree of richness, such as when trying to infer first-order logic
rules.
Let me know if there's anythign else you'd be interested in, and I'll see
what I can do.
bki...@cs.cmu.edu
> You received this message because you are subscribed to the Google Groups "NELL: Never-Ending Language Learner" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to cmunell+u...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.
>
Joseph Carneiro
Bryan Kisiel
interesting findings -- we've had a long-standing interest in adding more
inference capabilities to NELL, but, as usual, don't have the manpower to
pursue everything we can think of.
bki...@cs.cmu.edu
Joseph Carneiro
Bryan Kisiel
NELL has for quite a while now been able to automatically depromote
beliefs in the face of changing evidence. It's not as well-developed a
scheme as one might like to see. For instance some of the learning
algorithms, such as SEAL, have the ability to learn over time and to
revise past predictions in only a loose sense of the word. But others,
like CPL and CMC do more or less relearn their models from scratch on each
iteration and will make an attempt to revise their predictions todate, in
which case NELL can depromote a belief when the confidence score drops too
low.
Thusfar, the category ontology has only ever been extended by hand.
We've toyed with some ideas for automatic category ontology extension, but
nothing has ever been developed to the point of making it into the real
system. When NELL learns a new instance of a category, that is not itself
considered to be a subcategory, even though doing that has always been an
obvious sort of thing to want to do at some point.
bki...@cs.cmu.edu
On Wed, 13 Apr 2016, Joseph Carneiro wrote:
> Hello again,
>
> Two questions came up during our research. In the paper "Toward an
> Architecture for Never-Ending Learning" it states that once a belief is
> promoted it is never demoted. This paper is older, so we're wondering if
> this is still the case. Can NELL drop a belief in future iterations? The
> other question we had is can NELL learn new categories or does it only
> learn new relations? We know that NELL can learn new subcategories, but
> does this mean NELL is learning and adding a new category to its ontology
> or does it mean NELL is learning a subset relationship between categories
> that already exist in the ontology?
>
> Thanks again,
>
> Joseph
> Chirag
> Bedour
>
> On Friday, April 1, 2016 at 8:05:47 AM UTC-4, bkisiel wrote:
>>
>> If you need anything else, just ask. And feel free to share any
>> interesting findings -- we've had a long-standing interest in adding more
>> inference capabilities to NELL, but, as usual, don't have the manpower to
>> pursue everything we can think of.
>>
Olfert Rahbek
Great question!
Best, Olfert
Bedste hilsener/Kind regards,
Olfert Rahbek
WordMaps
Margrethevej 28, DK-2900 Hellerup
Att. Olfert Rahbek
skype: orahbek28
VAT no. 3756 6020
Homepage: https://wordmaps.org/
Press "go to app" for WordMaps application
NB: Optimized for Google Chrome and Firefox browsers
Read more here: http://www.samfundslitteratur.dk/wordmaps
Joseph Carneiro
Chirag Uttamsingh
Olfert Rahbek
Hi there,
One observation:
If we look at NELL's beliefs in the area of "holy place of the gods = museion" we soon start wondering about
the many occurrences of "stock exchange" and "home".
This might be a useful candidate for studying what and how to change in order to keep museion unpolluted by more mundane matters.
Best, Olfert
Bedste hilsener/Kind regards,
Olfert Rahbek
WordMaps
Margrethevej 28, DK-2900 Hellerup
Att. Olfert Rahbek
skype: orahbek28
VAT no. 3756 6020
Homepage: https://wordmaps.org/
Press "go to app" for WordMaps application
NB: Optimized for Google Chrome and Firefox browsers
Read more here: http://www.samfundslitteratur.dk/wordmaps
Fra: cmu...@googlegroups.com [mailto:cmu...@googlegroups.com]
På vegne af Chirag Uttamsingh
Sendt: 3. juni 2016 01:56
Til: NELL: Never-Ending Language Learner <cmu...@googlegroups.com>
Emne: Re: [cmunell] Integrating NELL with Expert System
Hey bkisiel,
Bryan Kisiel
Yes, that's the one. For what it's worth we have standing plans to
replace PRA with SFE, which is described in, "Efficient and Expressive
Knowledge Base Completion Using Subgraph Feature Extraction" although it's
tough to say when we might get a chance to do that.
bki...@cs.cmu.edu
On Thu, 2 Jun 2016, Chirag Uttamsingh wrote:
> Hey bkisiel,
>
> Is PRA described in the paper, "Random Walk Inference and Learning in A
> Large Scale Knowledge Base" the current rule learning system in NELL?
>
> At this point, I am not sure if an expert system can help with rule
> learning, since the expert system uses an ontology and if that ontology
> were to come from NELL, then we might end up with some type of circular
> reasoning. However, this is an avenue that I would like to explore and I
> want to make sure I am starting in the right place.
>
> Thank you for all your help.
>
> - Chirag
>
>>
>> Hi Joseph,
>>
>> NELL has for quite a while now been able to automatically depromote
>> beliefs in the face of changing evidence. It's not as well-developed a
>> scheme as one might like to see. For instance some of the learning
>> algorithms, such as SEAL, have the ability to learn over time and to
>> revise past predictions in only a loose sense of the word. But others,
>> like CPL and CMC do more or less relearn their models from scratch on each
>> iteration and will make an attempt to revise their predictions todate, in
>> which case NELL can depromote a belief when the confidence score drops too
>> low.
>>
>> Thusfar, the category ontology has only ever been extended by hand.
>> We've toyed with some ideas for automatic category ontology extension, but
>> nothing has ever been developed to the point of making it into the real
>> system. When NELL learns a new instance of a category, that is not itself
>> considered to be a subcategory, even though doing that has always been an
>> obvious sort of thing to want to do at some point.
>>