Sentiment Frames Extraction

Sentiment analysis is the highly demanded and rapidly developing branch of the Natural Language Processing. It is widely used in different fields of the NLP. Unfortunately, there are considerable problems in some studies in this branch. Most of them are very narrowly specialized and produce results that are not applicable to a variety of tasks. Due to the fact that most of the previous works are based on thematic datasets, this paper should shed some light on this problem with a different to a standard approach. Despite increasing popularity and usability of such Transformers as BERT and Open AI GPT-2 and their outstanding text predictions, sentiment analysis of non-topical texts is still rather a hard task even for them. The main idea of this work is to broaden the exising sentiment frames lexicon using classification models. In this research, we will implement the method of sentiment framework in a sentiment analysis, which is based on the work of Karnaukhova, Loukachevitch. Sentiment framework is the verb-mediated model of specific connections of predicates that relies on the idea that the predicate could affect the polarity of the subject and object of the sentence (e.g. “X wins Y” make X positive and Y negative). According to Deng and Wiebe, the verb could assign certain sentiment on the words that are connected to it. So, the frame acts as a set of encyclopedic, linguistic and cognitive knowledge. Using Twitter-based corpus of short texts we want to solve the problem of thematic and objective datasets. Based on this data we want to extract information about verbs and convert them into features for classification models. The data from short texts will be extracted by different methods. First one is combined tokenizer from NLTK combined with pymorphy2. Second one is more accurate and highly developed model based on BERT - deeppavlov. Using this different type of methods could give us a first look on the problem and results and after that deep and contextualized research. We believe that this new approach could give us information that could be useful in solving the problem of sentiment analysis. In future usage of sentiment frames combined with modern solutions for classification and embeddings has the potential to be a fresh look at this problem.