📄 6fanle Submission to the AmericasNLP 2026 Shared Task on Wixarika Image Captioning

Summary

In this paper, the 6fanle team describes the system it submitted to the Wixarika track of the Cultural Image Captioning Shared Task held as part of AmericasNLP 2026. The task is to describe culturally specific images in Wixarika. Only 70 labelled image–caption pairs were available (20 pilot, 50 development), which is too few to train an end-to-end model. The authors therefore built a modular pipeline that pivots through Spanish:

• CLIP retrieves visually similar examples, and their Spanish captions guide Qwen3-VL as it produces several Spanish candidates.
• The official Sheffield-compatible MT model converts these candidates into Wixarika.
• A character 5-gram language model reranks the Wixarika outputs.
• An automatic cleanup step replaces degenerate outputs with other candidates the pipeline had already generated.

The selected setup scored 19.1468 chrF++ in local 5-fold cross-validation, above the organizer baseline of about 17.7. On the official test it reached 19.1569 chrF++, and the team placed third overall. Error analysis shows remaining problems: leftover Spanish words, repetition and overly generic captions. The work suggests that pairing a high-resource pivot language with lightweight target-side reranking is a workable approach for extremely low-resource multimodal generation.

Key points

• Data:
  • 70 official pilot and development examples were used for validation, retrieval and LM training.
  • The test set had 201 images.
  • 9,960 AmericasNLP 2023 Spanish–Wixarika parallel pairs supplied Wixarika text for the character LM.
  • 65 images from the Wixarika Research Center, with captions rewritten into short form, served only as retrieval support.
• Components:
  • openai/clip-vit-large-patch14-336 retrieved the top-4 neighbours.
  • Qwen/Qwen3-VL-8B-Instruct produced 8 sampled candidates per image (temperature 0.85, top_p 0.95).
  • The MT model used beam size 5.
  • Everything ran on a single A100 40GB.
• Results:
  • Local 5-fold CV gave 19.1468 chrF++, with a fold standard deviation of 0.6601.
  • The official test gave 19.1569 chrF++ and 0.02145 CIDEr.
  • In human evaluation the mean rating was 2.48, placing the team third overall.
• Rejected variants:
  • Word-list output filter: 18.9689.
  • Restrictive prompt: 16.9403.
  • Decoding sweep: 18.4174 (2-fold screen only, so not directly comparable).
• Error analysis:
  • Cleanup changed 37 captions and cut flagged bad outputs from 20 to 7.
  • In the final output, 35 predictions still contained Spanish residue and 17 showed repetition or degeneration.
  • Three groups of identical captions were found.
  • The code is released on GitHub.