Contextualised sentiment analysis in the financial domain
Date
2021-07-02Author
Daudert, Tobias
Metadata
Show full item recordUsage
This item's downloads: 373 (view details)
Abstract
Sentiments and beliefs play an important role in actions and decisions in a market environment; for example, people disliking a brand will tend to avoid products of it or people aiming to reduce their carbon-dioxide footprint will tend to travel less. While these choices and their subsequent actions are driven by sentiment, they all have an economic effect. In our connected world, where society, economy, politics, and many others are linked to one another, one must consider the sentiments emanating from different parties (e.g. The Wall Street Journal, Twitter, analyst reports, and company reports). To determine a market sentiment, it is vital to analyse sentiments while considering their differences to ensure the consideration of different interpretations, expressions, and entity levels targeted by sentiment.
Driven by the emergence of the Internet as a means to share data and the invention of social platforms, an increasing amount of text data is produced every day, with a broader and more diverse set of authors. Furthermore, the Internet increased the availability of trading opportunities and allowed easier access to the markets by the general public. Layman, as well as journalists, politicians, and economists, can participate in the markets and publish their statements and opinions on the Internet via text. Online textual data sources provide a rich opportunity to harvest information about the public and are a prime target to analyse sentiments.
State-of-the-art Financial Sentiment Analysis approaches base their analysis only on a given text which comes with two bottlenecks. First, it makes the detection of implicit sentiments difficult and, second, the role of sentiment contagion is not considered. This thesis presents methods dealing with the mentioned bottlenecks by employing a contextual approach to Sentiment Analysis. We make the following contributions to Sentiment Analysis: First, we define the three novel concepts of sentiment conveyance, linking, and assigning, necessary to model textual sentiment contagion. Second, we investigate whether different data sources can be used to enhance Sentiment Analysis on each other. Third, we introduce the first graph neural network for Financial Sentiment Analysis using text and relationship features based on temporal, word, and entity information. Finally, we present a novel corpus (FinLin), tailored to contextual Financial Sentiment Analysis, which covers four data types from the same period and targets a selection of entities from the automobile sector.