Realizing Adaptive Questions and Answers
for ICALL Systems
Hidenobu Kunichika*1, Minoru Urushima*2, Tsukasa Hirashima*3 and Akira Takeuchi*2
*1
Dept. of Creation Informatics, Kyushu Institute of Technology, Japan
*2
Dept. of Artificial Intelligence, Kyushu Institute of Technology, Japan
*3
Dept. of Information Engineering, Hiroshima University, Japan
Abstract. Language training systems that provide learners adaptive questions on the
contents of stories require several capabilities such as semantic analysis, automated
question generation and diagnosis of learners' answer sentences. This paper presents a
method of selecting questions from a generated list to realize adaptive questions and
answers. Our method filters out similar questions, and then selects questions by
considering the difficulty, types and order. This paper also describes an evaluation of
our method. As the result of our experiment demonstrates, we have found that our
method generates a viable series of questions.
1. Introduction
It is common in second language learning, to answer questions on the contents of passages
after listening and/or reading them. Such questions and answers (QA) in the target language is
effective for acquiring practical skills because multiple language skills are required to answer
the questions, in particular to grasp the contents of the story and the questions, as well as to
compose answers. Many computer assisted language learning systems have been developed
[1]. Some are equipped with exercise functions which ask about the contents of sentences.
Most, however, use questions prepared beforehand [5, 6, 7]. Thus these have the problem that
such systems will present questions without considering the learner's level of understanding
because the number of prepared questions is limited.
In order to solve these problem, we are aiming to realize a QA function which provides
adaptive questions on the surface semantics of English stories prepared by authors or learners.
To realize the QA function, the following sub-functions are necessary: (1) to understand
English sentences and to extract syntactic and semantic information, (2) to generate
automatically various kinds of question sentences for presentation to learners who have
varying degrees of comprehension, (3) to select adaptive questions from a set of generated
question sentences, (4) to analyze learners' answer sentences and to diagnose errors and (5) to
offer intelligent help for the correction of errors and the acquisition of correct knowledge by
referring to the student models. In earlier studies, we have already implemented the sub-
modules for the functions (1), (2), (4) and (5). This paper proposes an adaptive method of
selecting questions for the function (3).
2. The outline of the QA function
Our QA function gives learners questions about the contents of a story. After studying the
contents of the story by reading and/or listening, they answer the questions. Aims of our QA
are both to train for conversation by using multiple skills through reading a story, listening to
or reading questions and composing answers and to give learners a chance to realize their own
state of understanding of, for example, vocabulary and grammar; and to practice usage
through QA. To reduce the burden of memorizing the content of the story and to concentrate
on composing sentences from memory, the length of any one passage in a single presentation
set at about 5 or 6 sentences and QA on the surface meaning of the story is sufficient.
The QA function generates as many questions about the story as possible [3], and then,
selects a suitable and purposive question.
for ICALL Systems
Hidenobu Kunichika*1, Minoru Urushima*2, Tsukasa Hirashima*3 and Akira Takeuchi*2
*1
Dept. of Creation Informatics, Kyushu Institute of Technology, Japan
*2
Dept. of Artificial Intelligence, Kyushu Institute of Technology, Japan
*3
Dept. of Information Engineering, Hiroshima University, Japan
Abstract. Language training systems that provide learners adaptive questions on the
contents of stories require several capabilities such as semantic analysis, automated
question generation and diagnosis of learners' answer sentences. This paper presents a
method of selecting questions from a generated list to realize adaptive questions and
answers. Our method filters out similar questions, and then selects questions by
considering the difficulty, types and order. This paper also describes an evaluation of
our method. As the result of our experiment demonstrates, we have found that our
method generates a viable series of questions.
1. Introduction
It is common in second language learning, to answer questions on the contents of passages
after listening and/or reading them. Such questions and answers (QA) in the target language is
effective for acquiring practical skills because multiple language skills are required to answer
the questions, in particular to grasp the contents of the story and the questions, as well as to
compose answers. Many computer assisted language learning systems have been developed
[1]. Some are equipped with exercise functions which ask about the contents of sentences.
Most, however, use questions prepared beforehand [5, 6, 7]. Thus these have the problem that
such systems will present questions without considering the learner's level of understanding
because the number of prepared questions is limited.
In order to solve these problem, we are aiming to realize a QA function which provides
adaptive questions on the surface semantics of English stories prepared by authors or learners.
To realize the QA function, the following sub-functions are necessary: (1) to understand
English sentences and to extract syntactic and semantic information, (2) to generate
automatically various kinds of question sentences for presentation to learners who have
varying degrees of comprehension, (3) to select adaptive questions from a set of generated
question sentences, (4) to analyze learners' answer sentences and to diagnose errors and (5) to
offer intelligent help for the correction of errors and the acquisition of correct knowledge by
referring to the student models. In earlier studies, we have already implemented the sub-
modules for the functions (1), (2), (4) and (5). This paper proposes an adaptive method of
selecting questions for the function (3).
2. The outline of the QA function
Our QA function gives learners questions about the contents of a story. After studying the
contents of the story by reading and/or listening, they answer the questions. Aims of our QA
are both to train for conversation by using multiple skills through reading a story, listening to
or reading questions and composing answers and to give learners a chance to realize their own
state of understanding of, for example, vocabulary and grammar; and to practice usage
through QA. To reduce the burden of memorizing the content of the story and to concentrate
on composing sentences from memory, the length of any one passage in a single presentation
set at about 5 or 6 sentences and QA on the surface meaning of the story is sufficient.
The QA function generates as many questions about the story as possible [3], and then,
selects a suitable and purposive question.
