CG3 in dialogue systems - Vasta and Sahka Lene Antonsen, Biret Ánne Bals Baal Saara Huhmarniemi, Trond Trosterud http://giellatekno.uit.no/oahpa/ 30. oktober 2008 1
CG3 in dialogue systems - Vasta and Sahka
Lene Antonsen, Biret Ánne Bals Baal
Saara Huhmarniemi, Trond Trosterud
http://giellatekno.uit.no/oahpa/
30. oktober 2008
1
http://victorio.uit.no/oahpa/morfa/
2
Pedagogical programs for learning sámi with QA - coming
Vasta: The program generates questions, the student can answerquite freely - with grammatical feedback
Sahka: A written dialogue between the program and the student- the answer decides the progress of the dialogue.
3
Vasta
4
Generating questions
5
Maid don lohket ikte? (What did you read yesterday?)Acceptable answers:
• Mun han lohken ollu áviissaid. (I HAN read manynewspapers.)
• Ikte mun gal lohken buori girjji. (Yesterday I GAL read agood book.)
• In lohkan maidege. (I did not read anything.)
• Ikte in lohkan. (Yesterday I did not read.)
6
Maid don lohket ikte? (What did you read yesterday?)Incorrect answers:
• Mun lohket ollu áviissaid. (Not agreement subj/verbal.)
• Mun lohken ollu áviissat. (Object should be in accusative.)
7
Steps
1. Analyse (morph-disambiguate) the question and answertogether
2. Common CG up until mapping. The disambiguation isincomplete, since we are careful with the errouneous input
3. Select the relevant reading
4. Make &err with assignment mapping rules
5. Give message to the student
8
Integrating the spellchecker in our ped program
We need a restricted spellchecker
Our normal spellchecker gives to many irrelevant suggestions
a. A pedogogically adjusted speller for all answers (cut theordinary speller with 50 % or more)?
b. A spell-checker for each question/answer?
9
Didactics more important than pragmaticsThe goal is to train morphology – therefore:
• No elipsis
• Finite verb compulsatory
• No inclusive 1st person dual and plural
• The answer I do not know is not accepted
10
Answer with the same verb
Solution: Sticky tag with regex (thanks to Tino)
Exceptional handling of pro-verbs
11
QDL is the delimiter between the question and the answer.
12
The errors are ordered:
1. spelling-error
2. verb: finit/infinit, what kind of verb,...
3. agreement subj/verbal
4. person/number and tense according to question
5. case of noun, according to question and argument of verb
6. case and type of numeral
7. PP: case and type of adp
8. agreement inside NP
9. time-expression, place-expression, some particles
13
Unintended lemmas, lexical levelHow to cope with unintended lemmas?
Problem e.g. viessut which can be a rare verb - Inf or finit, or acommon mis-spelling of viesut - Noun Pl Nom (= houses)Possible solutions:
• Remove problematic rare lemmas from the analyser (but notall of the problematic ones are rare)
• Make a lexeme-specific rule for the viesut N / viessut V pair
• Make a set of the problematic lemmas/word forms, andsubstitute or change the tag in the reading
14
Marginal morphological analysese.g. possessive suffixes> Remove them (when there is another reading)
When it is the only reading,e.g. Px & strong grade vs. Loc & weak grade:> Give a comment about it to the student
15
An example – from a test student
16
Nominative instead of accusative – a common errorThe question is how to deal with it when we cannot rely on theword order, and the subject may be ommitted
• The question asks for an object (but sometimes you mayeven answer without an Acc)
• Using sets:
• what are the required arguments of the verb?
• can it have HUMAN as an object?
17
HUMAN as an object?borrat (= to eat) - the HUMAN should be the subject, not theobjectlohkat (= to read) - you could read Ibsen
18
Sahka
19
What does the student answer?We need the answer to make a logical progress in the dialogue.
20
e.g. the answer should be an object:
21
Negative or affirmativee.g. if the student does not have a car, we do not want to askwhat car she has.Neg or aff as default – in case of difficulties in the analyse.
22
23
Namese.g. What is your name? What kind of car do you have?can be a N Prop (if in the lexicon) or a stringLIST QMRK = ? ;
24
Other possibilities for progressPicking the age with regexe.g. we do not want to ask a child if he has a car.
&dia-adult&dia-young&dia-child
25