The work, led by Abubakar Sulaiman Gezawa of Sanming University together with colleagues at Hanshan Normal University and Bayero University, tackles three persistent gaps in existing multimodal translation systems. The first is architectural: because captioning and translation are usually trained as independent modules, gradients cannot flow between them, so errors propagate irreversibly and visual context is lost after the first stage. The second is scaling: many multilingual systems require a separate decoder or substantial parameter growth for every additional language. The third, and arguably most subtle, is grounding: attention mechanisms in conventional systems may interact with visual features superficially, but they routinely fail to preserve fine-grained details such as gender, count, spatial relations and object attributes throughout the translation process.

The team’s answer is a unified architecture built around a Transformer decoder that supports two operating modes with shared weights. In captioning mode, the model looks at an image and generates an English description. In translation mode, a small learnable vector called a language token is prepended to the embedded English caption, steering the same decoder toward the desired target language. Because the captioning and translation objectives are optimized jointly through a combined loss, the visual encoder receives complementary gradients from both tasks, encouraging representations that are simultaneously good at recognizing objects and at preserving the attributes that translators need—exactly the information that a cascaded system loses the moment it hands its text to a second model.

The first core technical innovation is what the authors call cross-modal representation alignment. A ResNet-50 convolutional encoder, pretrained on ImageNet, converts each input image into a seven-by-seven grid of 49 visual tokens, projected into the same 512-dimensional space used by the language model. These tokens are then made available to every one of the six decoder layers through a tri-attention scheme: causal self-attention over previously generated tokens, cross-attention to the English caption’s encoder outputs, and cross-attention to the visual features, each with eight attention heads and independent query, key and value projections. The ordering is deliberate. Self-attention first establishes linguistic context; caption cross-attention grounds the representation in the semantic content of the source description; and visual cross-attention acts last, refining the representation with direct evidence from the image so that details such as plurality or gender can re-enter the computation even if the English caption under-specified them.

The second innovation is the language-token conditioning strategy. Instead of training a dedicated decoder per language, the model learns a single 512-dimensional embedding for each of five typologically diverse targets—Chinese, French, Arabic, Spanish and Hindi—spanning Sino-Tibetan, Romance, Semitic and Indo-Aryan families and four different scripts. Placed at position zero of the input sequence, the language token participates in self-attention, so every caption and generated token can attend to it, and it also joins the visual cross-attention stream, allowing language-specific visual interpretation. When generating Arabic, for instance, where gender agreement is grammatically mandatory, the token can help route visual gender information to the right morphological decisions. The authors report that learned tokens specialize: the French token attends strongly to determiner positions where gender must be resolved, while the Arabic token focuses on verb morphology. A shared vocabulary of 60,000 subword units, learned with SentencePiece across all languages, enables cross-lingual transfer of common subwords, particularly between French and Spanish.

The empirical results, obtained on MS COCO for captioning and the OPUS-100 parallel corpus for translation, show consistent gains over cascaded baselines. The unified model achieves a BLEU-4 score of 44.8 for English captioning on the MS COCO Karpathy test split, alongside a CIDEr of 153.2, and an average BLEU of 41.1 across the five target languages—improvements of 2.2 and 2.0 points respectively over cascaded systems. COMET scores, a neural metric that correlates better with human judgment of semantic fidelity, average 0.732 for the proposed model compared with roughly 0.698 for text-only baselines. Perhaps most striking is the efficiency: the language-token approach requires 218 million parameters in total, a 43 percent reduction compared with the 312 million needed when separate decoders are trained for each language, while actually producing slightly better average translation quality.

The gains are not distributed evenly, and that pattern is instructive. The largest improvements over cascaded baselines appear for Arabic and Hindi, languages with rich inflectional systems in which gender, number and case must be marked inside the word itself. In a cascaded pipeline, an English caption saying “two children playing” can lose the count information needed to generate the correct Hindi plural; the unified model preserves that cue through per-layer visual attention. French and Spanish, which mark gender largely through articles and function words that can sometimes be recovered from context alone, show more modest gains of about 1.9 BLEU. Manual inspection of 500 errors confirmed the mechanism: cascaded baselines produced 2.3 times more attribute errors and 1.8 times more spatial relationship errors than the unified model, which cut overall error rates by 58 percent relative to the cascaded approach.

An ablation study disentangled the contribution of each component. Removing cross-modal alignment caused the largest degradation, dropping CIDEr by 8.6 points and average BLEU by 1.6, while adding only 8 million parameters—evidence that the benefit comes from architectural integration rather than sheer capacity. Replacing joint training with a conventional two-stage scheme cost 1.8 BLEU, confirming that end-to-end optimization acts as a useful regularizer: the visual encoder and caption decoder develop representations that balance fluent English generation with preservable visual semantics. A late-fusion variant that injected visual features only at the decoder’s final layer fared worst among the ablations, underscoring the paper’s central claim that visual grounding must operate throughout decoding, not as an afterthought, to influence agreement, word order and inflection.

The authors are candid about the system’s remaining weaknesses. Visually similar objects such as cups and bowls are occasionally confused, a limitation partly traceable to the coarse seven-by-seven feature grid, where adjacent objects can blend within the same cell. Attribute hallucination persists in lower-resource languages; in one documented case, an image of two children with a red ball was rendered in Hindi with the wrong count, and a woman holding an umbrella became a man in Arabic. Spatial prepositions such as “behind” versus “next to” are sometimes flattened, because encoding three-dimensional depth from a two-dimensional projection remains genuinely hard. These failures occurred in only 4.7 percent of inspected test examples, however, and appeared two to three times more frequently in the cascaded systems the new model is designed to replace.

Beyond the benchmark numbers, the researchers frame the work in terms of biologically inspired computation, arguing that per-layer cross-modal modulation echoes the layered multisensory integration observed in human cortex, while the language tokens act as a sparse gating mechanism akin to experience-dependent linguistic plasticity. The practical implications reach further than the analogy: accessibility services for visually impaired users in diverse linguistic communities, multilingual educational content, and cultural heritage documentation all depend on descriptions that stay faithful to a visual scene across languages. With training completed in about 18 hours on four NVIDIA A100 GPUs and inference running beam search at modest width, the architecture is deliberately positioned for deployment rather than spectacle. Future directions the team outlines include depth-aware visual encoders for spatial reasoning, language-aware optimization for low-resource languages, and extension of the token-conditioning scheme toward zero-shot translation of unseen languages—steps that could turn a strong multilingual captioner into a genuinely universal visual translator.

Subject of Research: A unified multimodal neural model that integrates image captioning and multilingual machine translation with cross-modal alignment and language-token conditioning

Article Title: Cross-modal alignment with language tokens for unified multilingual image-to-text generation

Article References: Cross-modal alignment with language tokens for unified multilingual image-to-text generation. (n.d.). https://doi.org/10.3389/fncom.2026.1826455

Image Credits: AI Generated

DOI: 10.3389/fncom.2026.1826455

Keywords: multimodal machine translation, image captioning, cross-modal alignment, language tokens, neural machine translation, Transformer decoder, ResNet-50, MS COCO, OPUS-100, BLEU score, multilingual NLP, vision-language models