We read: A Survey of Ontology Learning Processes.
This week's meeting was held in the City Restaurant, at 0930 on Wednesday 21st.
[Notes scribbled during discussion]
General disagreement with the paper about needing ontology learning for SW; not sure that usefulness of claim about speeding up ontology creation/population.
For SW you need standards, so it has to be manual or semi-automated to check accuracy/meaningfulness. Focus on NLP in paper, which isn't SW... they don't seem to make much sense when they discuss SW. It's more machine reading / text disambiguation. Can't use it as a standard, because it would be faulty. If you have to agree on meaning you can't use automation, especially not for abstract concepts.
Meanwhile:
MA is interested in sensor data (in paper - strategy to extract concepts from data). But in SW you need to model anything.
MA needs ontology learning. Trying to find ways of telling whether one set of relations between concepts is better than another. How to abstract away from text/NL to just concepts? Won't have WordNet etc to help. SensorNet?
So much text data available, people can't resist taking advantage of it. They're too specialised. Every ontology learning paper is like this, apparently.
PP suggests looking for other terms for 'ontology learning' to find similar things from different perspectives.
MA: there's an ontology for streams of data. See Generating Data Wrapping Ontologies from Sensor Networks: a case study.
Good would be developing an ontology that reflects changes to the world (IRL) over time.
Are there intrinsic properties of ontologies are measurably good or bad?
Unsupervised semantic parsing - trying to extract semantic relations.
Unsupervised Ontology Induction From Text
eg. extract information from medical database
Evaulates by making ontology answer questions, like "what does IL-2 control?" Beats all other similar systems; 91% accuracy. They've done a few papers across different fields.
Unrelated:
This week's meeting was held in the City Restaurant, at 0930 on Wednesday 21st.
[Notes scribbled during discussion]
General disagreement with the paper about needing ontology learning for SW; not sure that usefulness of claim about speeding up ontology creation/population.
For SW you need standards, so it has to be manual or semi-automated to check accuracy/meaningfulness. Focus on NLP in paper, which isn't SW... they don't seem to make much sense when they discuss SW. It's more machine reading / text disambiguation. Can't use it as a standard, because it would be faulty. If you have to agree on meaning you can't use automation, especially not for abstract concepts.
Meanwhile:
MA is interested in sensor data (in paper - strategy to extract concepts from data). But in SW you need to model anything.
MA needs ontology learning. Trying to find ways of telling whether one set of relations between concepts is better than another. How to abstract away from text/NL to just concepts? Won't have WordNet etc to help. SensorNet?
So much text data available, people can't resist taking advantage of it. They're too specialised. Every ontology learning paper is like this, apparently.
PP suggests looking for other terms for 'ontology learning' to find similar things from different perspectives.
MA: there's an ontology for streams of data. See Generating Data Wrapping Ontologies from Sensor Networks: a case study.
Good would be developing an ontology that reflects changes to the world (IRL) over time.
Are there intrinsic properties of ontologies are measurably good or bad?
Unsupervised semantic parsing - trying to extract semantic relations.
Unsupervised Ontology Induction From Text
eg. extract information from medical database
Evaulates by making ontology answer questions, like "what does IL-2 control?" Beats all other similar systems; 91% accuracy. They've done a few papers across different fields.
Unrelated:
- Which is a word, oriented or orientated???
- Check out Calexico (band)
No comments
Post a Comment