Research to make AI understand "sarcasm" using deep learning

byJason Bolonski

Indirectly convey negative content in a positive wayIrony (Sarcasm)Is derived from an action meaning a mockery or an insult, an ancient Greek word "sarkasmós" (which scrapes meat with teeth), and has been used by human beings since ancient timesrhetoricis. Apart from the meaning of the word, the true feelings of a writer or speaker have been said to be unreadable by computers that can not understand the context and emotions. However, it is said that studies are being conducted to analyze massive amounts of textual data by deep learning, and to discriminate the sarcasm embedded in sentences from the relationship between words and meanings.

Snark Bite: Like an AI Could Ever Spot Sarcasm | NVIDIA Blog
https://blogs.nvidia.com/blog/2018/01/31/ai-detect-sarcasm/

He is the president of Indian Institute of Technology Patna and a professor at BombayPushpak BhattacharyyaIs leading a research team composed of his students, linguists and psychologists and conducting research to detect ironic comments and malicious remarks on the net by deep learning.


Bhattacharyya argues that "There is a reasonable reason to advance the analysis of irony". People in positions who must be concerned about the reputation of the public, such as the heads of state, politicians, celebrities, companies, are checking their reputations on Twitter and other social media. What is being used is analyzing the emotions of those who wrote from sentencesEmotion analysisIt is a technique called. However, Mr. Bhattacharyya says that "Ironicism clearly shows the movement of human emotions, but we can not correctly pick up sarcastic comments in computerized emotional analysis."

For example, if you can not judge that "It's not a good thing" that the life of the battery is the remaining 2 hours is because the computer has tweets saying "The battery of the mobile phone is two hours left and it's still the best." It can not be said that it detected it. According to Bhattacharyya's survey, roughly one-fifth of the sarcastic tweets were "tweets that included numbers," such as "I fell asleep for three hours and today is going to be a long day", but these He also said that he could not pick up much with general emotional analysis.


In order to notice the sarcastic story, I have to understand the context. However, Bhattacharyya leads to the conclusion that "contradictory phrasing and expressive emotional expression is the characteristic of irony". For example, a sentence saying "I love being neglected" means that the positive word "I love you" and the negative word "ignored" do not match. Bhattacharyya makes an algorithm to detect such inconsistency of words. It is a library specifically for deep learningCuDNNSpeed ​​up withTensorFlowFramework to build a neural network using NVIDIA GeForce GTX 1080 Ti GPU. And from a large number of tweets, film criticism, popular comedy drama in the 1990's "FriendsWe prepared a large amount of text data up to the dialogue of "and analyzed data using deep learning. As a result, Bhattacharyya's algorithm was able to detect the sarcasm more accurately than the previous method, especially with tweets containing figures, it was able to detect with accuracy of more than three times as high as 80% is.

Based on the data so far, we also developed Sarcasm Suite, a sarcastic search engine that runs on a browser basis, and SarcasmBot, who creates ironic comments on its own. A user sends SarcasmBot "Mr. Greg EstesWhat do you think about? "When asked, SarcasmBot said" I like Greg, because I really appreciate those who do not have a sense of responsibility at all "I responded to irony.