Tarun Chintada

Author directory

2026

Movie subtitle translation is inherently multimodal, yet text-only systems often miss visual cues needed to convey emotion, action, and social nuance, especially for low-resource Indic languages (English to Hindi, Bengali, Telugu, Tamil and Kannada). We present a case study on five full-length films and compare two lightweight visual grounding strategies: structured attribute summaries from a 5-minute sliding window and free-text summaries of inter-subtitle visual gaps. Our analysis shows that temporal misalignment between subtitles and frames is a major obstacle in long-form video, often rendering indiscriminate visual grounding ineffective. However, oracle selective grounding, which replaces only the lowest-quality 20-30
In conversational and cinematic subtitle translation, large language models often exhibit stylistic collapse, defaulting to neutral registers that flatten character-specific personas. Prompt-based attempts to recover style frequently introduce instruction drift and semantic degradation. We propose Orthogonal Style Injection (OSI), a parameter-efficient, inference-time intervention that removes global stylistic bias through orthogonal projection at an intermediate layer and injects a character-specific style vector derived solely from English source dialogue. We evaluate OSI for English-to-Bengali, Hindi, Telugu, Tamil, Kannada, Chinese, Japanese, Spanish, French, German, and Italian subtitle translation, spanning 11 typologically diverse target languages. Results show that zero-shot stylistic transfer depends strongly on pretraining alignment. Llama-3.1-8B yields gains in several European languages but degrades on Indic languages, whereas Qwen-2.5-7B transfers style more robustly across Indic and East Asian languages. To probe sociolinguistic mechanisms, we curate a parallel English-Bengali cinematic dataset as a focused morphological stress test. Automatic metrics and human evaluation show that OSI enables meaningful honorific and register shifts while largely preserving semantic fidelity, revealing a dataset-dependent trade-off between stylistic fidelity and translation quality in highly colloquial, morphologically constrained settings.