Looking into the Past: Evaluating the Effect of Time Gaps in a Personalized Sentiment Model

This paper concerns personalized sentiment analysis, which aims at improving the prediction of the sentiment expressed in a piece of text by considering individualities. Mostly, this is done by relating to a person’s past expressions (or opinions), however the time gaps between the messages are not considered in the existing works. We argue that the opinion at a specific time point is affected more by recent opinions that contain related content than the earlier or unrelated ones, thus a sentiment model ought to include such information in the analysis. By using a recurrent neural network with an attention layer as a basic model, we introduce three cases to integrate time gaps in the model. Evaluated on Twitter data with frequent users, we have found that the performance is improved the most by including the time information in the Hawkes process, and it is also more effective to add the time information in the attention layer than at the input.