How do we analyse written and spoken data and turn results into useful information?
- training, feedback, personalisation, tracking, better decision-making
- writing and feedback
- segue into Latent Semantic Analysis
- what can be scored? not just K-12 ELA essays
- application to formative and large-scale learning analytics
will talk about general methods and approaches
more conceptual than mathematical or procedural
turning open-ended language into performance data
- hand scoring is not feasible for large-scale learning analytics - writing, speaking, chat
- multiple choice and logged interaction data isnt always enough to get at skills and knowledge
- analysis of open-ended spoken and written expression is a key to access higher level cognitive processes
- techniques to allow data analytics access a wider range of assessment items
- improve student learning
- improve knowledge/ decision-making of instructors, administrators
- complexity of language can be distilled by mathematical methods (see Caroline's session yesterday for details)
- permits automatic assessment and feedback o written and spoken communication
- learn to score from human scored student responses
- measure the content and quality of responses by determining: the features that human scorers evaluate when scoring a response; how those features are weighed and combined to produce these scores
- obtain 200 - 1000s of pre-scored essays
- extract features from essays
- use machine learning to combine features to predict scores
Aspects of writing
content-based scoring - Latent Semantic Analysis to capture the 'meaning' of written English
Bell Labs - Yellow Pages problem - eg search for cars - get carpools etc - but not where to buy cars - have to search for automobiles. The synonymy problem.
What does that have to do with automated scoring?
- LSA reads lots of text
- learns what words mean and how they relate to each other
- result is a 'semantic space'
- every word is represented as a vector
- every paragraph represented as a vector
content scoring - every essay represented as a vector
new essays are placed based on the words they contain
demo interface
tools available
A score is not enough
- detect off-topic or highly unusual essays
- detect if it may not score an essay well
- detect larding of big words, non-standard words, swear words, too long, too short
- has been tested on millions of essays
- generally agrees with a single human reader as often as 2 human readers agree with each other
- situational judgment tasks and memo writing
- scoring physician patient notes
- language testing and translation
- comparison against - human scores, external validity criteria
- improvements in process/ visualisations - teacher scoring speed, ability to find students
Scenario-based writing - asking students to answer several questions about a hypothetical, yet realistic, scenario.
CLA - automated scoring.
Situated writing assessment
have worked with military on critical thinking, problem solving, leadership.
Automated assessment of diagnostic skills
actor trained - drs scored on process
Example
what helps in formative writing?(Writing Next, 2007; Carnegie Corporation, 2011)
- teaching students strategies for planning, revising and editing their compositions (effect size .82)
- teaching students how to assess their own writing (.46)
- explicitly and systematically teaching students how to summarise texts (.82)
- providing feedback
- monitor students' writing
WritetoLearn example
Scorecard
break
showed teacher dashboard - which students are having trouble with the essays?
impressive graphs showing improvement in assessment scores by the number of revisions. biggest gains in 'ideas', smallest in conventions (eg grammar) why? give more feedback on each iteration.
Automated scoring of spoken performance
- assessing understanding and spoken language proficiency in everyday and workplace contexts
- english fluency
- foreign language skills
- expression, use and comprehension of job relevant content
watch a video and describe what is happening
repeat a sentence
listen to a story and do a re-tell (fluency + comprehension skills)
others
describe a picture or graph - looking at vocabulary, language use, pronunciation and fluency
read a passage aloud
how it works
multiple pieces of information - wave form, spectrum, words segmentation
collect a lot of sound samples, human rated, train system to predict human ratings, run validation
What next?
- combined text _ context - need to be able to combine the textual situation with the contextual learning situation
- machine learning is moving fasst
- NLP is moving slower
- getting a number is not enough
No comments:
Post a Comment