2026-10-06T11:34:38Z https://api.figshare.com/v2/oai

oai:figshare.com:article/32751396 2026-10-01T19:00:31Z category_29416 category_29332 item_type_3 month_year_10_2026

<b>Parsing in Artificial Minds: A Comparative Study of Relative Clause Attachment and Semantic Bias Sensitivity in Arabic- and English-Trained Language Models</b> Shorouq Azzahrani (24249039) Linguistic structures (incl. phonology, morphology and syntax) Comparative language studies Humain AI Cross-linguistic Parsing Semantic Bias Sensitivity Arabic-English Comparison Syntactic Processing <p dir="ltr">This study investigates whether the recently released Arabic-trained language model Humain AI exhibits human-like sentence processing in resolving syntactic ambiguity in Arabic relative clause (RC) attachment, compared to general-purpose models (GPT-5.5 and Claude-Haiku-4.5). Since claims that Humain AI is optimized for Arabic language and culture remain unverified, it is unclear whether language-specific training produces genuine alignment with human Arabic parsing strategies or merely surface-level linguistic fluency. Native Arabic speakers typically exhibit high-attachment preferences in ambiguous RC constructions, whereas English speakers show low-attachment bias. Three language models were tested on 30 Arabic sentences containing RC attachment ambiguities, employing both syntactic and A/B prompts. Stimuli varied in semantic plausibility types: role-based, age-based, and logical contradiction biases. Contrary to expectations, Humain AI demonstrated categorical low-attachment preference identical to English models, failing to reflect human Arabic speakers' high-attachment tendency. However, across all models, sensitivity to semantic plausibility cues varied significantly. Claude-Haiku-4.5 exhibited the strongest overall semantic integration (up to 97% congruency in role bias with A/B prompts), while Humain AI and GPT-5.5 showed notably lower performance. Prompt formatting substantially influenced responses, with A/B prompts consistently eliciting greater semantic sensitivity. Findings challenge the assumption that Arabic-specialized training alone produces human-like language processing. Despite specialized training data, Humain AI exhibited English-like parsing biases, suggesting that large-scale pretraining may not authentically capture language-specific syntactic preferences. Results underscore that LLM linguistic fluency masks fundamental differences from human parsing mechanisms and highlight the critical role of prompt design in shaping model behavior.</p><p dir="ltr"><br></p> 2026-10-01T19:00:31Z Dataset Dataset 10.6084/m9.figshare.32751396.v4 https://figshare.com/articles/dataset/_b_Parsing_in_Artificial_Minds_A_Comparative_Study_of_Relative_Clause_Attachment_and_Semantic_Bias_Sensitivity_in_Arabic-_and_English-Trained_Language_Models_b_/32751396 CC BY 4.0