See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220902605

Math word question generation for training the students with learning
difficulties

Conference Paper · January 2011

DOI: 10.1145/1980022.1980069 · Source: DBLP

CITATIONS READS
12 448

2 authors:

Nandhini Kumaresh Balasundaram S.R.

Central University of Tamil Nadu National Institute of Technology Tiruchirappalli
23 PUBLICATIONS 415 CITATIONS 85 PUBLICATIONS 438 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Balasundaram S.R. on 09 November 2020.

The user has requested enhancement of the downloaded file.

 2nd International Conference and workshop on Emerging Trends in Technology (ICWET) 2011
Proceedings published by International Journal of Computer Applications® (IJCA)

Techniques for Generating Math Word Questions for

Learning Disabilities
K.Nandhini S.R.Balasundaram
Research Scholar Associate Professor
Department of Computer Applications Department of Computer Applications
National Institute of Technology National Institute of Technology
Tiruchirapalli-620015- India Tiruchirapalli-620015-India

ABSTRACT situations will help them in mastering their problem solving skills.
Math word problems resembles daily life situation. Solving math Instead of going for obsolete math word problems, problems that
word problem is a challenging task for normal students. It is still suit their need will be helpful to them.
more complex for those having problem in learning. Mastering in
solving math word problem can help them in leading daily life
activities. Technology plays a vital role in education for all, 1.1 Learning Disability
including the students being discriminated in normal schools due Not all learners have same capability. The degree of learning
to their different way of learning abilities.. The learning problem capability varies from learner to learner. About 5% to 8 % of the
cannot be cured but can be reduced by giving proper training in students in a class can have learning problem which can be
the required domain through an instructional strategy. Training identified by their poor performance in academic activities.
depends on number and type of questions. Due to the limitations Children with learning problems may have difficulty in one or
of getting questions from standard text books, the need for more of the following areas: memory, cognitive development, and
Automatic Question Generations becomes vital. This paper visual-spatial ability. Identifying the problem and giving proper
discusses two methodologies for question generation related to instructive or compensatory methods will reduce the degree of the
math word problems in turn can improve the training process.. problem. Problem in reading and carrying out basic computation
may account for the student‟s poor performance. Difficulties in
problem representation, the failure to identify relevant information
Categories and Subject Descriptors and the failure to perform proper operation may still affect their
H.5.2 User Interfaces- Training, Help and Documentation. performance [2]. Learning problem cannot be cured but can be
F.4.2 Grammars and Other Rewriting Systems problems-Grammar reduced by giving proper training in the domains where the
types (e.g., context-free, context-sensitive) learners lack.

General Terms 1.2 Math Word Problems

Human Factors, Languages. Word problem in mathematics resembles the daily life situation.
For example, consider a problem statement, “Mala has 5 biscuits,
Keywords her mother gave 3 more, and How many Biscuits Mala have
Learning problem, Math word problem, Automatic now?”. Here, the simple arithmetic operation addition is given in
Question generation, Shallow NLP based Template, terms of words, which every individual has to understand.
Context Free Grammar. Traditional teaching methodology exploits the rote memory in
solving word problems where the keyword present in the question
1. INTRODUCTION plays major role in identifying the arithmetic operation rather than
understanding the meaning of the problem. Training the students
Mathematics is integral component for all areas of daily life.
using keyword or procedural instruction will only improve rote
Solving math word problem is more significant for all, as it
learning and not the semantic learning. As of now there are eleven
resembles daily life aspects. When solving math word problems is
different meanings of addition and subtractions have been defined
sometimes challenging for even normal students, people with
learning difficulty may find it still more difficult. Children with [2]. There are four different problem types, ‘change‟,‟ „group‟,
learning difficulties are given training to solve such problems and „compare‟ and „part whole‟ that characterize most of the addition
this in turn may help them to lead independent life. Students with and subtraction problems. In each of them, three more subtypes
learning disabilities often exhibit serious deficits in mathematics. such as „Result unknown „,‟Compare unknown‟ and „Initial
Apart from that, ineffective instructional strategies may still unknown‟ exist [6].
worsen their poor problem solving capability. The students with 1. John has 5 Marbles. Jim gave him 3 more. How many marbles
disabilities experience considerable difficulty with problem does John have altogether?
representation or identifying relevant information, along with 2. John had 5 marbles. Jim gave him some more. Now John has 8
difficulties in reading and computation. By giving proper training marbles. How many did Jim give him?
through an instructional strategy the problem can be reduced. The
3. John had some marbles. Jim gave him 3 more. Now John has 8
number of word problems with varying complexities and

45
 2nd International Conference and workshop on Emerging Trends in Technology (ICWET) 2011
Proceedings published by International Journal of Computer Applications® (IJCA)

marbles. How many marbles did John have to begin with? idea is to generate the questions automatically based on question
Identification of “type of the problem” and the „Unknown type” templates which are created by training on many medical articles
are major factors in deciding the understandability level of the [18]. Andrenucci introduced a template based approach to
learners. The level of difficulty in solving word problems is generate questions [8]. The literature highlights very less research
influenced by the specific words used in the problem [9,10] and on math word problem question generation. Question generation
the extent to which the problem is embedded in contextually retaining the semantics is a challenging task. In this paper, two
realistic situations. Also the presence of extraneous information methodologies are discussed, one based on templates using
[11], and the number and type of calculations required to solve shallow NLP technique and the other based on Context free
the problem further increase the level of difficulty. Accelerating grammar.
the learners to perform operations based on semantics is the focus
of the work. Ineffective instructional strategies may worsen the 3. TRAINING METHODOLOGY
performance of students with learning disabilities. Schema-based A kind of pretest for assessing the learner‟s ability level is done
strategy instruction is known to benefit both special education by giving simple arithmetic problems to solve, or to test the
students [19, 20] and students at borderline in solving math word prerequisite in solving the math word problem like addition,
problems [18, 17]. subtraction, multiplication and division. After identifying the
difficulty of the learner, he/she is exposed to the training session
1.3 Automatic Question Generation that comprises of rigorous training. This training is based on
Question Generation is the task of generating reasonable automatic question generation phase followed by evaluation. The
questions. Question generation (QG) can be divided into deep overall flow is shown in Figure 1.
QG, such as‟ why‟, „what-if‟, „how‟ questions and shallow QG
such as „who‟, „what‟, „when‟, „where‟, „which‟,‟ „how
many/much‟, „yes/no‟ questions [23]. Methodologies to generate
questions can be broadly classified into categories such as
template-based [22], syntax-based [25 & 21] and semantics-based
[24]. Template-based approaches are mostly suitable for
applications with a special purpose. Syntax-based approaches are
rather effective, especially for short sentences. Semantic-based is
theoretically more interesting and practically challenging. For
training the students with learning difficulty, more number of
questions is needed which can also be taken from question bank.
Question bank based training/evaluation often loses its freshness
over a period, unless the questions are updated from time to time.
In this paper we propose an alternate method wherein questions
are generated from a template. Whenever a request is made to the
template, a new instance of the question is generated. The concept
is implemented for training /testing the students with difficulty in
solving math word problems.
Figure 1. Training Process
2. RELATED STUDY From the he user, the arithmetic competence level and reading
Mastering in math word problem solving can be achieved through difficulty level are given to Question Generation Engine. The
rigorous training. Training depends on number of exercises engine uses both template-based and grammar-based techniques
solved. Text books does not contain all types of Word problems.. to generate math word problems. The student is given training
The volume of exercises available with these materials may be using the generated questions through schema based approach.
less. Not all types of exercises are given equal importance. This And at the end of training phase the learner undergoes evaluation
motivation leads to apply technology for automatic math word to verify whether training is to be continued or the difficulty level
problem generation. Acquiring skills to solve math word problem has to be increased/decreased depending upon the performance.
can be achieved through training. Computer aided training has a The situation word problem usually consists of complex sentences
positive impact in promoting reading skills of the children. To an in turn having several simple sentences. Question generation
extent, technology has enhanced language processing and memory techniques such as shallow NLP template and grammar based
of the children [10]. Quite a few have proposed various approaches techniques are discussed in forth coming sections.
for solving word problems [12, 13]. The schema based strategy
was effective in substantially increasing the performance in word 3.1 Shallow NLP Based Templates
problems [4]. Problem solving through graphic representation
provides a prerequisite to understand and organize information for The Template based approach exploits shallow NLP technique to
later problem solution and to facilitate conceptual understanding generate more number of similar questions. The flow diagram is
[3]. When training depends on the availability of more questions, given in Figure 2. The sample question is given as input to the
then automatic question generation is a key component to achieve generator, the sentences in the questions are separated and then
it. Tao Li focuses on generating questions with multiple variables the separated tokens are attached with Part Of Speech tagging. To
on DAG and further extends it with a complex algorithm that will generate similar type of questions the noun phrases in the sample
generate more questions [5]. Questions generated are more questions are to be replaced by some other nouns from the list.
complex and require more mathematical skills to solve. The main And the questions can be modified using similar word or

46
 2nd International Conference and workshop on Emerging Trends in Technology (ICWET) 2011
Proceedings published by International Journal of Computer Applications® (IJCA)

synonyms of verb occurrences. [('John', 'NNP'), ('has', 'VBZ'), ('3', 'CD'), ('marbles.', 'NNP'),
('Jim', 'NNP'), ('gave', 'VBD'), ('him', 'PRP'), ('2', 'CD'), ('more.',
'NNP'), ('How', 'WRB'), ('many', 'JJ'), ('marbles', 'NNS'), ('does',
'VBZ'), ('John', 'NNP'), ('have', 'VB'), ('now', 'RB'), ('?', '.')]

A set of sample questions generated using template based

approach are given below:
1. Raja has 4 books, Ram gave him 5 more, How many
books does Raja have now?
2. Sonu has 3 bangles, Susma gave her 2 more, How many
bangles does Sonu have now?
3. Vishnu has 4 ben10 toys, Prakash gave him 2 more,
How many ben10 toys does Vishnu have now?

Although the basic concept behind using question templates for

generation of the questions is quite old , a generator using such a
grammar can ensure high-quality output provided that the input
semantic representation is reasonably coherent. The grammar
based question generation is discussed in the following section.

3.2 Context Free Grammar Approach

Context-free grammars are fundamental structures in computer
science. Recently, grammar engineering has been recognized as an
emerging field of software engineering [14]. Techniques for
generating sentences form a basic building block for grammar
engineering [15]. Context free grammars consist of rules as a
series of productions [1]. The aim of the work is to use the
production rules to generate math word problems. For the sake of
discussion, every question is defined in terms of three line word
problems for addition and subtraction. The question is separated
as three sentences, each with different set of production rules.
The production rules are stored in separate files. A parser
processes the production rules recursively in top down fashion
and builds sentences. Three types of nouns namely masculine
nouns such as {Ravi, Raju, Dilip, etc}, feminine nouns such as
{Mala, Meera, Mathu, etc.}, and neuter nouns such as {Marbles,
Biscuits, Toys, etc} are considered.
The Grammar Based Algorithm for Automatic Question
Generation is as follows
Algorithm
Procedure question-generation-CF (Production File)
/* Production File -contains set of production rules*/
Begin
Figure 2. Flow Chart 1. Call a method to add rule and to expand in specific order
using a context free object
The numeric values are replaced using random number generator 2. Expand the values using nouns till terminals
according to the complexity level. 3. Display the question using a suitable user interface
The output at each stage is given below. End
John has 3 marbles. Jim gave him 2 more. How many marbles
This approach preserves the semantics of the questions and aims
does John have now?
at generating different patterns of questions. The syntax trees for
['John has 3 marbles.', 'Jim gave him 2 more.', 'How many marbles sample questions are as follows (Figure [3-5]).
does John have now?']
Questions:
['John', 'has', '3', 'marbles.', 'Jim', 'gave', 'him', '2', 'more.', 'How',
1. John has 5 marbles, Jim gave him 3 more, and How
'many', 'marbles', 'does', 'John', 'have', 'now', '?']
many marbles does John have altogether?
2. How many marbles does John have now, if John had 5

47
 2nd International Conference and workshop on Emerging Trends in Technology (ICWET) 2011
Proceedings published by International Journal of Computer Applications® (IJCA)

marbles and Jim gave him 3 more?

3. Jim gave 3 more marbles to John, where John already Table 1.Questions and its types
has got 5 marbles. How many marbles does John have?
Qn. Question Type -Subtype
Type

1.1 John has 5 Marbles. Jim gave him 3 Join- Result

more. How many marbles does john
Unknown
have altogether?
1.2 John had 5 marbles. Jim gave him Join-Change
some more. Now john has 8 marbles. Unknown
How many did Jim give him?
1.3 John had some marbles. Jim gave him Join-Initial
3 more. Now john has 8 marbles. Unknown
How many marbles did john have to
begin with?

Figure 3. Syntax tree 1 4. RESULT AND DISCUSSION

In Generating questions using template approach, the number of
valid questions might be less due to the stereotyped nature and the
problems in noun and pronoun confusion. Moreover the generated
questions have to be validated manually by an instructor for
validity and for grammatical correctness. A deep natural language
experiences may not be seen in this approach. Whereas the
context free grammar, generates more number of valid questions if
the production rules are framed properly. A comparison of the
techniques based on the number of valid questions generated is
shown in Figure 6.

Figure 4. Syntax tree 2

.
Figure 6.Comparison graph
The learner is said to be competent enough for answering the
question only if he/she has chosen the correct problem type and
subtype of the answer and the correct answer. The figure 7 shows
the interface displayed to the user when they undergo training.

Figure 5. Syntax tree 3

The production rules and order of parsing are slightly modified to
accomplish the task. A sample set of questions retrieved from [6]
are shown in Table 1.
Figure 7. Evaluation Test Sheet

48
 2nd International Conference and workshop on Emerging Trends in Technology (ICWET) 2011
Proceedings published by International Journal of Computer Applications® (IJCA)

The understandability of the student can be measured by the

option selected during the test process through this interface. If
the problem type, the unknown type chosen are correct and the
answer given are also correct, then it indicates that the learner has
solved the problem after understanding. The understandability has
been improved when compared to previous training. If the option
chosen is wrong and the answer given is correct, then it means
that he/she lacks in understanding but good in arithmetic
operation. The objective of the training is to improve the problem
solving capability through semantic training. Figure 8,9 and10
show the performance of a student in solving a set of questions
after training in three types of word problems such as Group,
Change and Compare type. It is clear that arithmetic skills can be
improved through drill and practice and it takes more training and
practice to understand the semantics and solve the questions. Figure 10. Compare Type
A sample group of 15 students having learning difficulty studying However the student‟s performance is not only based on the
third to fifth grade were given training through template based question generation techniques. It only helps in mastering math
approach and were assessed for improvement. The result shows word problems. The key factor behind the success is because of
that there was a significant improvement in solving the number of the instructional strategy that is advised to be followed. The
questions correctly. But basically the question generation samples in this group are taught to solve the problem using
technique itself lacks natural language approach. Hence solving Schema based approach.
approach is simply keyword dependent. Table 3 shows the comparison of two approaches to generate math
words problems.
Table 2. Comparison of Question Generation Techniques
Shallow NLP Template Grammar Based Question
generation
Easy technique to generate Framing production rules for
similar type of questions each type is complex
Pronoun confusion may occur, Pronoun confusion can be
while replacing masculine, easily handled by proper
feminine and neuter nouns. production rules.
Questions generated are Question generation is based
basically of keyword approach on natural language approach
type, often misleading the which can be easily applied to
learner to rote learning. real world activities.

Figure 8. Change Type Training the difficulty learners using questions generated by
grammar based approach improves semantic learning. The
understandability level also has significant improvement while
solving the problems correctly.

5. CONCLUSION
Question generation is an essential component of learning.
Especially generating math word questions involves the need for
natural language interface. In this direction this paper has
discussed methodologies to generate various questions for people
with learning math word problems. Rigorous training is essential
to improve their skill level. However the instructional strategy
plays a key role in learning to solve math word problems.

6. REFERENCES
[1] John E. Hopcroft, Rajeev Motwani, Jeffrey D.Ullman,
Figure 9. Group Type 2001, Introduction to Automata Theory, Language, and
Computation, Pearson Education, Asia.

49
 2nd International Conference and workshop on Emerging Trends in Technology (ICWET) 2011
Proceedings published by International Journal of Computer Applications® (IJCA)

[2] Van de Walle, J.A. 2004, Elementary and Middle School Transaction on Software Engineering and Methodology
Mathematics: Teaching Developmentally, 5th Edition. New 14(3) 331–380.
York: Addison Wesley Longman, Inc. [15] Lammel, R, Grammar Testing, 2001, In: Proceedings of
[3] Asha Jitendra, Teaching Students Math Problem. 2002, Fundamental Approaches to Software Engineering
Solving Through Graphic Representation, Teaching (FASE‟01). Volume 2029. 201–216.
Exceptional Children, Vol 34, No 4, pp 34-38. [16] W. Wang, H. Tianyong, L. Wenyin, 2008, Automatic
[4] Asha Jitendra ,Caroline M.Dipipi,Nora Perron Jones. 2002, Question Generation for Learning Evaluation in Medicine,
An Exploratory Study Of Schema Based Word- Problem, LNCS Volume 4823, 2008, pp. 242-251.
Solving Instruction For Middle School Students With [17] Jitendra, A.K., DiPipi, C.M., & Grasso, E. 2001, The Role
Learning Disability, The Journal of Special Education, Vol Of A Graphic Representational Technique On The
36, No 1, pp 23-38. Mathematical Problem Solving Performance Of Fourth
[5] Tao Li and Sam Sambasivam, Automatically Generating Graders: An Exploratory Study. Australian Journal of Special
Questions In Multiple Variables For Intelligent Tutoring, Education, 25(1& 2), 17-33.
Issues in Information Sciences and Information Technology. [18] Jitendra, A.K., Griffin,C. C., McGeoy, K., Gardill. M.C.,
[6] Peterson, P.L Fennema, E. & Carpenter, T. 1989, Using Bhat, P.,& Riley, T. 1998, Effects Of Mathematical Word
Knowledge Of How Students Think About Mathematic. Problem Solving By Students At Risk Or With Mild
Educational Leadership, 46(4), pp 42-46. Disabilities. The Journal of Educational Research,91, 345-
[7] Yao –Ming Yen,Wen-Tsung Pai,Wen-Jeng Tzeng,Chin-Chin 355.
yen,Chi-Nung Chu,Yun-Lung lin, 2007, Research On [19] Jitendra, A., & Hoff .K. 1996, The Effects Of Schema-
Reasoning And Modeling Of Solving Mathematics Situation Based Instruction On Mathematical Word Problem Solving
Word Problems Of Primary School Proc. 9th IEEE Performance Of Students With Learning Disabilities. Journal
international Symposium on Multimedia. of Learning Disabilities 29, 422-431.
[8] A.Andrenucci, Sneiders,E. 2005, Automated Question [20] Jitendra. A., Hoff,K., & Beck,M.M., 1999, Teaching Middle
Answering :Review Of Main Approaches, Proc. 3rd School Students With Learning Disabilities To Solve Word
International Conference on Information Technology and Problems Using A Schema Based Approach . Remedial and
Applications Vol 1, pp 514-519. Special Education, 20, 50-64.
[9] Briara, D. J., & Larkin, J. H. 1984. An Integrated Model Of [21] M. Heilman and N. A. Smith.2009, Question Generation via
Skill In Solving Elementary Word Problems. Cognition and Overgenerating Transformations and Ranking. Technical
Instruction, 245-296. report, Language Technologies Institute, Carnegie Mellon
[10] Helwig, R., Rosek-Toedesco, M. S., Tindal, G., Heath, B., & University Technical Report CMU-LTI-09-013.
Almond, P. J. 1999, Reading As An Access To Mathematics [22] Jack Mostow and Wei Chen. 2009, Generating Instruction
Problem Solving On Multiple-Choice Tests For Sixth-Grade Automatically For the Reading Strategy Of Self-
Students. Journal of Educational Research, 93, 113-125. Questioning. In Proceeding of the 2009 conference on
[11] Parmar, R. S., Cawley J. F., & Frizita, R. R. 1996, Word Artificial Intelligence in Education, Amsterdam, The
Problem-Solving By Students With And Without Mild Netherlands.
Disabilities. Exceptional Children, 62, 415-429. [23] V. Rus and A. Graesser, 2009, The Question Generation
[12] Agaliotis, I. 2008, Strategies for Successful Word Problem Shared Task and Evaluation Challenge. 2009.
Solving By Students with Learning Disabilities. Strategies [24] Ivan A. Sag and Dan Flickinger., 2008, Generating
for Successful learning, 1(4). Questions with Deep Reversible Grammars. In Proceedings
[13] Fuchs, L. & Fuchs, D, 2007, Mathematical Problem Solving: of the First Workshop on the Question Generation Shared
Instructional Interventions. (pp. 397-414). Task and Evaluation Challenge, Arlington, VA: NSF.

[14] Klint, P., Lammel, R., Verhoef, C, 2005, Towards An [25] Brendan Wyse and Paul Piwek. Generating Questions from
Engineering Discipline For Grammar Ware ACM Openlearn Study Units. In Proceedings of the 2nd Workshop
on Question Generation In Craig,

50

View publication stats