In these slides we present a model that was intended to discriminate creative from non-creative news articles. In order to build the classifier, we have combined nine different measures using a stepwise logistic regression model. The obtained model was tested in two experiments: the first one tried to discriminate between news articles about the US 2012 Elections from different newspapers versus articles taken from The Onion (a website providing satiric news) on the same subject, while the second one evaluated the capacity of the model to generalize over different topics and text genres. The experiments showed that the system achieves 80% accuracy, but the lack of true positives from the second experiment raised the question of whether we really identified creativity or in fact we detected satire (as the assumption for the training corpus was that the satiric news from The Onion were also creative).
3. Introduction (I)
• Goal: Automatically identify creativity in a
text.
• How?
– Define the elements that characterize a creative
text.
– Determine the most important features that
explain creativity.
– Build a model for automatic creativity detection (a
classifier).
Creativity Detection in Texts ICIW201326.06.2013
4. Introduction (II)
• Definitions of Creativity:
– The ability to transcend traditional ideas, rules,
patterns, relationships, or the like, and to create
meaningful new ideas, forms, methods,
interpretations, etc. (Zhu, Xu, Khot, 2009).
– Creativity is typically thought of acting or the quality
of an unpredictable departure from the rules of
regular word formation (Renouf, 2007).
• Linguistic creativity = creativity in texts and
measures “new and creative ways of expressing a
given idea” (Veale, 2011).
Creativity Detection in Texts ICIW201326.06.2013
5. Other approaches
• Manual Identification: 21 creative writers, 4 human judges 105
rated tuples like: (word, sentence, creativity) (Zhu, Xu, Khot, 2009).
• Machine learning algorithm using a linear regression model with 17
features (Zhu, Xu, Khot, 2009).
• Jordanous (2012), - SPECS: three steps procedure for determining
whether a computational system can be defined as creative or not.
• Understanding and using metaphors (Kovecses, 2011; Veale, 2006)
and analogies (Veale, 2006) or on explaining the appearance of new
words from already existing ones (Lehrer, 2007).
• Creativity detection in song lyrics (Hu and Yu, 2011) – uses three
measures for identifying mood and creativity in a lyric.
Creativity Detection in Texts ICIW201326.06.2013
6. Creativity Measures
• A computational creativity measure should address
two aspects:
– Novelty: To what extent an item is different to the existing
samples of its genre?
– Quality: How good the item really is?
• We tried to capture these two criteria through nine
different measures: Type-to-Token Ratio, Word Norms
Fraction, Google Similarity Distance, Explicit Semantic
Analysis , Number of Named Entities, Named Entities
Score, Wordnet Similarity, Coherence measure, and
Latent Semantic Analysis (LSA) measures.
Creativity Detection in Texts ICIW201326.06.2013
7. Experiment Methodology (I)
• Extract the news articles
from the Web and save
them into the database.
• Apply NLP Preprocessing
techniques.
• Apply NLP Categorization
and Tagging techniques.
Creativity Detection in Texts ICIW201326.06.2013
Corpus &
Statistics
Web
Articles’
URLs
Text Preprocessing
Normalize Text
Segmentation
Tokenization
Stemming
Tokens &
Stems
Text
Extraction
URLs HTML
Text & Statistics
Categorization and
Tagging
Part-of-Speech Tagging
Named Entities Recognition
Chunking
Plain Text
Tokens &
Sentences
Named Entities
Corpus Acquisition
Text
Understanding
8. Experiment Methodology (II)
• Compute the value of
each measure for the
given text.
• Use Stepwise Logistic
Regression to select the
measures that best
describe creativity and
build the Classification
Model.
Creativity Detection in Texts ICIW201326.06.2013
Results
Wikipedia
Results in
CSV & ARFF
Formats
Google
Search
Wordnet
Compute Measures
Type-to-Token Ratio
Word Norms Fraction
Google Similarity Distance
Explicit Semantic Analysis
Number of Named Entities
Named Entities Score
Wordnet-Similarity
Coherence measure
LSA measures
Categorized
Text
Categorization
and Tagging
9. The Corpus
• 185 articles on the US Election news taken from:
– 67 articles from The Onion, and
– 118 articles from 12 news sites from all over the
world: UK (BBC, Wired, The Independent, The Sun),
Canada (CBC), Australia (News.com.au, The Australian,
Sydney Morning Herald), USA (Foxnews and
Huffington Post), South Africa (News24), and New
Zealand (The NZ Herald).
• We made the assumption that The Onion articles
are creative.
Creativity Detection in Texts ICIW201326.06.2013
10. Experiments (I)
• Two Experiments:
A. In order to assess the quality of our classifier was
tested on the news articles that we have extracted,
and
B. Intended to measure the capacity of the classifier to
adapt to different kinds texts.
• A – Classifier Evaluation
– Identify the mix parameters of the 9 measures for the
logistic regression model.
– Apply feature selection.
Creativity Detection in Texts ICIW201326.06.2013
11. Experiments (II)
Attribute Beta P value
Constant 1.830477 0.246
WordNet
Similarity -9.779 0
Named
Entities
Score 2.484793 0.0059
LSA cos.
Similarity
sentences 3.445448 0.0001
Number of
Named
Entities -2.89686 0.0053
Word
Norms
Fraction 3.585301 0.0378
Google Path
Similarity -3.25499 0.0538
Creativity Detection in Texts ICIW201326.06.2013
12. Experiments (III)
• Therefore, the classifier for discriminating
creative from non-creative text is given by:
– Pr (Y = 1 | X1, ... , X9) = F(1.83 + 3.585* X2 + 3.255 * X3 -
2.897 * X5 + 2.485 * X6 - 9.779 * X7 + 3.445 * X9)
• Where and X2, X3, X5, X6, X7, X9 are
the scores obtained by each text for Word Norms
Fraction, Google Similarity Distance, Number of
Named Entities, Named Entities Score, Wordnet
Similarity, LSA.
Creativity Detection in Texts ICIW201326.06.2013
13. Results (I)
• The obtained model was tested in a 10-fold
cross-validation:
• The accuracy for this experiment was 80.54%,
which is quite high, considering the difficulty
of this task.
Creativity Detection in Texts ICIW201326.06.2013
Values prediction Confusion matrix
Real
Predicted
Creative Non-creative Sum Precision Recall
Creative 46 15 61 0.754 0.6866
Non-creative 21 103 124 0.8306 0.8729
All 67 118 185
14. Experiments (IV)
• B – Adaptability Experiment
– Tried to use the built classifier to evaluate 20 book reviews taken
from The SFU Review Corpus (Taboada, Anthony and Voll, 2006).
– The reviews were independently evaluated by 3 master students:
1 = creative texts, 2 = mildly creative texts, 3 = non-creative texts
– Inter-rater agreement Kappa Statistic was too low (perceived
agreement was Po = 0.45) we considered binary classification:
• mildly creative texts = creative Po = 0.633 + considering the majority
class 12 out of 20 were considered creative
• mildly creative texts = non - creative Po = 0.733 + considering the
majority class 4 out of 20 were considered creative
– Since usually there are more non-creative texts than creative
ones, we considered the second situation (mildly creative texts =
non-creative)
– The classifier considered all the reviews as being non-creative
80% accuracy (missed the 4 positive samples, only 1 of these
being considered creative by all 3 students).
Creativity Detection in Texts ICIW201326.06.2013
15. Conclusions (I)
• We presented a model for discriminating creative from non-creative news
articles that was built combining nine different measures.
• The model could be improved by removing or changing the assumptions
that the The Onion articles are always creative.
• The feature selection revealed the following conclusions:
– The lack of creativity was best correlated with Word Norms Fraction which was
expected considering the definitions of creativity and of Word Norms Fraction.
Google Similarity Distance was in the same situation.
– Named Entities analysis showed that they are signs of a creative text as long as
not too many distinct such entities are used.
– Wordnet Similarity was the best evidence for creative texts, while LSA was
similar to the measures of Word Norms Fraction and Google Similarity Distance
in providing a measure for text “usualness” and therefore giving evidence of
non-creative texts. They also have similar weights in the final classifier. ESA had
no influence in the built classifier.
– Less coherent texts were expected to be more creative but coherence score
was found to have no influence in identifying creativity.
Creativity Detection in Texts ICIW201326.06.2013
16. Conclusions (II)
• The second experiment revealed that there are “levels” of
creativity: satire news articles may be more creative than
books reviews, in general.
• Both experiments had around 80% accuracy, showing that
there might be a possibility that the classifier adapts well
However, the lack of true positive examples from the second
experiment makes us be a little cautious in clearly stating
this fact.
• The classifier performed reasonably well at differentiating
articles from The Onion and from other news websites:
– Did we really identified creativity or we detected satire in fact?
– Increasing the size of the data set, and testing it further, could
shed some light on the decision of whether any of the two
assumptions stands and which of them is more adequate.
Creativity Detection in Texts ICIW201326.06.2013