Social media has become a dominant form of communication for individuals and for companies looking to market and sell their products and services. Properly understanding and responding to customer ON social media platforms is critical for success, but it is incredibly labour intensive. Hence, U.S. Computer science researchers have developed a technique that accurately detects sarcasm in a social media text. The team’s findings were recently published in the journal Entropy.
The team effectively taught the computer model to find patterns that often indicate sarcasm and combined that with teaching the program to correctly pick out cue words in sequences that were more likely to indicate sarcasm. They taught the model to do this by feeding it large data sets and then checked its accuracy.
Sarcasm has been a major hurdle to increasing the accuracy of sentiment analysis, especially on social media, since sarcasm relies heavily on vocal tones, facial expressions and gestures that cannot be represented in the text. Recognising sarcasm in textual online communication is no easy task as none of these cues are readily available.
The researchers proposed an approach consists of five components: Data Pre-Processing, Multi-Head Self-Attention, Gated Recurrent Units (GRU), Classification, and Model Interpretability. Data pre-processing involves converting input text to word embeddings, which is required for training a deep learning model. To this end, they first apply a standard tokeniser to convert a sentence to a sequence of tokens, then we employ pre-trained language models to convert the tokens to word embeddings.
These embeddings form the input to our multi-head self-attention module, which identifies words in the input text that provide crucial cues for sarcasm. In the next step, the GRU layer aids in learning long-distance relationships among these highlighted words and output a single feature vector that encodes the entire sequence. Finally, a fully connected layer with sigmoid activation is used to obtain the final classification score.
The researchers show the effectiveness of the approach by achieving state-of-the-art results on multiple datasets from social networking platforms and online media. Models trained using our proposed approach are easily interpretable and enable identifying sarcastic cues in the input text which contribute to the final classification score. They visualise the learned attention weights on a few sample input texts to showcase the effectiveness and interpretability of their model.
As reported by OpenGov Asia, other U.S. researchers have also developed an AI tool, called CitizenHelper to provide insights into online behaviour. This tool can sort through millions of tweets to identify behaviours that could assist emergency agencies and give them an understanding of the population’s attitudes.
The U.S. researchers specifically use this tool to gain insight into people’s response to COVID-19 in the Washington D.C., Maryland, and Virginia (DMV) area. The tool uses artificial intelligence (AI) techniques to filter the posts and then determine the relevance and information level of each tweet.
The head of the research team says that he and his team are extracting intelligence from social media posts on several key subjects. They include risks, symptoms, compliance with social distancing, and more relevant information on COVID-19 using a human-artificial intelligence (AI) teaming approach.
The tool helps these CERTs to scale work that would be difficult for humans to do alone. The head of the research team says that humans are good at contextual understanding to filter content but they cannot scale. Machines, on the other hand, are good at scaling, but they do not deeply understand the context very well. Hence, a human-AI teaming approach is invaluable. The algorithms need humans to help them improve their accuracy. CitizenHelper allows this very seamless interactive mechanism for humans and computers. The humans can provide feedback to the machine on what the machine has predicted.