--- title: "alexandru412 at MWE-2026 AdMIRe 2.0 — Alexandru-Marian 2026" description: "This system paper from team alexandru412 tackles the Text-Only track of the AdMIRe 2.0 Shared Task." authors: ["Cristea Alexandru-Marian"] year: 2026 venue: "Proceedings of the 22nd Workshop on Multiword Expressions (MWE 2026)" source: "https://aclanthology.org/2026.mwe-1.17/" doi: "10.18653/v1/2026.mwe-1.17" licence: "CC BY 4.0 (ACL Anthology)" licence_url: "https://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/papers/marian-2026-alexandru412-mwe-2026-admire-2/" --- [Elamur Research](https://md.research.elamur.ai/) · [this page in HTML](https://research.elamur.ai/free/papers/marian-2026-alexandru412-mwe-2026-admire-2/) [home](https://md.research.elamur.ai/) / [papers](https://md.research.elamur.ai/free/papers/) / [2026](https://md.research.elamur.ai/free/papers/2026/) / alexandru412 at MWE-2026 AdMIRe 2.0: Advancing Multimodal Idiomaticity Representation # alexandru412 at MWE-2026 AdMIRe 2.0: Advancing Multimodal Idiomaticity Representation **Authors:** Cristea Alexandru-Marian · **Year:** 2026 · **Venue:** Proceedings of the 22nd Workshop on Multiword Expressions (MWE 2026) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2026.mwe-1.17/) · [PDF](https://aclanthology.org/2026.mwe-1.17.pdf) · [DOI 10.18653/v1/2026.mwe-1.17](https://doi.org/10.18653/v1/2026.mwe-1.17) · [OpenAlex](https://openalex.org/W7140138909) **Citations:** 0 citations (openalex-snapshot · 2026-10-14) ## Summary This system paper from team alexandru412 tackles the Text-Only track of the AdMIRe 2.0 Shared Task. The task is to rank five candidate images by how well they depict an idiomatic or literal expression in context, across 15 languages. In this track, the images are seen only through their captions. The author tests whether a purely textual LLM can handle this multimodal problem. The approach fine-tunes Qwen-2.5-7B-Instruct with LoRA on the English and Portuguese training sets only. All other languages are handled zero-shot. Three ingredients support the system: - a list-wise prompt that asks the model to order all five captions at once; - random shuffling of caption order to counter positional bias; - machine translation of non-English contexts into English before inference. The system placed 3rd in the track. Ablations show that fine-tuning, translation and shuffling each add accuracy. Results remain weak for some languages, notably Turkic ones, Igbo and Ecuadorian Spanish, which the author calls a "Linguistic Family Performance Disparity." The work matters because it shows a single-GPU, text-only setup can be competitive while exposing where cross-lingual transfer of idioms breaks down. ## Key points - **Setup:** Qwen-2.5-7B-Instruct was fine-tuned with LoRA on attention projections (W_q, W_k, W_v, W_o) for 3 epochs on EN+PT data. The other 13 languages were inferred zero-shot. - **Main results:** The system ranked 3rd in the Text-Only track. - Best Top-1 scores were PT-PT 0.640 (nDCG 0.848) and EL 0.635. - The weakest was ES-EC at 0.104. - Top-1 scores for Turkic languages were also modest: TR 0.400, UZ 0.342, KK 0.333. - **Fine-tuning ablation:** Fine-tuning clearly beats the raw zero-shot model. For example, KA rose from 0.168 to 0.460 and PT-PT from 0.361 to 0.640. - **Shuffling and translation ablations:** - Dropping option shuffling lowers Top-1 in every language listed. For example, PT-PT fell to 0.455 and KA to 0.196. - Dropping test-time translation reduces UZ from 0.342 to 0.271 and IG from 0.194 to 0.142. - **Caveat:** The paper's figures are not fully consistent. - The text gives Igbo as 14.2%, but the main table shows 0.194. The 0.142 figure matches the no-translation run. - Table 2 is described as a zero-shot comparison, yet it actually reports a no-translation run. - The leaderboard leader ITUNLP scores higher than this system in most languages. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* Fields: [Linguistics & languages](https://md.research.elamur.ai/free/shelves/linguistics/) ## Cite The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{alexandrumarian2026alexandru412, title = {{alexandru412 at MWE-2026 AdMIRe 2.0: Advancing Multimodal Idiomaticity Representation}}, author = {Alexandru-Marian, Cristea}, booktitle = {Proceedings of the 22nd Workshop on Multiword Expressions (MWE 2026)}, year = {2026}, doi = {10.18653/v1/2026.mwe-1.17}, url = {https://doi.org/10.18653/v1/2026.mwe-1.17} } ``` - [← previous AlbanianLLMSafety: A Safety Evaluation Dataset for Large Language Models in Albanian](https://md.research.elamur.ai/free/papers/zaghouani-2026-albanianllmsafety-safety-evaluation-dataset-large/) - [next → Align and Shine: Building High-quality Sentence-aligned Corpora for Multilingual Text Simplification](https://md.research.elamur.ai/free/papers/sanchez-2026-align-shine-building-high-quality/) - [↑ back to papers of 2026](https://md.research.elamur.ai/free/papers/2026/) ## References 4 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Joanne Boisson, Luis Espinosa-Anke, and Jose Camacho-Collados. 2023. Construction artifacts in metaphor identification datasets. In Proceedings of EMNLP, pages 6581–6590. [find](https://scholar.google.com/scholar?q=Joanne+Boisson%2C+Luis+Espinosa%2DAnke%2C+and+Jose+Camacho%2DCollados%2E+2023%2E+Construction+artifacts+in+metaphor+identification+datasets%2E+In+Proceedings+of+EMNLP%2C+pages+6581%E2%80%936590%2E) 2. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and 1 others. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901. [find](https://scholar.google.com/scholar?q=Tom+Brown%2C+Benjamin+Mann%2C+Nick+Ryder%2C+Melanie+Subbiah%2C+Jared+Kaplan%2C+Prafulla+Dhariwal%2C+Arvind+Neelakantan%2C+Pranav+Shyam%2C+Girish+Sastry%2C+Amanda+Askell%2C+and+1+others%2E+2020%2E+Language+models+are+few%2Dshot) 3. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep Reproducibility Details To facilitate the reproduction of our results, we provide the specific hyperparameter configurations and experimental settings used in our final submission. [find](https://scholar.google.com/scholar?q=Jacob+Devlin%2C+Ming%2DWei+Chang%2C+Kenton+Lee%2C+and+Kristina+Toutanova%2E+2019%2E+Bert%3A+Pre%2Dtraining+of+deep+Reproducibility+Details+To+facilitate+the+reproduction+of+our+results%2C+we+provide+the+specific+hyperp) 4. Hyper-parameter Configuration We fine-tuned the model using the following settings: • LoRA Rank (r): 32 • LoRA Alpha (α): 64 • LoRA Dropout: 0.05 • Learning Rate: 1e-5 • Batch Size: 1 (with gradient accumulation steps = 4) • Optimizer: AdamW • Scheduler: Cosine with warmup • Max Sequence Length: 102… [find](https://scholar.google.com/scholar?q=Hyper%2Dparameter+Configuration+We+fine%2Dtuned+the+model+using+the+following+settings%3A+%E2%80%A2+LoRA+Rank+%28r%29%3A+32+%E2%80%A2+LoRA+Alpha+%28%CE%B1%29%3A+64+%E2%80%A2+LoRA+Dropout%3A+0%2E05+%E2%80%A2+Learning+Rate%3A+1e%2D5+%E2%80%A2+Batch+Size%3A+1+%28with+gradient+a)