Title: A system for generating teaching initiatives in a computeraided language learning dialogue
1A system for generating teaching initiatives in a
computer-aided language learning dialogue
- Nanda Slabbers
- University of Twente
- Netherlands
Alistair Knott University of Otago New Zealand
June 9, 2005 Nancy
2Overview
- Introduction
- Existing work
- Goals
- Formal goals
- Substantive goal
- Te Kaitito
- Authoring mode
- Student mode
- Conclusion
- Questions
3Introduction
- A dialogue system has to create two different
kinds of utterances - Responses not much content selection
- Initiatives content selection is very important
- It is common to analysethe task of
naturallanguage generationas a pipeline
4Existing Work
- Existing systems which focus on generating
initiatives - Used to accomplish a certain task.
- The difference is that a language learning
dialogue doesnt have to look like a tutorial
dialogue. - Existing systems which focus on language
learning - Most initiatives come from the student.
- In such systems it is hard to learn by adapting
utterances made by the teacher.
5Goals of the Generated Initiatives
- Content selection is the key process in
generating initiatives. - The process of content selection is typically
defined in relation to two goals - Formal goal creating a coherent text.
- Substantive goal achieving a certain effect on
the hearer (in our case teaching the student a
number of syntactic rules).
6Formal Goal - Possible Dialogue Acts
- Dialogue coherence cannot be defined at the level
of dialogue acts. - Therefore the system will only generate
initiatives of the following dialogue acts - Assertions
- Questions
- Genuine questions
- Teaching questions
7Formal Goal Topic Continuity
- We prefer initiatives which maintain the current
topic of the dialogue - Topic of utterance set of individuals and
predicates it introduces - Topic of dialogue combination of topics of last
utterances - Topic continuity overlap between topic of
dialogue and topic of possible initiative
8Formal Goal Mix of Dialogue Acts
- First we want to decide which dialogue act will
be taken - Normally we decide this based on the topic
continuity criterion. - If there are initiatives which score equally in
the topic continuity criterion, we prefer those
which belong to a dialogue act which has not been
recently used.
9Formal Goal Other Criteria
- After we have decided which dialogue act will be
taken, we have to compare the possible
initiatives belonging to that dialogue act. - To achieve this we introduce three dialogue act
specific criteria - Assertions strategy
- Genuine questions order of creation
- Teaching questions complexity
10Substantive Goal Syntax Criterion
- Every utterance is parsed and syntactic
representations are created. - Using these representations a set of target
syntactic rules can be created for each lesson. - Each lesson teaches a different set of rules,
beginning with simple rules and progressing to
more complex rules. - Using these rules initiatives which involve
target rules can be preferred to other
initiatives.
11Formal and Substantive Goals - Initiative
Generation Algorithm
- Our algorithm for generating initiatives has four
steps - Identify a set of possible topics.
- Generate a set of candidate initiatives.
- Create a shortlist of initiatives by scoring and
ranking them based on the criteria. - Assign a syntactic score and choose the
initiative with the highest total score.
12Te Kaitito
- Te Kaitito is a bilingual dialogue system meant
to teach the Maori language, but we will
demonstrate the English version. - The system supports two modes
- authoring mode to author new lessons.
- student mode to practice the lesson and learn
the target rules.
13Authoring Mode
- In authoring mode the user is assumed to be a
teacher. - The teacher can author a character by
- Telling facts about itself You are John
- Entering question generation rules How is a
person? - At the end the facts and questions are stored in
two separate knowledge bases. - The set of target syntactic rules is created
automatically.
14Authoring Mode Example Dialogue
- A Your name is John
- TK Okay
- A You are from Dunedin
- TK Okay
- A You are fine
- TK Okay
A How is a person? TK Okay A What is a persons
name? TK Okay A Where is a person from? TK Okay
15Student Mode Start of the Lesson
- Start of the dialogue
- Empty common ground
- Load knowledge bases and set of rules
- During the dialogue
- Dialogue consists of pairs of forward-looking and
backward-looking dialogue acts. - At transition relevant points the initiative is
passed to the student, but he or she can concede
the initiative to the system.
16Student Mode Example Dialogue
- S ltentergt
- TK How are you?
- S I am fine
- TK Okay
- S How are you?
- TK I am fine
- S ltentergt
- TK What is your name?
- S My name is Ann
- TK Okay
- S Who are you
- TK I am John
S ltentergt TK Where are you from? S I am from
Dunedin TK Okay S ltentergt TK I am also from
Dunedin S ltentergt TK Lets see if you
remember Who am I? S You are John TK Correct,
well done!
17Student mode End of the Lesson
- The previous dialogue continues until the system
has evidence that all target rules have been
assimilated by the student - Each target syntactic rule has an assimilation
score that is updated after an utterance. - When the score reaches a certain boundary value
the associated rule has been assimilated. - When all rules reach the boundary value, the
lesson ends successfully. - When the student does not learn the rule, even
when a correct application of the rule is shown,
the lesson may end unsuccessfully.
18Conclusion
- Te Kaititos teaching dialogues are useful to
practice conversational skills. - The system only has a simple grammar, but since
novice students also have a limited grammar, this
is satisfactory. - Adding initiatives is also beneficial
- When the student is lost, the dialogue can be
continued by simply pressing the Enter key. - The initiatives provide models of well-formed
sentences, so the student can modify these and
try them out.
19Questions