Dialogue and grammar
Document Sample


Goteborg University
Dialogue Systems Lab
Dialogue and grammar
Karin Cavallin, Robin Cooper
David Hjelm, Staffan Larsson
Göteborg University
Dialogue Systems Lab
Language Technology in Sweden
2004-05-12
Goteborg University
Dialogue Systems Lab
Speech recognition and NL
interpretation & generation
• Options for SR:
– Statistical language model
– Grammar model
• Options for NL interpretation:
– Word/phrase spotting
– Grammar & parsing
• Options for NL generation
– Template-based
– Grammar-based
• Can be combined in various ways
Goteborg University
Dialogue Systems Lab
Are grammars useful for dialogue
systems?
• Most working systems don’t use a
grammar!
• Instead:
– keyword or phrase-spotting techniques for
interpretation
– template-based generation
• Also, most working systems are built
more or less from scratch
– do not fully separate application-specific
knowledge from general and reusable
knowledge
Goteborg University
Dialogue Systems Lab
GoDiS
• A dialogue system built using TrindiKit and based on the
information state approach
• Implements a theory of Issue-based Dialogue
Management
• Some goals with GoDiS:
– Provide a domain-independent theory covering several
dialogue genres
– Modularity and reusability of system components
– Enable rapid prototyping of new applications
• Focus on dialogue management
• These goals have been achieved on the level of dialogue
management...
– ... but not for NL processing
Goteborg University
Dialogue Systems Lab
SR and interpretation in GoDiS
• Started out as text-based system
– Phrase spotting for interpretation;
strings to dialogue moves
– Templates for generation; dialogue
moves to strings
• Added SR component
– Did not have a corpus, so needed to
write a speech recognition grammar
by hand
Goteborg University
Dialogue Systems Lab
tvGoDiS
• GoDiS application for programming a
VCR over the telephone
– schedule recordings
– remove planned recordings
– TV card used for recording television
programs as video files
• Small corpus collected of users calling the
system
– approx. 100 sessions (calls) and 1100 utterances
– misrecognised utterances not yet transcribed
– GUI enables immediate access to corpus
Goteborg University
Dialogue Systems Lab
Example from tvGoDIS corpus
• TrindiGui displaying
a tvGoDiS session
• Audio files linked
with utterances
• Access to
information state at
different stages of
the dialogue
Goteborg University
Dialogue Systems Lab
Problems with GoDiS
• Mismatch between SR module and
interpretation module
– correctly recognised speech was
misinterpreted
– Leads to strange and unexpected
behaviour (unexpected interactions
with dialogue management)
Goteborg University
Dialogue Systems Lab
ASR/interpretation mismatch
S> Välkommen till videon! För
att få hjälp, säg hjälp. • jag vill spela in ett program
S> Låt oss se. Vad kan jag – Covered by SR grammar
hjälpa dig med? – Not covered by phrase spotting
lexicon
U> jag vill spela in ett program – However spela in and ett are
(0.6) recognized by interpretation
module
S> Vill du lägga till en
inspelning? • Correct interpretation:
– jag vill spela in ett program
U> ja (0.98) request(add_rec_job)
S> Låt oss se. Menar du från • Actual interpretation:
noll ett eller till noll ett? – spela in
request(add_rec_job
– ett
answer(time(1:00))
• System assumes ett specifies
either when the recording should
stop or end – thus the follow-up
question
Goteborg University
Dialogue Systems Lab
Problems with GoDiS
• Mismatch between interpretation
module and generation module
– System said something that led the
user to believe they could say X
– but the system does not understand
X
Goteborg University
Dialogue Systems Lab
Interpretation/generation mismatch
S> Du kan lägga till en • Help message suggests that
inspelning, ta bort en the user can say lista
planerad inspelning, och inspelningar
lista planerade inspelningar. • But help message is not up-
För att börja om, säg börja to-date, and neither is SR
om. För att sluta, säg sluta. grammar.
S> huvudmenyn. Vad kan jag • Instead interpretation
hjälpa dig med? module finds word
U> lista inspelningar (0.75) inspelningar
S> Ockey. gå till inspelningar. • The user is asked to choose
S> Låt oss se. Vill du lägga between roughly the same
till en inspelning , radera alternatives that was
dina inspelningar eller få presented in the help
information om dina message
inspelningar?
Goteborg University
Dialogue Systems Lab
• Bugs can of course be fixed...
– would take lots of time
– no guarantee that new bugs won’t be
introduced
– no guarantee that all bugs have shown up
• These solutions are not reusable
• Instead, we want a general and
principled solution
– enable reuse and rapid prototyping of
applications, including grammars
Goteborg University
Dialogue Systems Lab
Solution, part 1: single grammar for
SR, interpretation, generation
• No mismatches between the grammars
• Only one grammar needs to be written for each
application
• However:
– the different modules may require different kinds of
grammars, in different formats
• So:
– We need to be able to generate all 3 grammars from a
single grammar
• However, we may not want the exact same coverage for
SR, interpretation and generation grammars
– May not want to understand everything that can be
generated
Goteborg University
Dialogue Systems Lab
Solution, part 2: use GF!
• GF is a powerful tool for mutilingual grammar
development
• Can be used to generate grammars in various
other formats
• Subgrammars can be extracted from larger
grammars
– SR, interpretation and generation subgrammars may
have different coverage
• In addition, GF offers resource grammars in
several languages
– This promises to further decrease the time needed
for writing a new application grammar
Goteborg University
Dialogue Systems Lab
GF GoDiS Grammar
• Application grammars
– User grammar
– System grammar
– Shared grammar, used by user grammar
and system grammar
• Utility grammars
– Weekday, Time, Channel
• Application grammars may use
– Utility grammars
– Resource grammars
Goteborg University
Dialogue Systems Lab
GF VCR grammar
• One abstract syntax
– GoDiS dialogue moves as GF abstract trees
• Three concrete syntaxes
– Swedish
– English
– GoDiS dialogue moves in prolog notation
• Statistics:
– Abstract syntax : ~420 lines of code
– Concrete syntax :
• Prolog: ~440 lines of code
• Swedish: ~440 lines of code
Goteborg University
Dialogue Systems Lab
Parsing Swedish to abstract syntax
> p -cat=DMove -lang=Swe "spela in från kvart över tolv"
request (vcr_add_rec_job_args1 (startTimeToStoreReq
(startTimeToStore (time hour12 minute15))))
Goteborg University
Dialogue Systems Lab
Linearising abstract syntax to Prolog
notation
> l -lang=Prolog request (vcr_add_rec_job_args1
(startTimeToStoreReq (startTimeToStore (time hour12
minute15))))
request(add_rec_job , answer(start_time_to_store(12 :
15))
Goteborg University
Dialogue Systems Lab
Translating Prolog to Swedish
> t -cat=DMove Prolog Swe "request(add_rec_job ,
answer(start_time_to_store(12 : 15))"
spela in från tolv femton
spela in från tolv och femton
spela in från femton över tolv
spela in från femton minuter över tolv
spela in från kvart över tolv
Goteborg University
Dialogue Systems Lab
Time Grammar
• A multilingual grammar module
• Swedish, English, Prolog, started on
French and Catalan.
Goteborg University
Dialogue Systems Lab
Constructing concrete syntaxes for
time expressions isn’t always trivial
• Abstract: time hour12 Minute30
• Prolog: 12 : 30
• English: half past twelve / half twelve
• Swedish: halv ett
• klockan två
• klockan två och femton
• ?klockan fjorton
• two o’clock
• *two fifteen o’clock
• *fourteen o’clock
Corpus studies needed
Goteborg University
Dialogue Systems Lab
Another problem
• am or pm?
• på morgonen
• på eftermiddagen
• på kvällen
• på natten
Goteborg University
Dialogue Systems Lab
Combining GF and statistical
methods for robustness
• Use GF to generate a corpus to use as a
baseline for SLM training
• Use GF-generated SR grammars in
initial system and collect corpus to be
used for training SLMs for speech
recognition
• For recognized out-of-grammar string,
try to find nearest in-grammar string;
confirm this with user (Gorrell 2003)
Goteborg University
Dialogue Systems Lab
Conclusion
• Keeping track of separate SR,
interpretation and generation grammars
in a dialogue system is time-consuming
• GF offers the possibility of rapid
prototyping of grammars for use in
dialogue system applications
– Generate other grammar formats from GF
– Resource grammars for several languages
• GF can be combined with statistical
methods for robustness
Related docs
Get documents about "