Document-level Aspect Sentient Classification is a task to predict user’s sentiment polarities for different aspects of a product in a review. The authors propose using Hierarchical Reinforcement Learning to do the task as it’s more interpretable than other successful neural nets that perform well on DASC.
HRL is pretty cool because it makes (more) intuitive sense as to how it works compared to other neural net architecture. Instead of regular RL, there are different tiers of policies. For DASC, the authors propose using a high-level policy to find the relevant clauses. A low-level policy to select sentiment-relevant words inside the selected clauses, they used a sentiment rating predictor that provides the reward signals to both the clause and word selection policies.
The results of this method were comparable to state of the art methods of aspect sentiment classification. A cool thing that I appreciated the authors, including were ablation studies for HRL. They broke down how each component of the HRL architecture impacted the results.
It was interesting to note that negated sentences didn’t work well with this method. I think negation is an interesting problem because if a neural net truly understands a sentence/clause, then I’d expect it to understand negations easily.