SENSEI-LIF team Ranked 2nd at the Semeval sentiment analysis
The SENSEI team participated to the Semeval scientific evaluation campaign, under the sentiment analysis track 4.A (polarity detection
in tweets). Polarity detection consists in detecting whether the author expressed a positive or negative sentiment in a text. This task
is a basic building block for analysing social media conversations.
The system developed by SENSEI consists in a family of convolutional neural networks (CNNs) trained from multiple views of the input, and combined with a deep neural network. The different views are created by training various flavors of word embeddings, from lexical, syntactic and semantic evidence, on large datasets collected through the project. They are combined by extracting hidden layers of the CNNs, concatenating them as input of the final classifier. This type of embedding fusion helps characterizing correlations between hidden representations of the data which leads to more robust decisions
The SENSEI system ranked second out of 34 participants, with non significant difference with the first participant. Its detailed description will be available in the Semeval proceedings.