Content area
News has a profound influence on public opinion, shaping both short-term discourse and long-term social attitudes. Social media has amplified this influence, leading to concerns about selective exposure and echo chambers. Concomitantly, it has made public journalism more accessible than ever, creating a news overload. We present computational frameworks to analyze news and related social media discourse by uniting social science theories with language models (LMs) and multiagent social simulations. We employ two methodologies: (1) a data-centric approach that applies when data is available, and (2) a multiagent social simulation approach that applies when data is unavailable or inadequate. We analyze political slant in news based on sentiment toward presidential candidates in election-related news. We find that both traditional online news and social media exhibit political slant, with social media content being more sentiment-driven and exhibiting a stronger slant. User responses to news from politically divergent sources reveal a pronounced moral divide, highlighting an underlying information divide along partisan lines. We analyze conflict-related news by identifying moral and affective framing and apply the war and peace journalism framework to distinguish contrasting narratives. We analyze the coverage of the ongoing war in Gaza. The analysis reveals significant variation in framing across publishers, suggesting partisan news reporting. Moreover, we observe causal associations between publishers, indicative of agenda-setting effects, i.e., framing by one publisher appears to be influenced by the framing of another. Although the use of war and peace frames varies, all publishers incorporate both frames. In the simulation-based approach, we develop a multi-agent model to simulate online user behavior and information diffusion within a social network. We examine the effects of selective exposure to like-minded content on polarization dynamics. Results indicate that selective exposure being high increases network segregation, reduces opinion diversity, and lowers tolerance, though the effects are not strictly monotonic. Notably, low to moderate selective exposure mitigates segregation, enhance intra-community variance, and promote tolerance. However, these conditions yield lower reader satisfaction, incentivizing platforms to favor high selective exposure strategies. Furthermore, under high selective exposure, partisan content creators receive more positive user reactions than neutral ones, raising concerns that such incentives may encourage the production of more extreme content to maximize engagement.