Learning the Roles of Directional Expressions and Domain Concepts in Financial News Analysis

Rapid development of natural language processing technologies has paved way for automatic sentiment analysis and emergence of robo-readers in computational finance. However, the technology is still in its nascent state. Distilling sentiment information from unstructured sources has turned out to be a complicated and strongly domain-dependent problem. To emulate the human ability to recognize financial sentiments in natural language by using machines, we need to provide them with (i) necessary ontological knowledge on the relevant domain-concepts, and (ii) learning strategies that help the machines to combine this knowledge with the syntactic structures extracted from text. In this paper, we present a knowledge-driven tree kernel framework for sentence-level analysis of financial news sentiments. Comparisons with linear kernels and classical lexicon-based systems suggest that significant performance gains can be achieved by incorporating information on financial concepts and their grammatical context. The framework is decomposable into learning, knowledge and syntactic structure components. Contribution of each part is separately examined using a human-annotated phrase-bank with close to 5000 sentences collected across a number of financial news sources. The proposed sentiment analysis framework is flexible and can be applied also outside financial domain. To evaluate cross-domain performance, a further comparison of the algorithms is done with datasets from non-financial domains including movie reviews and general political discussions.