Researchers from the Caltech University are now successful in showing efficiency of machine learning programs to monitor social media conversations online. These programs can track the evolution of these conversations, which in turn can become an efficient method to identify online trolling.
This research project brings together experts from different domains. Anima Anandkumar is from the artificial intelligence domain whereas, Michael Alvarez is a professor of political science. After presenting their research at the 2019 Conference on Neural Information Processing Systems in Canada, they are planning to take it a step ahead. There are other experts in the team such as a postdoctoral scholar Anqi Lui and Nicholas Adams Cohen, a Ph.D from Stanford.
Preventing incidences of online harassment needs rapid identification of negative, offensive, and harassing social media posts. This is turn needs constant monitoring of interactions occurring online. Existing methods to gather such data from are completely automated and not entirely interpretable. Else, at times, they just rely on some static set of words or phrases. This has a risk of becoming obsolete pretty quickly and thus can be unusable after a point of time. Thus, these researchers believe neither methods are optimal.
Putting Words in Vector Space
To solve this issue, these researchers are now using the GloVe model. GloVe stands for Global Vectors for Word Representation. This model is able to discover and track new and more relevant keywords. It is a word-embedding model, which allows it to represent words in a vector space. The ‘distance’ between the words is then a measure of their semantic or linguistic similarity. Starting from one word, the model can then find a series of other words or phrases that are in close association with the first. This will present them with a huge cluster of words that are relevant and more interpretable.