#306 - Unable to link "Indonesia" #358
Conversation
Split query for candidate concept retrieval in two parts - one that matches the label exactly and one using full-text-search. Apply limit only for the full-text-search part.
|
|
|
|
|
|
reckart
added
the
Needs attention
Fixed a syntax error
|
|
|
Still WIP? |
|
Yes, can you try to comment out MaxQueryCostEstimationTime in Virtuoso's .ini file please? #306? |
|
@rcffc the max seems to be set to 400 seconds - that seems the pretty generous?! |
|
Ok, commented out and server is re-starting. This may take a few hours. |
|
|
reckart
removed
the
Needs attention
|
Since the PR is still WIP, I expect you don't need a review immediately. Please request one when you're done. I might still comment on the code from time to time. If something needs to be discussed, just add a comment. |
- pass both unmodified type string and processed mention to the candidate retrieval query builder - try matching both surface form and type string directly - first rank by edit distance to typed string, then by signature overlap score to make it easier to link instances with many candidates by typing - remove occurrences of GRAPH, it is unclear what it's benefits are - TODO reactivate Caching
|
|
| { | ||
| aString = RenderUtils.escape(aString).toLowerCase(Locale.ENGLISH); | ||
|
|
||
| String fullTextMatchingString = getFullTextMatchingQueryPart("string", aLimit); |
There was a problem hiding this comment.
In that case, why is this a parameter at all?
| " {", | ||
| " VALUES ?labelpredicate {rdfs:label skos:altLabel}", | ||
| " {", | ||
| " ?e2 ?labelpredicate ?" + aString + " @en .", |
There was a problem hiding this comment.
And why not hard-code the variable name in the same way that e.g. ?labelpredicate is hardcoded?
| " {", | ||
| " VALUES ?labelpredicate {rdfs:label skos:altLabel}", | ||
| " {", | ||
| " ?e2 ?labelpredicate ?" + aString + " @en .", |
There was a problem hiding this comment.
Sorry, I didn't notice the ? that introduces the variable name here. So escaping here doesn't make sense in deed. But I still wonder why the variable name is not simply hard-coded.
|
@rcffc seems to work reasonably. I believe that it would be nice if the "exact match" could be case-insensitive. E.g. one of the texts I have the Wired magazine is written as So I think the PR makes improvements, but it feels like we're not there yet. I would suggest to merge it and then to open new issues for case-insensitive search and also for unit tests that check if specific concepts are found when certain search strings are being used. |
|
@rcffc btw, is the caching still disabled? Somehow doing entity linking feels very slow. |
|
Caching has been reactivated. |
|
@rcffc are you sure that you can reach the KB server? |
I read that it would be expensive because we would need to use FILTER, comparing over all concepts in a knowledge base. (https://www.cray.com/blog/dont-use-hammer-screw-nail-alternatives-regex-sparql/) |
Yes. |
|
@rcffc Maybe case sensitivity is something that needs to be configured on the KB server (e.g. Virtuoso) such that the data is simply all indexed internally in lower case? |
|
@rcffc Does entity linking work for you on master? |
No. |
|
Do you mean REC Entity Linking Playground? Yes it works, but it does not use UKP_WIKIDATA. If I use my own project with UKP_WIKIDATA on blinky it does not load the annotations: |
rcffc
removed
the
Needs attention
Split query for candidate concept retrieval in two parts - one that matches the label exactly and one using full-text-search.
Apply limit only for the full-text-search part.
Cannot test it out at the moment, since the server is still down.