論 文Papers

CONFERENCE (INTERNATIONAL)

Word-level Sentiment Visualizer for Financial Documents

Tomoki Ito (UTokyo), Kota Tsubouchi, Hiroki Sakaji (UTokyo), Tatsuo Yamashita and Kiyoshi Izumi (UTokyo)

2019 IEEE Conference on Computational Intelligence for Financial Engineering and Economics (IEEE CIFEr 2019), 2019/3

Category:

自然言語処理 (Natural Language Processing) 機械学習 (Machine Learning) データサイエンス (Data Science)

Abstract:
It has a great demand for automatically visualizing word-level sentiment scores in financial documents in the form that even non-experts can briefly understand documents. In this paper, we aim to develop a method for automatically visualizing the original word-level sentiment (i.e., word-level sentiment before considering the contexts in a document) and the contextual wordlevel sentiment (i.e., word-level sentiment after considering the contexts in a document) of each term in a document. To achieve this aim, we develop a method for assigning both original and contextual word-level sentiment scores to words using the Layer-wise Relevance Propagation (LRP) method. The LRP based approach can consider the contextual information in assigning original and contextual sentiments to words, in contrast to the other approaches. Using synthetic and real financial textual datasets, we demonstrated the validity of our LRP based approach. Moreover, we propose two types of novel text-visualization frameworks: local word-level sentiment visualization (LWSV) and global word-level sentiment visualization (GWSV). The LWSV visualizes both original and contextual word-level sentiment of each term in a document. The GWSV visualizes both original and contextual word-level sentiments of documents in concept units. These types of text-visualization should be helpful for understanding financial documents quickly.
Download:

Word-level Sentiment Visualizer for Financial Documents(外部サイト/External Site Link)