Diagnosing Language Transfer in a Web-based ICALL that Self-Improves its

Student Modeler

Victoria Tsiriga & Maria Virvou Department of Informatics,

University of Piraeus,

80 Karaoli & Dimitriou St.,

Piraeus 18534, Greece,

vtsir@unipi.gr, mvirvou@unipi.gr

Adaptivity in Web-based Adaptivity in Web-based Tutoring SystemsTutoring Systems

Adaptivity is crucial in Web-based tutoring systems.

To be adaptive, a Web-based educational system should be able to draw inferences about individual students.

Therefore, the student modelling component is crucial for the purpose of adaptation.

Adaptivity in Web-based Adaptivity in Web-based Intelligent Intelligent Computer Assisted Language Learning Computer Assisted Language Learning (ICALL) (ICALL) SystemsSystems

In Web-based ICALL systems, the student’s prior knowledge of other languages is important.Language transfer is the interference resulting from the similarities and differences between the target language and other languages the student knows. According to some, Web-based system must adopt a more general scheme in order to accommodate the international nature of the Internet. Using a machine learning mechanism would allow an ICALL system to learn how a language may interfere in learning the target language.

Overview of Web-Passive Voice Tutor Overview of Web-Passive Voice Tutor (Web-PVT)(Web-PVT)

Web-PVT is an adaptive and intelligent Web-based tutoring system that aims at teaching non-native speakers the domain of the passive voice of the English language.

Web-PVT incorporates techniques from Intelligent Tutoring Systems and Adaptive Hypermedia to tailor instruction and feedback to each individual student.

Error CategoriesError CategoriesCategory of Error Example

Accidental slipsThe student may have deleted some words and/or

may have forgotten to complete the sentence.

Spelling mistakesThe student may have typed “mather” instead of

“mother”.

Article mistakesThe student may have typed “an book” instead of “a

book”.

Irregular verb mistakesThe student may have used the usual ending “ed” to

create the past participle of an irregular verb.

Absence or redundant presence of the agentThe student may have provided the answer “he is

said by people to be wealthy” instead of “he is said to be wealthy”.

Mistakes in the word that connects with the agentThe student may have written “she was killed by a

knife” instead of “she was killed with a knife”.

Verb tense conversion mistakesThe student may have typed “is cooking” instead of

“is being cooked”.

Singular/Plural mistakesThe student may have entered “English are spoken all over the world” instead of “English is spoken all

over the world”.

Errors in the special cases of the passive voiceThe student may have provided the answer “I was let

to go out” instead of “I was allowed to go out”.

Explanation about the cause of a Explanation about the cause of a mistakemistake

Language Transfer.

Overgeneralization of the target language rules.

Ignorance of rule restrictions.

Incomplete application of rules.

False concepts hypothesized.

Carelessness.

Categories of Error and Categories of Error and Language TransferLanguage Transfer

Language transfer may cause many mistakes.

Associating categories of error with language transfer would require eliciting the expertise of many human experts.

In Web-PVT, the association of the categories of error with language transfer is performed dynamically.

Acquiring Initial Information Acquiring Initial Information about the Student in Web-PVTabout the Student in Web-PVT

Direct Provision by the student: name, mother tongue, other languages s/he already knows, and self-categorization concerning how careful s/he

is when solving exercises.

A preliminary test to assess the knowledge level of the student in the domain.

Representation of the Student Representation of the Student CharacteristicsCharacteristics

The information acquired by the student in her/his first interaction with Web-PVT is represented in a feature vector:

<Student_Code, Name, Knowledge_Level, Carefulness, Mother_Tongue, Language1,

Language2, …>

Distance Weighted k-Nearest Distance Weighted k-Nearest Neighbor AlgorithmNeighbor Algorithm

It is used to estimate the student’s proneness to make each category of error.This is done using information about other students of the same knowledge level category, who speak the same languages as the new student.The contribution of each neighbor is weighted based on her/his distance from the new student.

Calculating Distance between Calculating Distance between StudentsStudents

valuesnominal arey x,if ),y ,x(overlap

else unknown, isy or x if 1,)y,x(da

otherwise 1,

y = x if 0,)y,x(overlap

Calculating Distance between Calculating Distance between StudentsStudents

n

1aaaayx )y ,x(d)s,s(

Defining k in the k-Nearest Neighbor Algorithm

In Web-PVT the number of k is defined to be the number of students that belong to the same knowledge level category with the new student. Students that belong to different knowledge level categories are not expected to have similar knowledge of the domain, irrespective of their other personal characteristics.

Estimating Error PronenessEstimating Error Proneness

2iq

i)s ,Δ(s

1w

k

1i

1=iixi

qx

w

)s ,goryError(Cate w )s ,goryError(Cate

i

k

Main PointsMain Points

Language transfer is important for ICALL systems due to the fact that students often use already acquired knowledge while learning a new subject.

Our approach to student modeling is based on recognized similarities of the new student with other students that have already interacted with the system.