A comprehensive study of domain-specific emoji meanings in sentiment classification

Abstract

The inclusion of emojis when solving natural language processing problems (e.g., text‐based emotion detection, sentiment classification, topic analysis) improves the quality of the results. However, the existing literature focuses only on the general meaning transferred by emojis and has not examined emojis in the context of investor sentiment classification. This article provides a comprehensive study of the impact that inclusion of emojis could make in predicting stock investors’ sentiment. We found that a classifier that incorporates domain-specific emoji vectors, which capture the syntax and semantics of emojis in the financial context, could improve the accuracy of investor sentiment classification. Also, when domain-specific emoji vectors are considered, daily time-series of investor sentiment demonstrated additional marginal explanatory power on returns and volatility. Further, a comparison of conducted cluster analysis of domain-specific versus domain-independent emoji vectors showed different natural groupings of emojis reflecting domain specificity when special meaning of emojis is considered. Finally, domain-specific emoji vectors could result in the development of significantly superior emoji sentiment lexicons. Given the importance of domain-specific emojis in investor sentiment classification of social media data, we have developed an emoji lexicon that could be used by other researchers.

Appendix: Details of Hk++

Hk++ (Olech and Paradowski 2016) adopts the Gaussian mixture model (GMM) where clusters (mixtures) are defined as Gaussian distributions (Bishop 2006). In GMM, clustering with n clusters is described with the following probability distribution function (pdf):

$$G\left(\psi ,\mu ,\varSigma \right)=\sum _{i=1}^{n}{\psi }_{i}N\left({\mu }_{i},{\varSigma }_{i}\right)$$

(16)

where $\mu$ is a set of all cluster centers (${\mu }_{i}\in \mu$), $\varSigma$ is a set of all cluster covariance matrices (${\varSigma }_{i}\in \varSigma$) and $\psi$ is a set of mixture weights (${\psi }_{i}\in \psi$) such that ${\psi }_{i}$ ranges from 0 to 1 (inclusive on both sides) and $\sum _{i=1}^{n}{\psi }_{i}=1$. $N\left({\mu }_{i},{\varSigma }_{i}\right)$ is a multivariate Gaussian distribution of the $i$-th mixture with center ${\mu }_{i}$ and covariance matrix ${\varSigma }_{i}$. Hk++ uses multiple GMMs and recursively (breadth-first) arrange them in a hierarchical structure. That hierarchical structure is an object cluster hierarchy, where every parent with its immediate children is represented by an extended GMM pdf function. In the extended version, GMM models the child nodes with the presence of an additional background component $N\left({\mu }_{B},{\varSigma }_{B}\right)$ representing the parent node:

$$G\left(\alpha ,\psi ,\mu ,{\mu }_{B},\varSigma ,{\varSigma }_{B}\right)=\alpha N\left({\mu }_{B},{\varSigma }_{B}\right)+\left(1-\alpha \right)G\left(\psi ,\mu ,\varSigma \right)$$

(17)

where the balance between a parent and its children is defined using parameter $\alpha$ that ranges from 0 to 1 (inclusive on both sides), ${\mu }_{B}$ is the center of a parent node, ${\varSigma }_{B}$ is the covariance matrix of a parent node, and the rest of variables are defined as in Eq. (16). The role of the background component is to capture sparsely distributed data, enabling the child clusters to be discovered on a filtered (denoized) set of data points (Olech and Paradowski 2016). Model parameters are estimated using the expectation maximization (EM) algorithm (Bishop 2006). This algorithm consists of two steps: reassignment of points to mixtures (expectation step), and recalculation of ${\mu }_{i}$, ${\varSigma }_{i}$ for every mixture (except for the background component) based on the updated assignment (maximization step). These steps are performed until convergence of the pdf function or up to the predefined number of iterations. In the implemented visual clustering procedure, the Hk++ parameters, including the number of iterations, are dynamically set up by the human operator.