Harnessing Context Incongruity for Sarcasm Detection (Joshi et al 2015)

Gist

• The key part of this paper is that incongruity e.g. clashes in sentiment are central to the detection of sarcasm
• "It must be noted that our system only handles incongruity between the text and common world knowledge (i.e. the knowledge that 'being stranded' is an undesirable situation and, hence, 'Being stranded in traffic is the best way to start my week' is a sarcastic statement)." (p 758)
• "This leaves out an example like 'Wow! You are so punctual' which may be sarcastic depending on situational context" (p 758)
• Explicit Incongruity is where there are polarity signifying words that make the clash in sentiment apparent
• Implicit incongruity is where there are phrases that imply a particular sentiment conventionally. These are the ones that seem the most interesting to see how they deal with them.

Dataset

Primarily focused on tweets.

• Tweet-A (5208 Tweets, 4170 sarcastic) Downloaded by looking for certain hash tags (#sarcasm, #sarcastic adn #notsarcastic) and then did a rough quality control check to make sure that they made sense, removing wrongly labeled examples.

• Tweet-B (2278 tweets, 506 sarcastic) manually labeled for Riloff et.al 2013. I suspect what they're doing here is trying to balance the class distributions for this since predicting sarcastic tweets using the Tweet-B dataset would be quite difficult.

Discussion board datasets

• Discussion-A (1502 discussion board posts, 752 sarcastic). Obtained from the Internet Argument Corpus (Walker et al. 2012). Manually annotated,. 752 sarc and non-sarc posts are selected randomly.

ML System

Detecting incongeruity

• Identifying phrases with implicity sentiment
• Obtained using algorithm given in Riloff et al. (2013) but extract both possible polarities for both nouns and verbs
• Keeping subsumed phrases "(i.e. `being ignored' subsumes 'being ignored by a friend')"
• Riloff et al. 2013 used these phrase as part of rules while this approach is a ML approach that uses them as features.

Features

• Unigrams
• Number of capital letters
• Number of emoticons and lol's
• Number of Punctuation marks
• Boolean feature indicating whether implicitly incongruous phrases were extracted.

Explicit Incongruity features

"""

• Number of times a word is followed by a word of opposing polarity
• Length of largest series of words with polarity unchanged
• Number of positive words
• Number of negative words
• Polarity of tweet based on words present """

Analysis

• Ran into errors with subjective things (Maybe this would be resolved if they wre able to look more closely at a user's history)
• Errors when there was incongruity but it was not within the text
• Incongruity due to numbers causes errors, here's the example they provide "going in to work for 2 hours was totally worth the 35 minute drive"
• Pieces of sarcastic text embedded in a larger non-sarcastic text were harder to identify.
• Politeness of sarcasm introduced difficulties.