This study proposes an annotated dataset for Feature-based Sentiment Analysis (FbSA) in the financial domain. The dataset is constructed based on a two-step annotation process, which consists of the SSP (Semi-automated Symbolic Propagation) methodology and supplementary annotation by a human annotator.
Financial domain research primarily relies on numerical data, but text data also provides crucial information that is difficult to quantify. Corporate articles offer various incidents about companies, and FbSA aims to categorize events related to companies and extract both positive and negative news from corporate articles.
Unlike sentiment analysis at the sentence level, FbSA requires advanced language knowledge to understand the semantic information for each element in a sentence. It also involves extracting relational information between vocabularies. Moreover, typical sentiment expressions are not commonly used in financial news data. Instead, various combinations of numerical nouns, predicates indicating ups and downs, and event nouns appear in different aspects, and the polarity also changes depending on the combination of these elements. Therefore, building high-quality annotated data for machine learning is a laborious task.
To achieve effective semantic information annotation, linguistic resources linked to the DECO Korean dictionary and LGG (Local Grammar Graph) were utilized. These linguistic resources consist of categorized patterns for entities, features, and sentiment expressions. They can be applied to the raw text to annotate single-word and multi-word expressions, as well as adjacent feature-sentiment expression pairs.
The linguistic resources constructed in this study reflect the aforementioned characteristics. The resources were augmented through a bootstrap method, which involved a cycle of semi-automatic annotation on the data and the addition of new expressions acquired during the annotator''s annotation process to the resources. It was confirmed that the final linguistic resources showed an f1-score of 0.90, demonstrating their contribution to increasing the efficiency of the actual annotation process.
The annotated data for the financial domain FbSA proposed in this study can be effectively extended to reflect changes in the stock market and the latest company and product information by utilizing structured linguistic resources. Furthermore, it is expected that the annotated data will be used for sophisticated stock price prediction in conjunction with numerical information.