1. sentiment classification: thumbs-up and down (or plus neutral)

2. sentiment association: to associate a sentiment with a topic or brand

3. fine-grained sentiment extraction: for example, who made the sentiment comment? about which topic or brand this sentiment is about (= 2 above)? how intense is the sentiment? what is the reason of the sentiment? Can the system associate sentiments not only with topics or brands (e.g iPhone), but also with a feature of a brand (e.g. screen) and how well they do so? In addition to sentiments related to emotions about agents (love/hate/happy/annoy etc), can system identify positive or negative evaluations of a topic/brand (cost-effective/poorly-designed) ?  How about the agents' needs and wishlist for brands?  How about agents' positive or negative action towards a brand (including consumers' purchase intent such as will buy; negative actions such as abandon; discontinue the use of)?  What are the brands' functionality (positive features (designed to do what)? Can system identify comparisons between brands/topics (iPhone is better than Blackberry)?

Most learning systems stop at 1 and sometimes at 2. We do all 3 based on deep parsing.

The most popular and easiest is the sentiment classification of documents based on keyword density: they perform well in domains where there are large labeled data available (e.g Amazon review; movie reviews etc), but they are too coarse-grained. They face challenges when they move to a new domain where labeled data are not sufficient for the algorithm to learn a classifier. The more severe challenge comes from 2 when comparative text mentions two brands in proximity with a detected sentiment (I prefer iPhone to Blackberry; iPhone is better than Blackberry) : because they are based on keywords and do not understand the sentence structures, they do not know how to associate the sentiment with iPhone or with Blackberry.

Finally for 3, so far no learning systems have even attempted that degree of fine-grainedness of sentiments in industry, but this is super important for a social media monitoring product which will then be able to support extracting "actionable intelligence" for decision makers. We are one of the first, if not THE first, to do 3. For fine-grained extraction, rules are more flexible to apply, especially after there is a parser built to support it.

Having said that, usually the QA (Quality Assurance) should still focus on 1 and 2, not so much on 3 for cost considerations. We want to make sure the benchmarks reflect the global picture of how well the system performs in sentiments.  As long as the global quality control is there, the fine-grained extraction in 3 cannot go too wrong. But in order to test every detail of sentiment related intelligence, there is huge cost that is required. We cannot afford that, but we perform in a self-adjusting mode: each difference the system makes for any development or change of the system, we developers are demanded to eyeball the results to decide if they are good catches or not. This way, we ensure 3 stays on track and makes improvement every day.

【置顶：立委科学网博客NLP博文一览（定期更新版）】