Exploiting Social Network Structure for Person-to-Person Sentiment Analysis

Exploiting Social Network Structure for Person-to-Person Sentiment Analysis Robert West Stanford University west@cs.stanford.edu Hristo S. Paskov Stanford University hpaskov@stanford.edu Jure Leskovec Stanford University jure@cs.stanford.edu Christopher Potts Stanford University cgpotts@stanford.edu Abstract Person-to-person evaluations are prevalent in all kinds of discourse and important for es- tablishing reputations, building social bonds, and shaping public opinion. Such evaluations can be analyzed separately using signed so- cial networks and textual sentiment analysis, but this misses the rich interactions between language and social context. To capture such interactions, we develop a model that pre- dicts individual A’s opinion of individual B by synthesizing information from the signed social network in which A and B are embed- ded with sentiment analysis of the evaluative texts relating A to B. We prove that this prob- lem is NP-hard but can be relaxed to an ef- ficiently solvable hinge-loss Markov random field, and we show that this implementation outperforms text-only and network-only ver- sions in two very different datasets involving community-level decision-making: the Wiki- pedia Requests for Adminship corpus and the Convote U.S. Congressional speech corpus. 1 Introduction People’s evaluations of one another are prevalent in all kinds of discourse, public and private, across ages, genders, cultures, and social classes (Dunbar, 2004). Such opinions matter for establishing rep- utations and reinforcing social bonds, and they are especially consequential in political contexts, where they take the form of endorsements, accusations, and assessments intended to sway public opinion. The significance of such person-to-person evalu- ations means that there is a pressing need for com- putational models and technologies that can analyze them. Research on signed social networks suggests one path forward: how one person will evaluate an- other can often be predicted from the network they are embedded in. Linguistic sentiment analysis sug- gests another path forward: one could leverage tex- tual features to predict the valence of evaluative texts describing people. Such independent efforts have been successful, but they generally neglect the ways in which social and linguistic features complement each other. In some settings, textual data is sparse but the network structure is largely observed; in oth- ers, text is abundant but the network is partly or un- reliably recorded. In addition, we often see rich in- teractions between the two kinds of information— political allies might tease each other with negative language to enhance social bonds, and opponents of- ten use sarcastically positive language in their criti- cisms. Separate sentiment or signed-network mod- els will miss or misread these signals. We develop (Sec. 3) a graphical model that syn- thesizes network and linguistic information to make more and better predictions about both. The objec- tive of the model is to predict A’s opinion of B using a synthesis of the structural context around A and B inside the social network and sentiment analysis of the evaluative texts relating A to B. We prove that the problem is NP-hard but that it can be relaxed to an efficiently solvable hinge-loss Markov random field (Broecheler et al., 2010), and we …

Content truncated to 500 words