“Majorly Adapted Translator”: Towards Major Adaptation in ITS

Tina Daaboul and Hicham Hage

Department of Computer Science, Notre Dame University - Louaize, Zouk Mosbeh, Keserwan, Lebanon

Keywords: Culturally Aware Learning System, Major Adaptation, Information Extraction, Intelligent Tutoring

Systems, Relation Extraction.

Abstract: Culturally Aware Learning Systems are intelligent systems that adapt learning materials or techniques to the

culture of learners having different “country, hobbies, experiences, etc.”, helping them better understand the

topics being taught. In higher education, many learning sessions involve students of different majors. As

observed, many instructors tend to manually modify the exercises several times, once for every major to adapt

to the culture, which is tedious and impractical. Therefore, in this paper we propose an approach to making

learning sessions adaptable to the major of the learner. Specifically, this work introduces an Artificial

Intelligent system, “Majorly Adapted Translator (MAT)”, which aims at translating and adapting exercises from

one major to another. MAT has two main phases, the first identifies the parts of an exercise that needs changing

and creates an exercise template. The second translates and adapts the exercise. This work, highlights the first

phase, the Feature Extract phase, which relies on our own relation extraction method to identify variables

which extracts relations specific to named entities by using dependency relations and shallow parsing.

Moreover, we report the performance of the system that was tested on a number of probability exercises.

1 INTRODUCTION

In recent years, culture became widely adapted

especially on the level of educational technology,

since an e-learning session can easily involve people

from different countries and cultures. A shift was

done in e-learning to become more culturally aware;

specifically in intelligent tutoring systems (ITS),

which are named “Culturally Adapted Tutoring

Systems (CATS)”.

In the case of CATS, the culture being adapted is

the “social culture” since it mainly involves factors

on the social level such as country. However,

learners taking the same topic might be from

different majors, in this case adapting to the “social

culture” of the learner is not enough, and the major

of the learner should be considered as well. As

suggested by Carnegie Mellon University, if a

learning session includes students of different

majors, it is preferable to either split them into

sections or to introduce, for each group, examples

that are “relevant to the major or appropriate to the

students’ ability” (Carnegie Mellon University,

2015).

Moreover, one of the common questions students

ask is “when will I ever use this in the real world?”

(Briggs, 2014). Students are usually more concerned

in knowing how a certain learning material applies

to their major rather than other majors. For example,

a civil engineering student uses the probability topic

in expecting how much capacity a large container

can hold (Prudchenko, 2017) whereas a computer

science student uses the same topic in determining

how a certain program will act (The University of

Chicago, 2017). This is important because learning

material should be designed in a way that has a

lifelong effect on students and prepares them to their

future career as mentioned by Kneale (Kneale,

2009). Several researchers highlight the importance

of making learning material relevant to the student’s

major. As (Azi et al., 2008) mentions, most people

learn by relating the material to what they previously

know. Relevance is a key component to motivate

learners and help them maintain a good memory of

the material they are learning. The neurologist Judy

Willis gives an example of memorizing long

vocabulary words, she says that if they “don’t have

personal relevance or don’t resonate with a topic

about which the student has been engaged, they are

likely to be blocked by the brain’s affective (or

emotional) filters” (Briggs, 2014).

Currently, in multi-disciplinary classes, adapting

Daaboul, T. and Hage, H.

“Majorly Adapted Translator”: Towards Major Adaptation in ITS.

DOI: 10.5220/0007296904510457

In Proceedings of the 11th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2019), pages 451-457

ISBN: 978-989-758-350-6

451

to the major is taken into consideration by human

instructors, were some tend to manually modify the

same exercise several times, in order to cover all the

contexts to suit learners from different majors.

Despite helping learners better understand the

concepts being taught, this method has several

shortcomings: students do not relate to all the

examples, it is very tedious, time consuming, and

not yet supported by e-learning systems.

The purpose of this paper is to lay the

foundations for an ITS, to adapt the topics being

taught to the major of the students in an automated

and efficient way. For this, we propose “Majorly

Adapted Translator (MAT)” an intelligent system

that “translates” exercises authored by an instructor

from one concept to another while maintaining the

same structure of the exercise. This system contains

two phases, the first is “Feature Extract (FE)” which

identifies the parts of an exercise that need to be

translated and adapted to other concepts and the

second is “Translate” which is used to translate the

exercise into other concepts. This paper highlights

the initial phase of the system FE which is based on

the Natural Language Processing (NLP) method

“Information Extraction”. The main purpose of FE

is to identify the parts that can be translated and

obtain the mathematical form of the exercise to be

used in the next phase “Translate” for translating the

exercise to other concepts. For this purpose, we have

created our own “Relation Extraction (RE)” method

that extracts relations of domain specific words by

relying on their dependency (grammatical) relations

with the rest of the words in the text. Currently,

MAT focuses on translating exercises written in

human language related to the “Probability” branch

of mathematics, since this is a common domain

taught to students of different majors such as

mathematics, biology, business, computer science,

and engineering, to name a few. The paper is

organized as follows: the next section provides an

overview of CATS, Section 3 details our approach,

Section 4 highlights the testing and results, and

Section 5 concludes this work and presents the

future works.

2 STATE OF ART

Culture or “the programing of the mind” (Hofstede,

1997) affects the way people think, act, and even

their understanding of certain matters, which all

goes back to what is in their cultural background. In

early stages, culture in the subdomains of e-learning

was adapted on the level of nation by taking the

“human culture perspective” which is based on

Hofstede’s method who studied the dimensions of

culture in organizations from the human’s

perspective (Hofstede, 1997). These dimensions

represent emotions such as pride, teamwork spirit, or

ability to accept criticism which are common among

certain cultural groups. They are used as the bases

for dealing with the learner since they are assumed

to be inherited from the learner’s cultural group in

which he/she unconsciously acts upon when

interacting with the system (Blanchard and Frasson,

2005) (Vartak et al., 2008). “Culturally Adapted

Tutoring Systems (CATS)” is an example of a

system that relies on Hofstede’s method. Other than

adapting culture in learning techniques, several

systems adapted culture in the pedagogical resources

given to the learner such as in mathematical tutoring

systems (Melis et al., 2009) and authoring tools

(Vartak et al., 2008). In later stages, scientists

claimed that it is not enough to rely only on the

“nation” culture as this results in many “cultural

stereotypes” (Ogan et al., 2014). Thus, other factors

should be considered as well such as “technological

factors” (Nye, 2014), “collaborative filtering” (Eboa

et al., 2010) and “Instructional Cultural

Contextualization (ICON)” (Mohammed and

Mohan, 2014) which learns from analyzing the

learner’s preferences. Later, (Gasparini et al., 2010)

introduced an e-learning system that is more learner-

personalized, i.e., it considers factors such as

personal, culture, technological, and pedagogical

perspective of the user (how much he/she knows

about the topic). Currently, many systems adapt the

personalization concept in e-learning making a

learning session more student-centered such as

(Khemaja and Taamallah, 2016) and (Klašnja-

Milićević et al., 2017). However, to the best of our

knowledge, no work has been done to adapt to the

major of the learner.

3 APPROACH

“Majorly Adapted Translator (MAT)” contains two

major phases. The first “Feature Extract (FE)” which

identifies the parts of the exercise that needs to be

translated and adapted to other concepts. The

second, “Translate” translates the exercise into other

concepts. In this work we highlight the first phase,

“Feature Extract”. Specifically, Feature Extract

defines the structure of an exercise through

transforming it into a template. This template

includes the list of variables and the mathematical

form of the exercise. This template will later be used

ICAART 2019 - 11th International Conference on Agents and Artiﬁcial Intelligence

452

in order to translate the exercise.

Consider the following sentence: “4% of the

resistors are defective”. The value 4% should remain

unchanged, whereas “resistors are defective” should

be contextualized and adapted. We define a variable

as “a core part of an exercise that has stable

keyword(s) having a related dynamic value which

could be changed without affecting the structure of

the exercise”. A variable is formed of a “keyword”

(the “4%” in the previous example) which is an

expression that has a mathematical implication

depending on the domain of the exercise and a

“value” (the “resistors are defective” in the previous

example) that refers to all the words linked to the

keyword in a way that defines it. The aim is to find

the keyword, look for its value and contextualize it

into other concepts.

Identifying variables is obtained by extracting the

relations of the keywords. However, the challenge is

for MAT to be able to find the “correct” value of the

keywords since not all terms linked to a certain

keyword are its correct value. In this paper, we

contributed in creating a “Relation Extraction (RE)”

method which extracts relations specific to named

entities, by relying on the dependency relations of

these named entities with other words in the text. As

shown in Figure 1, FE undergoes several steps,

detailed in the rest of this section. Note that we

consider an exercise as composed of two parts: the

given where all the details are provided, and the

questions that the learner must answer.

Figure 1: Feature Extraction Process.

3.1 Tokenization

The initial step in FE is “Tokenization”; it divides

the “given” of the exercise into “chunks of

words/tokens” which are a group of words

belonging to the same grammatical type. FE

identifies chunks as either a “keyword chunk” or a

“non-keyword chunk” and labels each with its

grammatical type which is used as a key to identify

the relations among the chunks. Like any other

Information Extraction (IE) task, the initial step is to

split the text into smaller parts, i.e., splitting the

given into sentences and then to clauses in which

each undergoes the process of identifying the

variables. Splitting into clauses is helpful since the

smaller the text, the easier it is for the dependency

parser, used in the next step, to extract accurate

dependency relations. Next each clause is divided

into tokens using a “word tokenizer” and the part-of-

speech tag of each token is retrieved which will be

used to group them into chunks.

Next, the Stanford NER (Named Entity

Recognition) tool is used to identify the “Keyword

Chunks”. Stanford NER is a statistical parser created

by Stanford University, trained on previously

annotated text that labels sequence of words

referring to “name of something” (Manning et al.,

2014) (Finkel et al., 2005). Since NER is a domain

specific tool, it was required that we customize the

Stanford NER to identify keywords related to the

domain of MAT, i.e., mathematical or probability

expressions. Stanford NER can identify numbers

such as percent, integer, and money. Whereas other

types of numbers and mathematical/probability

expressions that Stanford NER cannot read, are

identified using “Stanford Regex NER”; a rule based

NER system that uses Java regular expressions in

order to label named entities. Accordingly, we have

trained the Regex NER to identify specific

probability expressions which we created by

developing our own Java regular expressions such

as: “2 out of five” is identified as “2/5”, and the

words “probability, random, standard deviation,

average, median, etc.” are identified as keywords.

Furthermore, all the chunks that are not

identified as keyword chunks are considered non-

keyword chunks and parsed using a shallow parser.

Chunking non-keywords at early stages is important

because it simplifies the work of analyzing the

relations between a keyword and its value. Consider

the phrase “20 have a laptop computer” where “20”

is the keyword; instead of obtaining the relations:

“20” linked to “laptop” and “laptop” linked to

“computer”; having “laptop computer” as one chunk

of type “Noun Phrase”, directly leads to the relation

that “20” is linked to “laptop computer”. In order to

obtain similar forms of chunks such as noun phrases

and verb phrases, we developed our own grammar

customized for MAT to perform the shallow parsing.

This grammar contains “tag patterns” which were

derived from studying various grammatical rules and

from looking at frequently encountered chunks

“Majorly Adapted Translator”: Towards Major Adaptation in ITS

453

related to a value of a variable found in

mathematical text. Finally, we reserved for each

chunk, its type, the original token value, its

normalized form, and most importantly the index of

each token in the chunk in order to link it later on

with the indexes of the dependency relations.

3.2 Retrieve Dependency Relations

After identifying the keyword chunks and the non-

keyword chunks, the second step is to extract the

dependency relations among these chunks. This step

relies on the Stanford Core NLP dependency parser

(Manning et al., 2014). It uses its “enhanced”

dependency relations that consists of a governor (or

the head) and a dependent that depends on the

governor. In order to analyze the dependency

relations between chunks, we studied various

dependency type relations. One of the most

important relation types to mention is the “core

argument” type which is a relation that includes a

subject and an object (Universal Dependencies

Contributors, 2016). Consider the phrase “three have

defects”, the following relations are obtained:

“three” is the subject of “have” represented as

nsubj(have,three) and “defects” is the object of

“have” represented as dobj(have,defects). This step

helps the system in the next stage to link certain

chunks to the keywords according to their

dependency relations, forming up a “variable”.

3.3 Identify Variables

This step is the most important step in FE, as it

detects the parts of an exercise that needs to be

translated into other concepts, i.e., the “variables”.

“Identify Variables” takes as an input the list of

chunked phrases obtained from “Tokenization” and

the list of dependency relations determined by the

“Retrieve Dependency Relations” step. As

mentioned previously, not all relations of a keyword,

are supposedly its value. Hence, the challenge is to

obtain the correct value of the variables; FE

performs this through three main steps:

First, “Getting the linked Relations” extracts all

the linked relations of the keyword(s) and omits

those that are not. A chunk can be directly linked to

a keyword chunk or indirectly through other chunks

linked to the ones related to the keyword. Moreover,

a value can be found in more than one chunk, so the

dependency relations among these chunks are to be

considered as well. All the dependency relations

obtained from this step are called “linked relations”.

Moreover, MAT considers a predefined priority list

of keywords; this list was carefully created based on

scanning similar examples, such that it gives the

shortest path leading to the value and accurate

results. Accordingly, the system starts with the

priority keyword, gets all its directly linked relations

and then completes to get the relations linked to the

direct chunks, i.e., the indirect relations. This

process stops when the system cannot find linked

relations any further. Consider the phrase “Eleven of

the motors are free of defects” where “eleven” is a

keyword of type “Number” and “free” is a keyword

of type “Negation”. According to the predefined

priority list, MAT starts with “eleven” and obtains

the direct relations: “eleven” is the nominal of

“motors” and “free” is the subject of “eleven”. Since

“free of defects” is one chunk, thus “eleven” is

linked to “free of defects”. The indirect relation

obtained is that the verb “are” is the copula that

connects to the subject “free” and “free” is the

nominal of “defects”.

Second is the “Translating Relations” step which

helps the system understand the relations among the

dependencies and how to use them as instructions to

identify the variables. It does that by translating the

linked relations based on “the type of relation” and

on the “type of the related chunks”, as well as

eliminating relations that are considered useless. In

the previously mentioned example, “Eleven of the

motors are free of defects” the obtained dependency

relations are translated as:

1. “nmod:of(11,motors)” “Number is

NounPhrase”

2. “nsubj(11,free)”  “Number has Subject

Negation”

3. “cop(free,are)”  “Negation has Equality”

4. “nmod:of(free,defects)”



“Negation has

NounPhrase”

Third, “Variable Identification” aims to identify

the correct value of a keyword by understanding the

translated linked relations obtained from the

previous steps. FE defines an algorithm that extracts

the “value” for different types of keywords.

Essentially, while searching for the “value”, the

system undergoes two levels of search, the first

searches in the direct relations and the second in the

indirect relations. Throughout both levels of search,

the system performs three actions: “Search”,

“Continue”, and “Add”. Based on the type of the

related chunks and the type of the relation, the

system decides whether to add the word as a value

or ignore it and continue the search. As the system

analyzes the dependency relations, it relies on

general grammatical rules inspired from previously

ICAART 2019 - 11th International Conference on Agents and Artiﬁcial Intelligence

454

chunked variables of similar structure. For example,

when a keyword is linked to a “verb” by the relation

“subject”, the system understands that there should

be an “object” so it continues to search for it, if the

encountered object is a noun, then it is added to the

“value”. In the previously mentioned example, the

keyword “11” is linked to a negation “free” by the

relation subject, so FE adds “free” to the value and

continues to add the chunks related to “free” as part

of the value, forming up the value “free of defects”

for the keyword “11”. Furthermore, when MAT

identifies the value, it defines the type of it. A value

can be a regular value type (denoted as “Value”) or

of the types: “Key Is”, “Number”, or “Relation (i.e.,

negation or intersection)” type. According to the

type of the value, FE determines the variables. In the

above mentioned translated results, FE identifies

“motors” and “are free of defects” as values of “11”

of value type “Key Is” and “negation” respectively.

Thus, the obtained variable is “11 (of the motors)

=free of defects”.

3.4 Translate into Mathematical Form

The final step, “Translate into Mathematical Form”

generates the mathematical form of the exercise.

First, it translates the extracted variables, then the

given part of the authored exercise, and finally the

questions associated with the exercise. Basically,

MAT tries to understand the given as if trying to

solve it, taking advantage of the structure of the

variables and related mathematical rules. As

probability exercises vary in type and in the way

they are solved, MAT defines different algorithms

for different categories of probability exercises. For

instance, for probability exercises of type “basic

probability”, the system searches for the “sample

space”, the events and the possible negation or

intersection relations. Following are examples of

obtained variables: (1): 20.0 (electric motors); (2):

11.0 (of the motors) = free of defects; (3): 8.0 = have

defects on the exterior finish; (4): 3.0 = have defects

in their assembly. MAT “translates” (3) and (4) as

events A and B because they are related to the

keyword with the value type “value”; (2) is

identified as a “relation value” since it has the value

negation (Not A and Not B), and (1) is the sample

space. After translating the variables, MAT displays

the given in mathematical form using predefined

probability symbols. For example, it associates each

event with its number value, e.g., Event A= 8.0.

Also, it identifies terms that have mathematical

implications, e.g., the word “at least one” implies a

union P(AUB). Thus, the translated mathematical

form is displayed as: “Consider the Sample Space

20.0, let the events A and B, 8.0 is the number

having A, 3.0 is the number having B, 11.0 is the

number having NOT A & NOT B.” Finally, MAT

translates the questions associated with the exercise

to mathematical form using the same approach.

4 TESTING AND RESULTS

The performance of Feature Extract (FE) is

evaluated based on the percentage of exercises it can

produce a valid template for and the various types of

probability exercises it can cover. Based on the

recommendation of statistics and probability

professors, and after researching commonly used

statistics/probability books (such as (Scheaffer et al.,

2010) and (Grinstead and Snell, 2009)) and other

online resources, we decided to cover the following

topics for the testing: Basic Probability (20

questions), Permutation & Combination (24

questions), Conditional Probability (14 questions)

and Distributed Probability (12 questions). The

reason behind choosing these categories is that they

are the most common and preliminary to several

majors requiring probability courses. As such, a total

of 70 different exercises were selected to cover the

various question formats and topics. The exercises

were first translated to mathematical form using

MAT. Then, with the help of probability instructors,

the translated forms were compared and evaluated

either as correctly translated, partially translated

(containing some minor errors) or incorrectly

translated. Overall, MAT succeeded in extracting the

templates and translating correctly to the

mathematical form 84% of the exercises. Moreover,

MAT was able to partially translate 14% of the

exercises and incorrectly translated only 2% of the

exercises.

4.1 Results and Findings

Figure 2, shows the detailed results of each tested

category. As observed, FE performed well in

translating correctly exercises of the different

categories. As for the exercises that were partially

and incorrectly translated, we observed several

issues in which the majority of them could be fixed.

First of all, in the “Basic Probability” category,

MAT correctly extracted and translated 90% of the

exercises and it performed well in identifying events

and sample spaces, and distinguishing between

different probabilities symbolic forms.

As for the “Conditional Probability” category, in 86%

“Majorly Adapted Translator”: Towards Major Adaptation in ITS

455

Figure 2: Detailed Results.

of the exercise FE was able to detect conditional

probabilities in the form of P(A|B) whereas in 14%

of the exercise, where the given was more complex,

it was only able to translate it partially. In the

“Distributed Probability” FE also performed well, as

it was able to correctly translate 92% of the

exercises. Finally, in the “Permutation and

Combination” category FE was able to correctly

translate 95% of the exercises.

The shortcomings of the FE component were due

mainly to the following reasons. First, FE assumed

that all the variables were in the given. In the cases,

where not all the events were presented in the given,

but some in the questions, MAT incorrectly

translated the exercises. This issue can be solved by

performing further text analysis and extracting the

keywords that were not found in the given. Second,

FE had some difficulties in exercises having

unidentified keywords such as “X members attend

Sport”. This can be solved by further extending the

FE to recognize single letters to be keywords.

Finally, FE failed to identify “list of words” which

infers a number, for example the phrase “the vehicle

can go straight, turn right, or turn left” which infers

3 choices, denoted as n=3. This can be solved by

training FE to identify list of consecutive words

separated by a comma as “values”.

Overall, the performance of FE is considered to

be acceptable, especially that the FE phase can

successfully handle complex terms including

numbers (“two out of five” is recognized as 2/5),

units (such as “10 ohm resistors”), negations (such

as “neither”, “nor” or “free of”) and other

ambiguous terms ( “A+” is recognized as a blood

type).

5 CONCLUSION AND FUTURE

WORKS

This paper contributes first in providing arguments

in order to expand the field of AIED towards

adapting to the major of the learner. The proposed

system, “Majorly Adapted Translator” (MAT) is

designed for that end. Indeed, MAT adapts to the

students’ major by translating exercises from one

concept to another according to their major. The

system consists of two parts, “Feature Extract”

which identifies the parts of an authored exercise

that must be changed (i.e., variables) and

“Translate” which translates these to different

concepts. In this work, we highlight the Feature

Extract phase, which relies on our own relation

extraction method to identify variables which

extracts relations specific to named entities by using

dependency relations and shallow parsing.

The system was tested on 70 different exercises,

which were selected to cover the various question

formats and topics from the Statistics and

Probability domain. MAT was successful at properly

extracting the templates and translating into

mathematical form 84% of the exercises. Moreover,

in 14% of the exercises MAT was partially

successful, and the reasons for these limitations were

determined.

As future works, first we plan to increase the

performance of the FE phase by addressing the

limitations highlighted in the testing section.

Second, the translate phase only translates to

mathematical form. As such, it should be extended

to translate other majors in order to further test the

system with actual learner. In addition, we plan to

investigate extending the system to translate various

topics other than probability such as linear

programming problems and numerical analysis.

REFERENCES

Azi, J. I., Aboderin, A. O., Nkom, A. A., and Schweppe,

M., 2008. Dialogue on Integrating Indigenous

Contents and Instructional Technology Application in

Nigeria: The “i-CLAP” Model Experience. In

EMERGE'08, Emerge Conference, South Africa.

Blanchard, E. G. and Frasson, C., 2005. Making

Intelligent Tutoring Systems culturally aware: The use

of Hofstede’s cultural dimensions. In IC-AI'05,

International Conference on Artificial Intelligence,

Las Vegas, pp. 644-649.

90%

86%

92%

95%

14%

75%

80%

85%

90%

95%

100%

Basic

Probability

Conditional

Probability

Distributed

Probability

Permutation

Combination

Correctly Translated Partially Translated

Incorrectly Translated

ICAART 2019 - 11th International Conference on Agents and Artiﬁcial Intelligence

456

Briggs, S., 2014, How To Make Learning Relevant To

Your Students (And Why It’s Crucial To Their

Success), viewed 23 September 2018,

<https://www.opencolleges.edu.au/informed/features/how-

to-make-learning-relevant/>.

Carnegie Mellon University, 2015. Solve a Problem: The

course draws students across majors, viewed 20

September 2016,

<https://www.cmu.edu/teaching/solveproblem/strat-

backgroundknowledge/backgroundknowledge-

02.html>.

Eboa, F. M., Courtemanche, F., and Aïmeur, E., 2010.

CAMPERE: Cultural Adaptation Methodology for

Pedagogical Resources in E-learning. In CATS'10,

Third International Workshop on Culturally-Aware

Tutoring Systems, Montreal, pp. 37.

Finkel, J. R., Grenager, T., and Manning, C., 2005.

Incorporating Non-local Information into Information

Extraction Systems by Gibbs Sampling, In 43

Annual

Meeting on Association for Computational Linguistics,

Ann Arbor, pp. 363-370.

Gasparini, I., Bouzeghoub, A., De Oliveira, J. P. M., and

Pimenta, M. S., 2010. An adaptive e-learning

environment based on user´s context. In CATS'10,

Third International Workshop on Culturally-Aware

Tutoring Systems, Montreal, pp. 1 -12.

Grinstead, C. and Snell, J., 2009, Introduction to

Probability. American Mathematical Society.

Hofstede, G., 1997 Cultures and Organizations: Software

of the Mind, McGraw-Hill, London.

Khemaja, M. and Taamallah, A., 2016. Towards Situation

Driven Mobile Tutoring System for Learning

Languages and Communication Skills: Application to

Users with Specific Needs. Journal of Educational

Technology and Society, vol. 19, no. 1, pp. 113-128.

Klašnja-Milićević, A., Vesin, B., Ivanovic, M., Budimac,

Z., and Jain, L. C., 2017. E-Learning Systems-

Intelligent Techniques for Personalization,

Springer,Cham.

Kneale, P., 2009. Teaching and Learning for

Employability. In Fry. H., Ketteridge, S., & Marshall,

S. (eds.), A Handbook for Teaching and Learning in

Higher Education, 3

edition, Routledge, NewYork,

pp. 99-112.

Manning, C., Surdeanu, M., Bauer, J., Finkel, J., Bethard,

S., and McClosky, D., 2014. The Stanford CoreNLP

natural language processing toolkit. In 52

Annual

Meeting of the Association for Computational

Linguistics: System Demonstrations, pp. 55-60.

Melis, E., Goguadze, G., Libbrecht, P., and Ullrich, C.,

2009. Culturally adapted mathematics education with

ACTIVEMATH. AI and Soc, vol. 24, no. 3, pp. 251–

265.

Mohammed, P. and Mohan, P., 2014. Dynamic

contextualization of educational context in intelligent

learning environments using ICON. International

Journal of Artificial Intelligence in Education, vol. 25,

no. 2, pp. 249–270.

Nye, B., 2014. Intelligent Tutoring Systems by and for the

Developing World: A Review of Trends and

Approaches for Educational Technology in a Global

Context. International Journal of Artificial

Intelligence in Education, vol. 25, pp. 177–203.

Ogan, A., Walker, E., Baker, R., T. Rodrigo, M., Soriano,

J., and Castro, M., 2014. Towards understanding how

to assess help-seeking behavior across cultures.

International Journal of Artificial Intelligence in

Education, vol. 25, pp. 229–248.

Prudchenko, K., 2017, What Kind of Math Is Expected of

a Civil Engineering Student?, viewed 23 May 2017,

<http://education.seattlepi.com/kind-math-expected-

civil-engineering-student-2159.html>.

Scheaffer, R., Mulekar, M., and McClave, J., 2010.

Probability and Statistics for Engineers. 5

Edition,

Cengage Learning.

The University of Chicago, 2017, Math Needed for

Computer Science, viewed 23 May 2017,

<https://csmasters.uchicago.edu/page/math-needed-

computer-science>.

Universal Dependencies Contributors, 2016, Universal

Dependencies v2, viewed 20 January 2017,

<http://universaldependencies.org/>.

Vartak, M. P., Almeida, S. F., and Heffernan, N. F., 2008.

Extending ITS Authoring Tools to be Culturally

Aware. In CATS'08, Culturally Aware Tutoring

Systems, Montreal, pp. 101-105.

“Majorly Adapted Translator”: Towards Major Adaptation in ITS

457