| With the rapid development of Internet technology and e-commerce, dramatically increasing online product reviews have become an important opinion resource which is useful for consumers making purchase decisions, enterprise promoting reputation and managing the market. Therefore, automatiely and efficiently mining opinions of product features has grown up to be a popular research topic in the field of sentiment analysis. However, due to the diversities and ambiguities of Chinese language, especially the non-normative expression of user comments, it is faced with many difficulties and challenges.To solve the problem of identifying product features and their opinions in the sentiment analysis field, we researched theoretical approachs and proposed new algorithms of a domain affective ontology to build a feature-based affective space model, which can enhance the ability to understand the semantics of product reviews. It is very useful for easier identifying features and their affective words and improving the judgement accuracy of sentiment polarity of feature-affective pairs. The main works done in this paper are as follows:1. The representation model of domain affective ontology is proposed. Our ontology model, which regards both product feature words and affective words as concept nodes, is used to describe three semantic relations:the relation "part of between a product and its components, the relation "attribute of between a product and its attributes, the relation "associated" between product features and affective words. This representation model is useful for clearly describing the semantic relations of our concepts.2. During extracting work of ontology concepts in the domain affective ontology, this paper mainly studies the collocation extraction of feature words and affective words and the algorithm of feature word similarity for feature concepts clustering. First, we start the construction by collecting some seeds and degree adverbs and design the POS patterns to match the feature words and affective words in review texts. Experiments on37300review sentences was performed to extract the feature-affective pairs, and the results show that our method is very effective in work of concept extraction of our ontology and provides an important basis for identifying sentiment polarity of feature-affective pairs. In order to construct the sharing concept of features, we proposed the algorithm about feature words similarity to cluster all the feature words by their same meaning in context.3. During extracting semantic relations of our ontology, we focus on deciding the association relationship between feature words and affective words. It means that we studies the detail method which identifies the sentiment polarity of feature-affective pairs to eliminate the domain dependence of some affective words. We considered the polarity labels of review sentences and negations to retroactively predict the sentiment polarity of feature-affective pairs. Experiments on our review corpus show the average precision are promoted as19.50%compared to the Ku’s method which determines sentiment polarity of an unknown affective word based on the polarity of individual characters. Note that the individual characters compose the affective word.4. We generate an instance of our domain affective ontology named "laptop" affective ontology. In order to take full advantage of our ontology knowledge to improve the precision and recall, this study puts forward the node matching rules of our affective ontology for sentiment analysis. According to the analysis of different situations, this paper makes corresponding matching rules. Especially, we come up with an algorithm to infer implicit features in target review text. At last, compared with raditional method which based on HowNet, experimental results show that the sentimental prior knowledge of our "laptop" affective ontology can provide the evidence for eliminating the domain dependence of sentiment word and identifying the implicit features. |
PDF Full Text Request