swlee: Structured information extraction by Triplet Extractor

When we develop a software, we have to take care of "error handling". Soon-Woong Lee, our graduate student, now works for a software SIET to extract structured information from unstructured text data through triplet extraction. His general idea is about parsing the sentences, extracting triplets, mapping predicate to WordNet, using WN2SUMO to link the predicate of triplet to SUMO class. The advantage of this approach is to map the fact triplet to SUMO event class. Then we may know about what is going on from the viewpoint of "event" plus their argument's noun (or others). But the possible fault could be in the arguments that is possible to be linked to SUMO classes.

What to do could be as follows:

1. Error handling for e.g., parsing error, terminology recognition error, no entry in WordNet, etc.

2. SUMO/MILO/Domain Ontology should be utilized. Now only SUMO (somwhat old version) is used. Now there are many domain ontology in SUMO webpage.
3. ISA extractor should be plugged into this software. In general, SIET should have all functions beforehand even if the real implementation could have some technical problems. Particulary, ISA extractor is very important for all situation.
4. If there is no entry in SUMO and WN, how to do that: That is, now there is an assumption that all predicates will be mapped to WN and finally to SUMO class. It will not be possible always. If there is no WN entry, the most similar words should be found. First, list up such words. Then we may find it by classification or clustering lexically. Collocation map could be useful.
5. Structure? for DBPedia, OWL - SUMO property, relation: The title of software includes "structure extraction". What does "structure" mean here? If you think about DBPedia structure, it should be exhaustive for the given entry word. If you think about OWL-SUMO property, the goal should be to achieve the property filling of SUMO. Another thing: if you extract "relation" from unstructured text, how do you map it to "relation"?
6. SIET, this software has good UI and finally will be better. Because it is not now for commerical software, it should be academic - that means, it should have a tutorial value to teach students. Thus show the internal process and interal results like parsing, triplet, error diagnosis, Sparwl handling, etc., etc.

7. plug-in: There will be more functions in SIET. Now the current version cannot cover all of them. Many other programs will be plugged into SIET. If there is a standard interface, it is easier.
8. UIMA is one of platform for unstructured information management based on Eclipse Java. Why don't you utilize them?
9. in-memory data: now there is very good mechanism to generate a symbolic node from the predicate with arguments (is it included?) But it is based on in-memory. The serial number will be broken when the program is stopped and halts. When SIET gets another file (sentences), it is impossible to be integrated with already generated stuffs.
10. serial-number persistency and equivalence - file integration should be considered.
ambiguity and equivalence checking: There are many ambiguites when WN is mapped. But now keep all of ambiguites. Resolution of ambiguiteis is different problem after all of arguements is properly dealt and mapped into the relevant SUMO/MILO/Domain ontology classes.
11. Noisy problem: now it is expected that there are much noisy problem. It is great problem.

12. Korean text should be running on SIET.

Time Table:

1. before AND workshop

2. before ACL short paper

3. before ISWC