A student team developed a machine learning software designed to manage to break one of the key challenges in natural language processing: the recognition of sarcasm.
The software was called TrueRatr, a collaboration of students at Cornell University who participated in a technological challenge, and Bloomberg. The project was originally designed to detect ironic negative reviews online selling products, but the tool had its open source, which means that it can gain new applications in the future (maybe a sarcasm detector on Facebook? It would be useful …) .
The complex work goes against the old research on what defines sarcasm. Previous attempts were based on words and phrases that served as clues, such as “yeah, right” (which would be similar to a “Oh, sure,” the good Portuguese sarcastic). However, new research is based on the “feeling of exchange”, that is the use of positive and negative words simultaneously.
Christopher Hong, mentor of the student team, which had already researched the sarcasm detection, gives the example of the following sentence: “I love it when you yell at me.” “Worship” is a positive feeling, while “yelling at me” is something negative. So the sentence should denote sarcasm.
With this sentimental basis, the system was trained and can be tested. The result was 71%, which, while not great, is better than a guy a coin, indicating that it is not guesswork or luck. However, the comparison base was relatively small, with only 50 joint assessments and 50 ratings sarcastic products on Amazon.
The developers decided to turn TrueRatr into something useful for the average consumer: a tool to filter and correct the distortion caused by sarcastic ratings in Mac OS X and iOS applications. The algorithm analyzes the reviews and try to eliminate those considering sarcastic to cause the index to be more honest. For example: the application of Uber, when removed sarcastic analysis, had increased its score of 3.5 stars to 4. At the same time, Grand Theft Auto: Chinatown fell from 4.5 to 3.9 stars.
As the algorithm is still not 100% reliable, these analyzes are still questionable. However, the open source TrueRatr, which can be accessed here, allows developers to collaborate worldwide. Also, it can access the much larger text samples that will allow artificial intelligence improves over time to recognize sarcasm.
Via Ars Technica
No comments:
Post a Comment