Methods and features[edit]
Existing approaches to sentiment analysis can be grouped into three main categories: knowledge-based techniques, statistical methods, and hybrid approaches.[45] Knowledge-based techniques classify text by affect categories based on the presence of unambiguous affect words such as happy, sad, afraid, and bored.[46] Some knowledge bases not only list obvious affect words, but also assign arbitrary words a probable "affinity" to particular emotions.[47] Statistical methods leverage elements from machine learning such as latent semantic analysis, support vector machines, "bag of words", "Pointwise Mutual Information" for Semantic Orientation,[6] semantic space models or word embedding models,[48] and deep learning. More sophisticated methods try to detect the holder of a sentiment (i.e., the person who maintains that affective state) and the target (i.e., the entity about which the affect is felt).[49] To mine the opinion in context and get the feature about which the speaker has opined, the grammatical relationships of words are used. Grammatical dependency relations are obtained by deep parsing of the text.[50] Hybrid approaches leverage both machine learning and elements from knowledge representation such as ontologies and semantic networks in order to detect semantics that are expressed in a subtle manner, e.g., through the analysis of concepts that do not explicitly convey relevant information, but which are implicitly linked to other concepts that do so.[51]
Open source software tools as well as range of free and paid sentiment analysis tools deploy machine learning, statistics, and natural language processing techniques to automate sentiment analysis on large collections of texts, including web pages, online news, internet discussion groups, online reviews, web blogs, and social media.[52] Knowledge-based systems, on the other hand, make use of publicly available resources, to extract the semantic and affective information associated with natural language concepts. The system can help perform affective commonsense reasoning.[53] Sentiment analysis can also be performed on visual content, i.e., images and videos (see Multimodal sentiment analysis). One of the first approaches in this direction is SentiBank[54] utilizing an adjective noun pair representation of visual content. In addition, the vast majority of sentiment classification approaches rely on the bag-of-words model, which disregards context, grammar and even word order. Approaches that analyses the sentiment based on how words compose the meaning of longer phrases have shown better result,[55] but they incur an additional annotation overhead.
A human analysis component is required in sentiment analysis, as automated systems are not able to analyze historical tendencies of the individual commenter, or the platform and are often classified incorrectly in their expressed sentiment. Automation impacts approximately 23% of comments that are correctly classified by humans.[56] However, humans often disagree, and it is argued that the inter-human agreement provides an upper bound that automated sentiment classifiers can eventually reach.[57]
Evaluation[edit]
The accuracy of a sentiment analysis system is, in principle, how well it agrees with human judgments. This is usually measured by variant measures based on precision and recall over the two target categories of negative and positive texts. However, according to research human raters typically only agree about 80%[58] of the time (see Inter-rater reliability). Thus, a program that achieves 70% accuracy in classifying sentiment is doing nearly as well as humans, even though such accuracy may not sound impressive. If a program were "right" 100% of the time, humans would still disagree with it about 20% of the time, since they disagree that much about any answer.
On the other hand, computer systems will make very different errors than human assessors, and thus the figures are not entirely comparable. For instance, a computer system will have trouble with negations, exaggerations, jokes, or sarcasm, which typically are easy to handle for a human reader: some errors a computer system makes will seem overly naive to a human. In general, the utility for practical commercial tasks of sentiment analysis as it is defined in academic research has been called into question, mostly since the simple one-dimensional model of sentiment from negative to positive yields rather little actionable information for a client worrying about the effect of public discourse on e.g. brand or corporate reputation.[59][60][61]
To better fit market needs, evaluation of sentiment analysis has moved to more task-based measures, formulated together with representatives from PR agencies and market research professionals. The focus in e.g. the RepLab evaluation data set is less on the content of the text under consideration and more on the effect of the text in question on brand reputation.[62][63][64]
Because evaluation of sentiment analysis is becoming more and more task based, each implementation needs a separate training model to get a more accurate representation of sentiment for a given data set.