A Controllable Text Generation Framework Based on Appraisal Theory and Support Vector Machine for Enhanced Contextual Alignment

The recent progress of generative models like GPT-4, though achieving unprecedented fluency in text, are still limited in controlling evaluative resources including attitudes, judgment, appreciation, etc., which risks output misalignment with the surrounding context settings in applications like news writing, teaching, and commercial advertisement, etc. To fill this gap, we present a controllable text generation architecture based on Systemic Functional Linguistics (SFL) in particular, appraisal theory, for the controllability and adaptability optimizing of evaluative resources. The designed architecture builds a hierarchical evaluative lexicon in terms of affect, judgment, and appreciation to optimize evaluative expressions adaptively and meanwhile a semantic evaluation measure called Appraisal Adaptability Index (AAI), based on cosine similarity to assess the alignment degree between output and intended evaluative goal, is also introduced. A two-step optimization scheme is designed, where (1) a prompt engineering embedded evaluative parameter at input-level helps guide the generator's output; and (2) a reinforcement post-generation process is performed with AAI as the reward signal in Proximal Policy Optimization (PPO) setting to refine the output. To objectively evaluate the alignment accuracy of evaluative expressions, a SVM-based evaluation module that classifies whether the generated evaluative expressions are appropriate or not from domain perspective is suggested. Compared to the baseline generative models, experimental findings indicate, on average, a 5% increase in AAI score (p < 0.01), and the results of human study demonstrate the enhanced contextual relevance and persuasiveness of the adapted generation. Theoretically, this paper develops a new “linguistic constraint + evaluation-informed optimization” framework, thus advances the literature of controlling-generation; practically, the framework can be used for domain-adaptive, context-aware text generation, and its usability scope covers more areas, like tailored education, public speech, targeted promotion, and etc; In addition, it will serve as a baseline to future related studies on cross-modal appraisal adaptivity in multi-modality generation.