Extant research highlights voter backlash against previous regimes, yet little is known about how political elites in new democracies frame their discourse to distance themselves from the regime's ideological legacy. How do they employ ideological references, and what sentiment do they attach to them? How often do they refer to the former regime and to what extent do they frame such references in ideological terms? To address these questions, we analyze parliamentary speeches, which offer a continuous flow of political discourse and capture the dynamics of ideological bias during democratic transitions. Using the ParlSpeech corpus, we apply Deep Transfer Learning to speeches from Czechia (1993–2016), emerging from a Left-Wing regime, and Spain (1996–2018), transitioning from the Right-Wing Francoist regime in the late 1970s. We fine-tune state-of-the-art RoBERTa and DeBERTa transformer-based models to integrate domain-specific insights from human coders (Laurer et al., 2024). Namely, we employ quasi-sentences categorized into the 56 policy areas of the Comparative Manifesto Project coding scheme. This enables us to detect key dimensions such as attitudes toward democracy, the constitution, minorities, communism, and economic policy. Each sentence’s “sentiment” is then analyzed to uncover ideological framing. To ensure robustness, we complement our analysis with additional methods, including Random Forest classification (e.g., Djourelova, 2023) and Latent Dirichlet Allocation (e.g., Di Cocco & Monechi, 2022). This approach allows us to systematically examine how political elites adapt discourse in response to ideological stigmatization and ideological legacy in new democracies.