The Sentic Demo shows how SenticNet can be exploited for concept-level sentiment analysis. In particular, the demo adopts an ensemble of machine-learning and knowledge-based techniques to perform sentence polarity detection. The demo employs sentic patterns for structuring natural language clauses into a sentiment hierarchy, which is then exploited to infer the overall polarity label (positive versus negative) associated with the input sentence. A sentence such as “This car is expensive but nice”, for example, is supposedly positive, as it expresses a favorable sentiment of the speaker, who is probably going to purchase the product. However, the sentence “This car is nice but expensive” is negative, as it expresses the reluctance of the user to purchase the product. Despite the latter contains exactly the same concepts as the former, the polarity is opposite because of the adversative dependency, which shows that the bag-of-concept model is not always applicable.
Please note that the task of subjectivity detection is not addressed. Hence, the demo assumes that the input sentence is opinionated. Also, please use non-inflected form of auxiliary verbs, e.g., "do not" in stead of "don't" and "i am" in stead of "i'm". Finally, the current version of the demo does not deal with comparative sentences such as "I love iPhone but Android is so much better". The same demo is also available as a real-time Twitter sentiment analysis service: SenticTweety.
Before polarity detection can be performed, multi-word expressions need to be extracted from text. Below is a demo of the concept parser, which the Sentic Demo exploits to quickly identify commonsense concepts from free text without requiring time-consuming phrase structure analysis. From a sentence like “I am going to the market to buy vegetables and some fruits” the parser extracts concepts such as go_buy, go_to_market, market, buy_vegetable, buy_fruit, and some_fruits. The parser makes use of linguistic patterns to deconstruct natural language text into meaningful pairs, e.g., ADJ+NOUN, VERB+NOUN, and NOUN+NOUN, and then exploits commonsense knowledge to infer which of such pairs are more relevant in the current context. For demo purposes, the output is limited to 15 concepts maximum.
Aspect mining is also a key task for sentiment analysis. Below is a demo of the aspect parser, which the Sentic Demo exploits to extract product or service features (aspects) from natural language sentences so that better polarity detection can be achieved. The detection of opinion targets, in fact, is key in correctly calculating the polarity of a sentence in which antithetic opinions about different aspects of the same product are expressed. From a sentence like “the touchscreen is good but the battery lasts very little”, for example, the aspect parser extracts touchscreen and battery.
Finally, the demo enables multimodal sentiment analysis by blending the conceptual and affective information associated with natural language through different modalities. In particular, the module adopts an ensemble feature extraction approach by exploiting the joint use of tri-modal (text, audio, and video) features to enhance the multimodal information extraction process. In preliminary experiments using the eNTERFACE dataset, the system is able to achieve an accuracy of 87.95%. Some sample videos are available here.