Generating educational assessment items from Linked Open Data The case of DBpedia Muriel Foulonneau [email protected]
Generating educational assessment items from Linked Open Data
The case of DBpedia
Muriel Foulonneau
“To Really Learn, Quit Studying and Take a Test” (NYT, Jan, 2011)
Formative assessment
Self-assessment
Items are expensive
Creating, reusing, sharing test items
05/2011 2ESWC 2011
Why generating items?
� Security issue
Adding variability to an item
no expected variation of the construct
� Model-based learning
Generating items from knowledge represented as a model
the construct is modified for each item
30/05/2011 Presentation Tudor 3
Assumption on model-based learning
INTERESTING BECAUSE
- Can enable adaptive learning paths
- Independent from particular representations of learning resources
CONSTRAINTS
A domain model must exist
- Can enable adaptive learning paths
- Bring experts together to design a model of what learners should learn
LIMITATIONS
- Experts are difficult to mobilize for a long modeling exercise
- What about specialized /professional knowledge?
- How to ensure the evolution of the model?
30/05/2011 Presentation Tudor 4
The LoD Cloud as a source of knowledge
� Existing data sources
no need to gather experts
� Including knowledge which is not well codified in curricula
Knowledge gathered from experts as well as non experts
� Many datasets added or modified all the time
Can reflect evolution of the knowledge
30/05/2011 Presentation Tudor 5
Using LoD for model-based learning
30/05/2011 Presentation Tudor 6
Limitations of model-based learning
LoD as a source of knowledge
� Experts are difficult to mobilize for a long modeling exercise
�Existing data sourcesNo need to gather experts
�What about specialized /professional knowledge?
�Including knowledge which is not well codified in curricula
Knowledge gathered from experts as well as non experts
�How to ensure the evolution of the model?
�Many datasets added or modified all the time
Can reflect evolution of the knowledge
Objectives of the experimentation
� Are there limitations to the use of Linked open Data as a knowledge model for learning ?
• Is this feasible?
• Are the datasets relevant?
• How much quality control is needed?
Test on factual knowledge for simple choice items
30/05/2011 Presentation Tudor 7
Semi-automatic item generation
� Manual definition of an item template
� Automatic generation of variables
30/05/2011 ESWC 2011 8
Stem variables
options
key
Auxiliary information
Existing strategies
• Algorithms• X: Value range: 3 to 18 by 3
• Natural language processing• vocabulary questions and cloze questions
• Structured datasets• Vocabulary questions from the WordNet dataset
• Model extraction then question generation• From natural language (or model creation by experts)
� Mostly used in mathematics and scientific subjects • where algorithmic definition of variables is easier
� And for L2 learning
Challenge to generate other types of variables
• Additional information, historical knowledge, feedback…
30/05/2011 Presentation Tudor 9
The QTI item generation process
30/05/2011 Presentation Tudor 10
QTI Item template
IMS Question & Test Interoperability Specification
XML serialization using JSON templates
30/05/2011 ESWC 2011 11
<choiceInteraction responseIdentifier="RESPONSE" shuffle="false"maxChoices="1">
<prompt>What is the capital of {prompt}?</prompt><simpleChoice
identifier="{responseCode1}">{responseOption1}</simpleChoice><simpleChoice
identifier="{responseCode2}">{responseOption2}</simpleChoice><simpleChoice
identifier="{responseCode3}">{responseOption3}</simpleChoice></choiceInteraction>
Get the knowledge from LoD
SELECT ?country ?capitalWHERE {?c <http://dbpedia.org/property/commonName> ?country .?c <http://dbpedia.org/property/capital> ?capital}LIMIT 30
30/05/2011 ESWC 2011 12
SPARQL query to generate capitals in Europe
Never possible to generate an item from a single triple because of constraint to find appropriate labels
Label
Generating item distractors
i.e., incorrect answer options
Strategies
- Instances of the same class
⇒ Creation of a variable store⇒ Random selection of distractors
Next step: Attribute-based resource similarity (can be instances of a different class)
=> use of semantic recommender system
30/05/2011 ESWC 2011 13
Item data dictionary
30/05/2011 ESWC 2011 14
Generation of the QTI-XML item
30/05/2011 ESWC 2011 15
Publication on the TAO platform
� TAO is an open source e-assessment platform based on semantic technologies.
Used for diagnostic, formative, large-scale assessment, including national school monitoring, OECD PISA/PIIAC surveys, competence assessment for unemployed ….
� Supports imports
of IMS-QTI items
30/05/2011 Presentation Tudor 16
Different types of questions
Q1: queries uncontrolled datasets
Q2: queries revised ontology
Q3: ����queries historical information
Q4: queries a linked data set to add item feedback
Q5: queries medical information
30/05/2011 ESWC 2011 17
Q1: What is the capital of { Azerbaijan }?
� Infobox dataset
� 3 were not generated for a country (Neuenburg am Rhein, Wain, and Offenburg)
� “Managua right|20px”
� Two distinct capitals were found for Swaziland (Mbabane, the administrative capital and Lobamba, the royal and legislative capital)
30/05/2011 ESWC 2011 18
Q2: Which country is represented by this flag ?
� Use of FOAF and YAGO
� Transactional closures
<http://dbpedia.org/class/yago/EuropeanCountries><http://dbpedia.org/class/yago/Country108544813>
� 6/30 URIs did not resolve to a usable picture (HTTP 404 errors or encoding problem).
30/05/2011 ESWC 2011 19
Q3:Who succeeded to { Charles VII the Victorious } as ruler of France ?
� YAGO ontology
� 1 was incorrect (The three Musketeers)
� Multiple labels for the same king
Louis IX, Saint Louis, Saint Louis IX
� One item generated with options having inconsistent naming:
Charles VII the Victorious, Charles 09 Of France, Louis VII
30/05/2011 ESWC 2011 20
Q4:What is the capital of { Argentina }? With feedback
� Uses the linkage of the DBpedia dataset with the Flickr wrapper dataset
� The Flickr wrapper data source was unavailable
� No IPR information
30/05/2011 ESWC 2011 21
Q5: Which category does { Asthma } belong to?
� Retrieves diseases and their categories
� SKOS and Dublin Core, Infobox dataset for labels
� SKOS concepts are not related to a specific SKOS scheme
� Categories of diseases from Skeletal disorders to childhood. => the correct answer to the question on Obesity is childhood.
30/05/2011 ESWC 2011 22
Data quality challenges
From Q1, 53,33% were directly usable
neither a defective prompt nor a defective correct answer nor a defective distractor .
Benchmark from unstructured content between 3,5% and 21%.
Issues• Ontology issue
• Labels
• Inaccurate statements
• Data linkage (resolvable URIs)
• Missing inferences
30/05/2011 ESWC 2011 23
Chance that an item
will have a defective distractor =
Data selection
Item difficulty
- can change even with variables not related to the construct (cognitive issues)
- Can change according to the distractors
- => need to establish a framework to assess the difficulty of the construct AND of the item in general (including the relevance of the distractors for instance)
- Psychometric model: what do we know about previous test takers? What can we infer from their performance?
- Ad hoc model: can an a priori difficulty assessment be performed or ����inferred?
30/05/2011 ESWC 2011 24
Future work
� Assessing models on Linked Open Data as a source of knowledge for supporting formative assessment and the learning process
� Improving the selection of distractors by integrating dedicated similarity approach (from a semantic recommender system)
� A wider variety of assessment item models
� A framework to assess the difficulty of items
� Authoring interface for item templates
30/05/2011 ESWC 2011 25