Pablo Gervás

Computational narrative—the automatic sifting, composition and interpretation of stories—needs raw material: event logs rich enough to yield tellable stories. Social simulations are an attractive source but only if their logs are embodied, eventful and interconnected so that concurrent threads share the coordinates a telling must weave. We present a dependency-light multi-agent social simulation built to that specification as a synthetic log generator for downstream narrative tasks. Its distinguishing feature is that perception follows from position: agents encounter one another by co-location, and every event records the set of agents present to witness it. That explicit, queryable who-perceived-what relation—left disembodied or implicit by prior simulations—is what focalised composition needs, and recovering it is the gap this work addresses. Agents negotiate shared activities and exclusive relationships directly, driven by affect and per-agent temperament. Run in a calibrated, seeded free mode, social structure—couples, jealousy, affairs, break-ups—emerges unscripted from local interaction: across 150 random casts the structural invariants hold in every run, while dramatic outcomes vary widely from seed to seed, and a disjoint held-out population of 150 fresh seeds reproduces every calibrated band. A narrative-interest diagnostic instruments the logs, separating routine from active runs and reading off which dramatic configurations occur.

In this paper we study how external affective information should be integrated into a compact transformer-based text classifier. Rather than treating affective features as signals to be appended directly to the representation, we examine whether their contribution should be controlled through lightweight fusion mechanisms. The comparison focuses on scalar-gated fusion versus plain concatenation, using DistilBERT as the textual backbone and four affective resources: the NRC VAD Lexicon, VAD-BERT, Ekman-style emotion scores, and SenticNet. The evaluation is conducted on two English corpora with different label structures: a seven-class MentalHealth dataset and the fine-grained GoEmotions benchmark. Across both corpora, scalar gating consistently matches or outperforms concatenation in terms of Macro-F1. On MentalHealth, scalar gating improves all directly comparable configurations. On GoEmotions, it achieves the best overall Macro-F1 and improves most matched comparisons. Beyond static feature integration, we introduce affective flow (EmoFlow) representations derived from VAD-BERT, which model the evolution of valence, arousal, and dominance across segments of a text. These dynamic representations do not surpass the strongest static lexical resources in absolute performance, but they provide consistent improvements within the VAD-BERT family, particularly when combined with scalar gating or cross-attention. Our contribution is twofold. First, we show that a lightweight scalar gate provides an effective and interpretable mechanism for adaptively integrating low-dimensional affective side information into transformer-based classifiers. Second, we introduce affective flow representations that explicitly model how affect evolves within a document, enabling the analysis of both adaptive resource selection and intra-document affective dynamics. Together, these results suggest that the key issue is not only which affective resources to use, but also when and how they should influence the model.

AI‐Cinema is a hybrid neural‐symbolic framework addressing the critical challenge of preserving cultural authenticity in Arabic movie scenario generation. The framework integrates transformer‐based neural language models (AraT5‐base and AraGPT2‐medium) with symbolic reasoning encoded in OWL‐DL ontologies and SWRL rules to ensure linguistic fluency, narrative coherence, and cultural preservation. AI‐Cinema introduces a three‐tier architecture comprising a data layer, cultural embedding layer, and scenario generation layer. Central to its design is an attention‐based cultural embedding mechanism leveraging ArabicVerbNet (12,500 culturally annotated verbs) and ArabicNameNet (3653 names with regional annotations), complemented by a mathematically grounded harmony function that dynamically balances neural generation with symbolic constraints. In experiments on 2740 Arabic narratives, AI‐Cinema achieves a BLEU‐4 score of 32.76 (± 0.6), representing a 5.0% relative improvement over AraBERT‐Gen ( p < 0.01, paired bootstrap test) and a 27.5 percentage point gain in cultural preservation metric (CPM) compared with MARBERT‐Gen (82.3% vs. 54.8%). The framework maintains 92.3% dialectal accuracy across Modern Standard Arabic and six regional dialects, with explicit evaluation of code‐switching scenarios. Expert evaluations by 30 Arabic literature scholars demonstrate strong interannotator agreement (Fleiss’ κ = 0.78, p < 0.001), with 87% of evaluators rating generated narratives as culturally authentic (score ≥ 4 on a 5‐point scale). Current limitations include reduced performance on underrepresented dialects (Yemeni: 85.6% and Sudanese: 86.1%) and complex code‐switching scenarios (76.3% for 3+ dialects). The framework incorporates transparent labeling mechanisms for AI‐generated cultural content to address authenticity concerns. All resources are publicly available at https://github.com/Mossab82/AI-Cinema .