Towards Semantic Affect Sensing in Sentences Alexander Osherenko
Goal –Language independent approach for affect sensing in textual corpora containing spontaneous emotional dialogues Method –Extracting features and evaluating resulting datasets by standard data mining approaches considering language independence
Overview Properties of classified corpora Feature extraction Results: –SAL –AIBO –SmartKom Conclusions Outlook
Properties of classified dialogues Corpora may be in different languages No obvious signs of emotional meaning. Utterances are short. Can be grammatically incorrect, contain repairs, repetitions and inexact wordings. Can convey contradicting emotional meaning. Utterances are interdependent (can be seen as a continuous stream of information).
Feature extraction Most frequent (stemmed) utterance words in the current corpus (in most cases only seventh/eighth of the whole list) History as most frequent (stemmed) words in the current and n previous utterances (ditto) No dependence on an affect words‘ list e.g. Whissell‘s dictionary of affect
Dialogue corpora SAL –QUB (Cowie 2006) AIBO –Univ. of Erlangen (Batliner et.al. 2004) SmartKom –Univ. of Munich (Steininger et al. 2002)
SAL Instance – utterance in transliteration 670 in FEELTRACE annotated utterances Agreement – % 3 affect states English corpus FEELTRACE scores (mapped onto classes pos./neutral/neg.)
Evaluation for a three class problem in SAL SMO in WEKA Cross-validation (10 fold) Overall number of words rev.precisionrecallfMeasure#wordshistory maj cc dr em jd
AIBO Instance – paragraph in transliteration 3990 instances Sparse transliteration texts (commands to AIBO) 4 affect states German corpus
Evaluation for a four class problem in AIBO SMO in WEKA Learning/testing sets (1738/2252 resp. 2252/1738) Only words’, not history features Overall number of words – 488 (!) precisionrecallfMeasure#wordshistory
SmartKom Wizard of Oz scenario Instance – turn 817 annotated instances 11 user states German corpus
Evaluation for n class problem in SmartKom SMO in WEKA Cross-validation (10 fold) Overall number of words different affect states#whistoryPRfM joyful- strong joyful- weak surprisedneutralhelplessangry- weak angry- strong joyfulsurprisedneutralhelplessangry joyfulneutralhelplessangry joyfulneutralproblem no problemhelplessangry no problemproblem not angryangry
Conclusions Higher number of words and longer history don’t induce better classification, rather their combination Extracted features can serve as a basis (AIBO results – sparse data, repetitious content) Erroneous classification could have been caused by the discrepancy between the rating and the corresponding text Language-independent features
Outlook Further feature extraction (combination, history of POS groups?) Studying erroneous instances (esp. in SMARTKOM) Multimodality (prosodic/lexical) Application for journalistic articles e.g. movie reviews Is 100% precision the goal?