2. • Introduction
• Nichesourcing Web Document Quality Assessments
• User studies
• Conclusion and Future Work
Outlin
e
Capturing the Ineffable: Collecting, Analysing, and Automating Web Document Quality
Assessments
4. Web Document Quality Assessment
• Source criticism
• Methodological practice from the humanities
• e.g., from the American Library Association:
• How was the source located?
• What type of source is it?
• Who is the author and what are the qualifications of the
author in regard to the topic that is discussed?
• When was the information published?
• In which country was it published?
• What is the reputation of the publisher?
• Does the source show a particular cultural or political bias?.
• How does it apply to Web sources?
Capturing the Ineffable: Collecting, Analysing, and Automating Web Document Quality
Assessments
5. Web Document Quality Assessment
What is the quality of each of these documents?
Capturing the Ineffable: Collecting, Analysing, and Automating Web Document Quality
Assessments
Authoritative source ✓
Accurate ✓
Precise ✓
Complete ✓
Neutral (?)
Blog Post (?)
Accurate (?)
Precise (?)
Complete (?)
Neutral ✗
6. • We adapt source criticism to Web documents &
aim at automating the process of quality
estimation by:
• Gathering quality assessments (mostly from experts).
• Looking for markers (document features) that correlate
with them.
Quality and Quality
Markers
Capturing the Ineffable: Collecting, Analysing, and Automating Web Document Quality
Assessments
7. Objectives
• Analyse the consistency of quality assessments.
• Are quality assessments consistent among users, over
time, etc.?
• Analyse user ability to interpret document
features.
• Can the users estimate the quality of a document from
its sentiment or trustworthiness level?
• Analyse the predictability of quality assessments.
• Can we automatically estimate the quality of a
document?
Capturing the Ineffable: Collecting, Analysing, and Automating Web Document Quality
Assessments
9. • Dataset: documents about vaccinations
• Initially, 50 docs, various sources (blogs, authorities, etc.)
• Features
• Information (automatically) extracted from documents
using AlchemyAPI & Web of Trust.
• Entities, Topics, Sentiment, Emotions, Trustworthiness.
• Quality dimensions
• Overall quality, accuracy, completeness, precision,
trustworthiness, readability, neutrality.
Dataset, Features, and Quality
Dimensions
Capturing the Ineffable: Collecting, Analysing, and Automating Web Document Quality
Assessments
10. • Setup:
• 6 documents per participant.
• Random selection.
• Even distribution of assessments.
• Scenario:
Suppose you are asked to write an article
about debate on vaccinations triggered by the
measles outbreak in 2015 at Disneyland in
California.
WebQ: Nichesourcing Web Quality Assessments
Capturing the Ineffable: Collecting, Analysing, and Automating Web Document Quality
Assessments
11. • Documents are anonymized.
• Users choose documents that meet their quality
criteria based on features only.
• All feature values are shown, alone and together.
WebQ: Task 1
Capturing the Ineffable: Collecting, Analysing, and Automating Web Document Quality
Assessments
12. • Read each of the 6 articles.
• Assess it.
• Rate completeness, accuracy, etc.
• Likert scale 1-5.
• Annotate the article to explain the ratings
• Articles are proxied & annotated through AnnotatorJS.
WebQ: Task 2
Capturing the Ineffable: Collecting, Analysing, and Automating Web Document Quality
Assessments
13. User Studies
Capturing the Ineffable: Collecting, Analysing, and Automating Web Document Quality
Assessments
14. • User Study 1
• Participants: 20 last-year UvA journalism
students.
• Duration: 60’.
• User Study 2
• Participants: 20 RMA media scholars.
• Duration: 45’.
• Improvements (learnt from user study 1).
Setup
Capturing the Ineffable: Collecting, Analysing, and Automating Web Document Quality
Assessments
15. • Data collected:
• 104 (US1) + 47 (US2) assessments.
• 238 (US1) + 89 (US2) annotations.
• No significant difference between Use Cases
(Wilcoxon signed-rank test).
• Assessments are assimilable.
• Assessment predictability (SVC)
• Up to 63% accuracy (5-classes)
• Up to 89% accuracy (2-classes)
• Promising predictability. We will try other algorithms.
Results
Capturing the Ineffable: Collecting, Analysing, and Automating Web Document Quality
Assessments
16. • Highest correlation with overall quality:
• Accuracy
• Trustworthiness
• Precision
• Completeness
• Given the task at hand, neutrality is not relevant.
• Weak correlation task 1 - overall quality (task 2).
• Users were mostly unable to interpret those features.
Results
Capturing the Ineffable: Collecting, Analysing, and Automating Web Document Quality
Assessments
18. • We collected Web document quality
assessments.
• WebQ – Nichesourcing application.
• 2 user studies with experts.
• Clear defined task.
• Controlled dataset.
• We analysed the assessments, and automated
their prediction.
• The task matters more than subjectivity.
• Assessments are quite uniform and coherent.
• Features in isolation are not very meaningful.
• The application setup is important.
Conclusion
Capturing the Ineffable: Collecting, Analysing, and Automating Web Document Quality
Assessments
19. • We plan to and are currently working on:
• Extending the dataset (currently ~1,500 documents).
• Scaling up the experiments and gathering more
assessments.
• Involving laymen via crowdsourcing.
• Extending the analyses.
• Utilising other automated reasoning approaches.
(Current and) Future Work
Capturing the Ineffable: Collecting, Analysing, and Automating Web Document Quality
Assessments