On 12 April 2013 10:12, And Rosta
<and....@gmail.com> wrote:
Given that Linas got a wh-question-answering system working, I suppose that mean that the system given input like "What is Sophy likely to eat" can work out that "Sophy" is subject of "eat" and "what" is object of "eat". Is that right?
A large part of question normalization is done by RelEx. For this example, it gives:
_predadj(Sophy, likely)
_obj: the object being eaten is $qVar. Yayyy!
_to-do: the eating that will be done is likely. Yayyy!
HYP: the eating is hypothetical (it hasn't been done yet). Correct again.
_predadj: "Sophy is likely". Booo. That's just wrong (would have been right for "Smith is late")
There are two sources for errors. RelEx the "relation extractor", is extracting its relations from the link-grammar parse. If link grammar is wrong, then relex gets confused. If link-grammar is right, relex could still be wrong. (In this case, its relex that flubbed the _predadj)
The extraction is being done from this:
+--------------------Xp--------------------+
| +--------------Bsw--------------+ |
| | +------Paf------+ | |
+---Wq--+-Rw-+--SIs-+ +--TOf-+-I-+ |
| | | | | | | |
LEFT-WALL what is.v Sophy[!] likely.a to.r eat ?
The [!] says that "Sophy" is not in the link-grammar dictionaries, and instead, link-grammar is guessing it's part-of-speech rules. By contrast, "Sophie" is in the link-grammar dictionaries, as a female given name. This has some side effects: The sentence, with "Sophie", is parsed correctly on the first try. With "Sophy", the first parse is wrong, and the above is the second parse. The first parse boiled down to "What is Sophy?" which would have generated an unexpected response from the QA system: something like "Sophy is a dog".
I'm adding Sophy to the dictionaries now. This nicely illustrates one of my complaints: the endless need to maintain the dictionaries.
-- Linas