Sarcasm Detection | A classification problem

I do not post any links to the papers because whatever I summarize is full and sufficient to get started.

Paper 1

  1. Three types of sarcasm: Positive sentiment, negative sentiment, no sentiment. Sentiment here can be classified as ironic and further as euphemism, analogy etc.
  2. Identify sarcasm by bragging, in speech, verbal irony(change in tone) and un-situational, being bitter towards entity, predefined patterns of irony, hashtag based.
  3. Raise of sarcasm: Failed expectation, pragmatic insincerity, negative tension and the presence of a victim. 2 sets of features observed: implicit and explicit incongruity-based features (incongruity=unpleasant). Using word extensions, number of flips, pointedness and word-shaping. Also POS sequence and semantic imbalance as features.
  4. Responses: Smile/laughter, non-verbal(eye tracking), literal reply, change of topic or no response.
  5. Other ques to identify sarcasm: Like prefix, embedded in actual words
  6. Representation proposed in that paper: 6 tuple <S, H, C, u, p, p’>. Speaker S generates an u erance u in Context C meaning proposition p but intending that hearer H understands p’
  7. Cross checks experiences or past situations ie. appraisal that it is an unpleasant event, ridicule in tone/words, drop not
  8. Two machine translation systems [based on Moses and RNN] to obtain non-sarcastic interpretations of the sarcastic sentences.
  9. (i) Tweets: one line, usually sarcastic or non. (ii) Long text: mix of sarcastic and non. here, writer’s attitude must be captured also called opinion mining about entities.
  10. Approaches: (i) Identify sarcasm through specific evidence. (ii) Odd words out of context — negative statement in positive sentence/+ve verb in -ve situation and its order (iii) +ve adjective, -ve noun (iv) Measure relatedness between WordNet based similarity.

Below are few lines from the paper

“semantic similarity, emoticons, counterfactuality, etc. introduce features related to ambiguity, unexpectedness, emotional scenario”

“include seven sets of features such as maximum/minimum/gap of intensity of adjectives and adverbs, max/min/average number of synonyms and synsets for words in the target text, etc.”

Algorithms

  1. SVM-based classifier and {binary}Logistic Regression with chi-square test to get discriminating feature or SVM-Perf
  2. SVM-HMM to incorporate sequence nature of output labels in conversation, SEARN as sequence labelling algorithm.
  3. Naive Bayes, Decision trees, AdaBoost, Random Forest
  4. Ensemble-based classifiers (bragging, boosting), Fuzzy Clustering

Problem: say chinese homophonic sentences

DL approaches

Word embeddings, CNN, LSTM, recursive SVM, Deep Convolutional Network.

Future enhancements

ISSUES: Manual or distant supervised datasets, skewed data in labeled datasets, sarcasm and irony.

  1. semi-supervised pa ern discovery.
  2. for now, types of sarcasm are not correctly handled.
  3. Future work can benefit from reporting which types of sarcasm are proving to be difficult for different approaches.
  4. like-prefixed and illocutionary are not much explored.
  5. DL based architectures are not much famous.

--

--

Love podcasts or audiobooks? Learn on the go with our new app.

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store