Upload
maya-stafford
View
217
Download
0
Tags:
Embed Size (px)
Citation preview
WP5: Validation of developed functionality
Report on the final Experiments
Anne de Roeck
Diane Evans
James Gray
The Open University - UK
OutlineIntroductionBusiness hypothesesStudent Scenario (QUIZ)Results Desktop ActivitiyResultsTutor ScenarioResultsFinal Conclusions
Overall HypothesisThe augmentation of e-learning systems (in this
case ILIAS) with NLP and semantic web technologies
increases the effectiveness of learning and teaching,
and in particular,
Increases the effectiveness of teachers and learners in locating relevant learning objects in the context of learning related tasks.’
IntroductionChanges were made to the methodology after the
previous tests and review.
Measuring learning
Performing Quantitative tests across the scenario activities
Defined Hypotheses and formulated tests to measureIntroduced Mitigation controlsAlso included value judgements as supporting analysis
Restriction to 2 scenarios (1 student & 1 tutor)Scenarios built on experiences of previous trials
Data captured electronically
Student Hypotheses
Retrieval of Content (Student)1.1 The addition of functionalities based on
NLP increases the retrievability of learning objects in terms of the relevance of the content. [Quant test]
1.2 The addition of functionalities based on NLP increases the effectiveness of learners in locating relevant learning objects in the context of answering Quiz questions. [Quant test]
Support for Learning
2.1 The developed tools and ontology help to support learners to better, more effectively grasp the terminological and conceptual space which defines a certain domain of knowledge. [Quant test]
Multi-lingual study
3.1 The ability to easily retrieve content in more than one language supports learning activities of students who are studying in a Multi-lingual situation e.g. the language of study is not their native language.
[Quant test]
Support for Learning Paths and Personalised Learning
4.1 Learners are able to build individual learning paths by entering key terms of concepts they need to learn. [Qual. test]
4.2 Learners are able to classify and order learning material (in the form of documents) they have placed within their personal desktop by using concepts and keywords linked to each document
[Qual. test]
Cont..
4.3 The facility for a learner to classify and order learning material supports self-guided learning.
[Qual. test]
4.4 The facility for a learner to classify and order learning material supports self-guided learning by enabling the user to create a meaningful / linked path through the content. [not tested]
Basis of Student scenarioThe student scenario focussed on:
Demonstrating improvement in learning (as requested by the Reviewers)
The LT4eL search functions i.e. Semantic, Definition, Keyword and Ontology Browsing.
Multi-lingual search and retrieval associated with these search functions
The Use of the Personal Desktop to support individual learning
The scenario centred around a multiple choice QUIZ
Structure and rationale of the student scenario
• Target - 1st year undergraduate humanities students• Topics - Information exchange: protocols and markup languages
You will shortly be studying a short course covering 'Information Exchange on the Internet'.
This is a subject on which you have little or no previous knowledge.
Your tutor has prepared some research questions for you in the form of a quiz to enable you to gain a basic grounding in key aspects.
Depending on your group you will either use internet facilities such as Google and Wikipaedia etc or you will have access to different ILIAS search facilities to assist your search, including the LT4el searches and the structured concept browser.
Experiment design - students
To assess whether technologies "increase the effectiveness of learning", students needed to do some learning:
Learning experience -
Quiz to encourage students to engage with learning objects
Learning experience -
Quiz to encourage students to engage with learning objects
Assessing the learning
Assessment Activity
Pre-Test
Assessment Activity
Pre-Test
Learning experience(quiz)
Learning experience(quiz)
Assessment Activity
Post-Test
Assessment Activity
Post-Test
The answers provided by the student at each of the 3
stages were marked
Measuring the effect of our technology
We introduced a control group with a different learning experience:
Learning experience(quiz)
ILLIAS / LT4eL
Learning experience(quiz)
ILLIAS / LT4eL
Learning experience(quiz)
Internet / Google
Learning experience(quiz)
Internet / Google
TARGET CONTROL
To compare NLP technologies with plain search in the same repository
PreTestPreTest
Quiz Part 1 - ILIASQuiz Part 1 - ILIAS
Quiz Part 2 - LT4eLQuiz Part 2 - LT4eL
PostTestPostTest
PreTestPreTest
Internet/Google Internet/Google
PostTestPostTest
TARGET CONTROL
Test ATest A
Question Set AILIAS
Question Set AILIAS
Controlling for question differences (eg is the post-test easier than the pre-test?):
Question Set BLT4eL
Question Set BLT4eL
Test BTest B
Test BTest B
Question Set B
ILIAS
Question Set B
ILIAS
Question Set A
LT4eL
Question Set A
LT4eL
Test ATest A
Quiz Part 1
Quiz Part 2
PreTest
Post-Test
Test ATest A
Question Set AWWW
Question Set AWWW
Question Set BWWW
Question Set BWWW
Test BTest B
Test BTest B
Question Set BWWW
Question Set BWWW
Question Set A
WWW
Question Set A
WWW
Test ATest A
Group 2 Group 3 Group 4Group 1
TARGET CONTROL
Pre & Post-TestsFive questions testing understanding of relationships between concepts
Pre-test and Post-test questions were paired, to test the same area
Pre-test and Post-test questions were alternated in different groups to ensure fairness
QuizSeven questions with learning materials in the same topics as the test questions designed to encourage engagement.Each question followed by supplementary questions to ascertain:
whether student already knew the answersource document (including language)which tool (if any) helped find the answer
Quiz is time-limitedQuestion groups alternate between student groups
Results of Student Experiments
Main Student Scenario
Searching
Learning
Multilinguality
Student Scenario:Searching
Hypothesis 1.1: The addition of functionalities using / based on NLP increases the retrievability of learning objects in terms of the relevance of their content.
Hypothesis 1.2: The addition of functionalities using / based on NLP increases the effectiveness of learners in locating relevant learning objects in the context of answering Quiz questions.
Student: Searching
Test that mean scores in Part 2 of the quiz (LT4eL) are greater than mean scores in Part 1 of quiz (plain LMS) for the target group.
Student: Searching
Searching - Test 1
Test 1
Quiz 1 Quiz 20
1
2
3
4
5
3.44
4.07
Improvement in quiz scores
Students scored higher in the quiz when they had access to LT4eL technologies to help them.
Student: Searching Test 1 Results
Was it the technology that helped?
Question differences?
no - rotating banks of questions
Order effect?
might expect students to be learning, thus achieve better results in second half of quiz.
Test 1a: test by looking for improvement in control group
Student: Searching Test 1a
Hypothesis 1.1 -
Gains achieved in both groups
Control group scored better in both quizzes
Target Control0
1
2
3
4
5
3.44
4.074.07
4.57
Improvement in quiz scores Quiz 1
Quiz 2
Sco
re
Student: Searching Test 1a Results
Searching - Conclusions so farStudents scored higher in the quiz when they had access to LT4eL technologies to help them.
Partly explained by order effect - students do better in second half of quiz anyway
Students using resources of World Wide Web score higher than those restricted to learning objects within an SMS
LMS score improvement appears slightly better than WWW score improvement
Student: Searching Conclusions 1
Searching - test 2Test that for any given question in Part B of the quiz, students who choose to use LT4eL functionality score more highly for that question than those who do not use LT4eL functionality.
Note: In Part 2 of the quiz, students are free to use whichever search method they feel will be most effective. This test is to assess whether those who use LT4eL methods score more highly than those who do not.
Student: Searching Test 2
Only three questions answered better using LT4eL functionality
Q1 Q2 Q3 Q4 Q5 Q6 Q70%
20%
40%
60%
80%
100%
Score per question by functionality
Full text
LT4eL
Question
Sco
re
Student: Searching Test 2 Results
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Full text search
Keyword search
Semantic search
Concept browser
Definition finder
Success in answering quiz questions by functionalityF
un
cti
on
ali
ty
Success
Student: Searching Test 2a Results
(target groups only)
Student: Searching Opinions 2
What did they dislike about Semantic Search?It didn't return relevant results.
because it doesn't find what I am searching for
its too vague
i didn't use it a lot as the results were chaotic
i find it not much to the point for the types of research i usually do
It is a bit too much to offer this much search methods
the name semantic is confusing
i liked this type best.. it was the easier to find the relevant information
Student: Searching Comments: Semantic
What did they dislike about the Concept Browser?
It didn't return relevant results
It was too slow for my part and did not give any additional value
it's a roundabout way of searching
was not eay to use it. maybe this was because i did not fully understand how the concept browser worked
I am not sure what the concept browser is
don't know what it is.
for content questions this might be a relevant search method. However, less relevant when studying a language
I like that method the most but it wouldn't be useful in my studies - English philology
it helps a lot to understand given topic/term
Student: Searching Comments: Concept Br
0% 10% 20% 30% 40% 50%
full text search
keyword search
semantic search
concept browser
definition finder
Most useful functionality - students' opinionsF
un
ctio
nal
ity
Percentage finding this the most useful
Student: Searching Opinions 1
Searching - Test 3 Test that for any given question in Part B of the quiz, students who choose to use LT4eL functionality and answer correctly, do so more quickly than those who answer the question correctly using Full Text Search
Notes:
In Part 2 of the quiz, students are free to use whichever search method they feel will be most effective. This test is to assess whether those who use LT4eL methods are faster than those who do not.
Only consider students who searched for an answer, and answered correctly
Student: Searching Test 3
Students answered faster using LT4eL technologies than using full text search.
Student: Searching Test 3 result
Q1 Q2 Q3 Q4 Q5 Q6 Q70
30
60
90
120
150
180
210Average time to answer questions
Full text
LT4eL
Question
Sec
on
ds
Test 3 resultStudent: Searching
Q1 Q2 Q3 Q4 Q5 Q6 Q70
30
60
90
120
150
180
210Average time to answer questions
Internet
Full text
LT4eL
Question
Sec
on
ds
Searching - conclusions
Using Semantic Search, Concept Browser and Definition Finder to find answers, students score better than using plain text search
Students appreciate all functionality, especially the Definition Finder;
Students' ratings of Semantic Search and Concept Browser, though favourable, showed that these techniques were less familiar to them.
Students answered faster using LT4eL technologies than using full text search; speed of searching seems to improve with practice.
ConclusionsStudent: Searching
Hypothesis 2.1: The developed tools and ontology help / support learners to better / more effectively grasp the terminological and conceptual space which shape a certain domain of knowledge.
Student: Learning Test 1
Test: Assess learning by testing that the mean score in the post-test is higher than the mean score in the pre-test
Assessment Activity
Pre-Test
Learning experience(quiz)
Assessment Activity
Post-Test
Learning has taken place. This validates an important part of the methodology.
Pretest Post test0
1
2
3
4
5
2.53
3.58
Improvement in test scores
Student: Learning Test 1 - result
Learning - test 2Test that mean improvement in test scores for target sample (using ILIAS and LT4eL) is higher than mean improvement in test scores for control group (using Internet to find content).
Note: This test is intended to show whether learning is more effective in the target sample than in the control sample.
Student: Learning Test 2
Learning appears slightly greater in Control group; but they started from slightly lower base. But not statistically significant
Target Group Control Group0
1
2
3
4
5Improvement in test scores
Pretest
Posttest
Sco
re
Student: Learning Test 2 - result
Learning - Test 3Test whether students who use the LT4eL functionality in the LMS learn more than those who do not.
Student: Learning Test 3
Did not use LT4eL functional-ity
Used LT4eL func-tionality some-times (1-4 times)
Used LT4eL func-tionality frequent-ly (5-7 times)
0%
20%
40%
60%
80%
100%
Test improvement by Quiz functionality
No of students dropping marks
No of students keeping same marks
No of students gaining marks
Student: Learning Test 3 - result
Learning - Conclusions
Learning took place in target and control group
Students in the target group who used LT4eL functionality extensively, learned more than those who stuck to plain text search.
Student: Learning Conclusions
MultilingualityHypothesis: The ability to easily retrieve content in more than one language supports learning activities of students who are studying in a Multi-lingual situation.
Student: Multilinguality
Notes:• We considered test persons who are capable to read and
understand documents in other languages than their native. • In normal situation we expect also students who are studying in
other language that their native (exchange students) to make use of the multilingual facilities and find documents in their language
• For our tests this was however not replicable as we would not have been able to ensure a relevant sample size
For the quiz:
Allow students to select most appropriate documents to answer questions, irrespective of language
Students will retrieve greater linguistic diversity of documents
Students who find documents in other languages will score highly
FindingDocuments
Getting correct
answers
Student: Multilinguality Tests
Multilinguality - Test 1Searching
Test: Allow students in a quiz to select the most appropriate document to answer the questions, irrespective of language.
We expect those using NLP-based functionalities to select a higher proportion of documents in a non-course language
Test 1Student: Multilinguality
Target Quiz 1
Target Quiz 2
Control Quiz 1
Control Quiz 2
0%
20%
40%
60%
80%
100%
Language of document retrieval
Course language
Non-course language
Student: Multilinguality Test 1 result
Conclusions so farTarget group retrieved more non-course language documents when they had access to LT4eL functionality, showing that multilingual access supported the learning activity
No significant increase in non-course language retrieval for control group, suggesting this is not an order effect
Student: Multilinguality
Test: test that, of students who are studying in a non-native language, those who retrieve second-language documents in Quiz Part 2 (either target or control) score more highly for those questions than those who do not.
Note: This is intended to show that multilingual retrieval is helpful in the context of answering quiz questions
Student: Multilinguality Test 2
Multilinguality - Test 2Getting the right answer
Document in main course language
Document in other language
0%
20%
40%
60%
80%
100%
74% 71%
Percentage of correct answers
Student: Multilinguality Test 2 result
Those retrieving documents in a non-course language score statistically the same for those questions as those who retrieved documents in their main course language.
These results depend on the quantity and quality of information in different languages available within the repository being searched (or across the Internet for control group students).
It is possible that students retrieving in a second-language document are less fluent in that language, hence have more difficulty assimilating the information within the document. In 328 cases (48%) the language of retrieval was English. It is also possible that students only searched in a second language when they had failed to find the answer in their course language.
Student: Multilinguality
Multilinguality - test 3Does our functionality aid multilingual retrieval?
Student: Multilinguality Test 3
Full text Keyword Semantic Concept browser
Definition finder
0
20
40
60
80
100
Language of documents by functionalityDocuments in other language
Documents in course language
Search method
Nu
mb
er o
f d
ocu
men
ts
Student: Multilinguality Test 3 result
Student ScenarioQualitative evaluation of multilingual Search
Multilingual search would be useful for my studies?
71% (157 Students) STRONGLY AGREED
12% (25 Students) DISAGREED
Lower results for students in countries where large collections of materials in their native language are available (EN, DE) or in countries were usually materials (at least for the particular domain of the test) are in English (NL)
Student: Multilinguality Opinions
Multilinguality conclusions
Students retrieved more non-course language documents when they had access to LT4eL functionality, but this did not result in higher quiz scores.
LT4eL functionality helped retrieval of multilingual content - particularly Concept Browser, Definition Finder and Semantic Search
Student: Multilinguality Conclusions
Overall conclusions - StudentsUsing Semantic Search, Concept Browser and Definition Finder to find answers, students score better than using plain text search
Students most appreciate Definition Finder; they found it harder to see the benefit to them of Semantic Search and Concept Browser
Students answered faster using LT4eL technologies than using full text search
Learning took place in target and control group
Students in the target group who used LT4eL functionality extensively, learned more than those who stuck to plain text search.
Students retrieved more non-course language documents when they had access to LT4eL functionality, but this did not result in higher quiz scores.
LT4eL functionality helped retrieval of multilingual content - particularly Concept Browser, Definition Finder and Semantic Search
Students Overall conclusions
Personal Desktop experiments
Create classifications and organise their content within these sections
Use LT4eL functionality to locate other resources in own and course languages.
Organise these into structure
Student tasks required them to:
Evaluate the usefulness of these features for their study
Students were able to classify documents into topic sections within the personal desktop.
Using the concepts and keywords linked to each document helped them do the classification
Students were able to order documents within the Sections
RESULTS
Student ScenarioMultilingual support with the Personal Desktop
Finding, classifying and ordering of multilingual documents within the personal desktop is possible using the NLP based functionality
?
62% (10% strongly) AGREED that they were able to classify and order documents in more than one language based on key concepts
75% (24% strongly) AGREED that they found documents in more than one language using concept and topics
Student: Multilinguality Opinions
Students agreed that:Features would help students organise their study
Multiple course modulesIndividual directed study
RevisionResearchHaving classified and ordered documents,
they have a clearer understanding of the relationships between the topics
Tutor Hypotheses
Support for Teaching in a multi-lingual context - finding resources
The addition of functionalities based on NLP increases the effectiveness of tutors in identifying supplementary resources to support students who are studying in a multi-lingual context. [Quant test]
The addition of functionalities based on NLP increases the effectiveness of tutors in identifying documents that are relevant, reliable, and appropriate for level..
[Quant test]
Content Preparation
Using the Keyword Extractor based on NLP increases the quality and consistency of keywords assigned to repository content and helps tutors who are expected to perform this task. [Quant test]
Using the Keyword Extractor based on NLP enables tutors to produce a set of keywords faster than those tutors who perform the task unsupported
[Quant test]
Basis of Tutor Scenario
The tutor scenario focussed on:Multi-lingual search the LT4eL functions of Semantic search and ontology browsingThe Keyword Generator for new content
Both in English and in native language
Tutor scenario design
The tutor scenario was based on the need to
find content to support students’ learning
and to add additional content to the repository
Tutor tasks
Internet Resources Internet Resources
LMS resourcesLMS resources
Locate Content
Group1 - ILIASGroup1 - ILIAS Group2 - LT4eLGroup2 - LT4eL
Search Topics
Students will learn which markup languages are appropriate for holding data.
Data formats on the Web
Students will be able to identify ways to construct links to internal and external resources.
links in markup languages
Students will be able to identify which protocols are required to transfer files over the internet
file transfer and internet protocols
DescriptionTopic
Session 1 - Set A topics
Students will learn some of the different picture formats & how to select an appropriate one for different purposes.
picture formats on the Web
Students will be able to identify how to present data in tabular form on web pages
tables in markup language
Students will be able to identify which protocols are used for sending & receiving emails
email and internet protocols
DescriptionTopic
Session 2 -Set B topics
Add New ContentAdd New Content
Group1 - ILIASGroup1 - ILIAS Group2 – LT4eLGroup2 – LT4eL
Tutor Scenario - Results
Searching
Multilinguality
Keyword addition
Hypothesis: The addition of functionalities based on NLP increases the effectiveness of tutors in identifying documents that are relevant, reliable, and appropriate for level.
Test within the same repository
Use tutors' own assessments of relevance, reliability and appropriateness for level.
Tutor: Searching Test
Rele-vance Native
Rele-vance English
Reliabili-ty Native
Reliabili-ty Eng-lish
Level Na-tive
Level English
0
1
2
3
4
5
Quality of documents
ILIAS score
LT4eL score
Attribute
Sco
re
Tutor: Searching Results
Searching - conclusion
LT4eL functionality makes no difference to the tutors' assessments of the Relevance, Reliability or Level of the documents found when searching within the repository.
Tutor: Searching Conclusion
Hypothesis: The addition of functionalities based on NLP increases the effectiveness of tutors in identifying supplementary resources to support students who are studying in a multi-lingual context.
i.e. tutors can find content in two languages
Tutor: Multilingual
In more detail...
Tutors are better able to find second language documents using LT4eL than using Internet search / plain ILIAS search in the same repository.
Tutors find location of second language documents using LT4eL is relatively faster than using Internet searches / standard ILIAS.
Tutors find location of second language documents using LT4eL is relatively easier than using Internet searches / standard ILIAS in the same repository.
Tutor: Multilingual Tests
Tutors are better able to find second language documents using LT4eL
Internet searches
ILIAS searches
LT4eL searches
0%
20%
40%
60%
80%
100%
Success of multilingual searching
Two Documents Not Found
Two Documents Found
Tutor: Multilingual Test 1
Tutors find location of second language documents using LT4eL is relatively faster than using Internet searches / standard ILIAS
Tutor: Multilingual Test 2
Internet Ilias LT4eL0%
20%
40%
60%
Fastest document to find
First Language
Both Same
English Language
Tutors find location of second language documents using LT4eL is relatively easier than using Internet searches / standard ILIAS in the same repository.
Tutor: Multilingual Test 3
Internet ILIAS LT4eL0%
20%
40%
60%
Easiest document to find
1st language docs easier
Both docs the same
English docs easier
Multilingual searches are more successful on the Internet than within repository.
Within the repository, LT4eL functionality does not improve ability to find multilingual content
Within the repository, when two documents are found, tutors using are more likely to find both documents together (same speed, same ease).
Tutor: Multilingual Conclusions
KeywordsUsing the Keyword Extractor based on NLP increases the quality and consistency of keywords assigned to repository content and helps tutors who are expected to perform this task
Test: Tutors using the keyword generator produce a more consistent set of keywords than those who do not
Tutor: Keywords Test 1
Number of distinct keywords Number of words unique to one tutor
0
20
40
60
80
100
Consistency of keywords
Ilias
LT4eL
Tutor: Keywords Test 1 results
Keywords - test 2Using the Keyword Extractor based on NLP enables tutors to produce a set of keywords faster than those tutors who perform the task unsupported.
Tutors assessed their own time to complete the exercise:
Tutor: Keywords Test 2
ILIAS LT4eL 0
1
2
3
2.04
1.41
Time taken to assign keywordsA
vera
ge
co
de
d t
ime
Tutor: Keywords Test 2 results
Keywords - test 3Using keyword extractor gives:
more consistent set of keywords
faster results
But are tutors using the extractor just assigning fewer keywords?
Tutor: Keywords Test 3
ILIAS LT4eL0
1
2
3
4
5
6
7
8
6.94 6.97
Average words per tutor per document
Tutor: Keywords Test 3 results
English Own language0%
20%
40%
60%
80%
100%
Did you think the keyword generator helped you to assign keywords to the document?
Helped a lot
Helped a little
No help
Tutor: Keywords Opinions
base = 23 base = 27
Keywords - Conclusions
Tutor: Keywords
Using the keyword extractor yields a much more consistent set of keywords than manual
Using the keyword extractor, tutors assign keywords 30% faster than not using it.
About 90% of tutors said the keyword generator helped them
Conclusions
Overall conclusions - Tutors
Within the repository, LT4eL appears to make finding multilingual documents together slightly more likely (same speed, same ease).
LT4eL functionality makes no difference to the Relevance, Reliability or Level of the documents found
Using the keyword extractor yields a much more consistent set of keywords than manual
Using the keyword extractor, tutors assign keywords 30% faster than not using it.
Tutor Overall Conclusions
Evaluation - Conclusions
WP5 Participation in Experiments
Participants in Student Scenario
22637387675Total per group
36661311Romania
244389Portugal
31551011Poland
275499Netherlands
41771413Malta
244488Germany
244488Czechoslovakia
19 (23) 2566 (*10)Bulgaria
TotalGroup 4Group3Group 2Group1
8 languages and 226 students
Participants in Student Personal Desktop Activity
12092016182713116
TotalRomaniaPortugalPolandNetherlandsMaltaGermanCzechBulgaria
8 Languages and 120 students
Participants in Tutor Scenario
71Total
11Netherlands
12Romania11German
10Portugal10Czech
11Poland6Bulgaria
7 Languages and 71 tutors
Not a trivial task!8 partners
Pre-test, post-test, Quiz-A, Quiz-B, Desktop questionnaire, Tutor questionnaires - all translated into host languages (and verbatim answers translated back to English)
152 data files to download, clean, and load into database
Analysis database contains: 18 tables, 585 different fields, 2483 records. The total number of non-blank data items is 66,725.
Over 6,500 lines of code to manipulate, condition, load and analyse that data.
Overall HypothesisThe augmentation of e-learning systems (in this case
ILIAS) with NLP and semantic web technologiesincreases the effectiveness of learning and teaching,
and in particular, Increases the effectiveness of teachers and learners in locating relevant learning objects in the context of learning related tasks.’