In recent years, since technology has revolutionized human life, plenty of data and information are published on the Internet each day. These shared data contain sentiments (i.e., positive, neutral or negative) toward various topics. Hence, there is a growing need for some techniques and tools which make anticipations about the sentiment of documents. Some sentiment analysis tools extract sentiments for words or sentences separately; moreover, these tools might generate noisy results. Moreover, some documents may contain noisy sentences(not relevant sentences to the context of the document). Therefore, we present some fusion methods to fuse all separated sentiments to have a noise robust value as the whole document sentiment. The proposed robust sentiment analysis tool is called RSF. Initially, the sentiment of sentences are extracted by CoreNLP; afterward, due to the histogram of extracted sentence sentiments, documents are divided into two separated groups. The first group consists of documents which have one main concept and some noisy signals. On the other hand, the second group consists of documents with two main concepts. Afterward, the separated sentiments are fused by using non-linear and linear operator with different loss functions. Moreover, an approach is proposed to evaluate the different loss functions utilized in the fusion step. This approach is based on spaCy and neural network which calculate the train and test errors of different fusion methods to determine the most efficient loss function. Eventually, a news corpus of English news of ISNA is introduced in this paper.

Keywords