LightSIDE Carolyn Penstein Ros Language Technologies Institute Human-Computer Interaction Institute School of Computer Science
With funding from the National Science Foundation and the Office of Naval Research 1 lightsidelabs.com/research/
2 3 Click here to load a file
4 Select Heteroglossia as the predicted category 5
Make sure the text field is selected to extract text features from 6 Feature Space Customizations Punctuation can be a stand in for mood
you think the answer is 9? you think the answer is 9. Bigrams capture simple lexical patterns common denominator versus common multiple
Trigrams (just like bigrams, but with 3 words next to each other) Carnegie Mellon University POS bigrams capture syntactic or stylistic information the answer which is vs which is the answer
Line length can be a proxy for explanation depth Feature Space Customizations Contains non-stop word can be a predictor of whether a conversational contribution is contentful
ok sure versus the common denominator Remove stop words removes some distracting features Stemming allows some generalization Multiple, multiply, multiplication
Removing rare features is a cheap form of feature selection Features that only occur once or twice in the corpus wont generalize, so they are a waste of time to include in the vector space Feature Space Customizations
Think like a computer! Machine learning algorithms look for features that are good predictors, not features that are necessarily meaningful Look for approximations If you want to find questions, you dont need to do a complete
syntactic analysis Look for question marks Look for wh-terms that occur immediately before an auxilliary verb Click to extract text features
10 Select Logistic Regression as the Learner 11
Evaluate result by cross validation over sessions 12 Run the experiment
13 14 Stretchy Patterns
(Gianfortoni, Adamson, & Ros, 2011) A sequence of 1 to 6 categories May include GAPs Can cover any symbol
GAP+ may cover any number of symbols Must not begin or end with a GAP 16
17 Now its your turn! Well explore some advanced features and error analysis
after the break! 18 Error Analysis Process Identify large error cells
Make comparisons Positive Negative Ask yourself how it is similar to the instances that were
correctly classified with the same class (vertical comparison) How it is different from those it was incorrectly not classified as (horizontal comparison)
Error Analysis on Development Set 20
Error Analysis on Development Set 21 Error Analysis on Development Set
22 Error Analysis on Development Set 23
Error Analysis on Development Set 24 25
26 Positive: is interesting, an interesting scene Negative: would have been more interesting, potentially interesting,
etc. Whats different? 27
28 29 30
31 32 * Note that in this case we get no benefit if we use feature selection over the original feature space.
33 Feature Splitting (Daum III, 2007) General
General Domain A Domain B
Why is this nonlinear? It represents the interaction between each feature and the Domain variable Now that the feature space represents the nonlinearity, the algorithm to train the weights 34 can be linear.
Healthcare Bill Dataset 35 Healthcare Bill Dataset
36 Healthcare Bill Dataset 37
Healthcare Bill Dataset 38 Healthcare Bill Dataset
39 Healthcare Bill Dataset 40
Healthcare Bill Dataset 41 Healthcare Bill Dataset
42 Healthcare Bill Dataset 43