Analyzing the Impact of Typological Features on Cross-lingual Transfer for Generative LLMs

ACL ARR 2025 July Submission328 Authors

Abstract: Recent advances in large multilingual language models have enabled impressive zero-shot cross-lingual transfer capabilities. However, their performance remains uneven across languages, with underrepresented languages often lagging behind. While prior research has explored typological similarities to explain performance disparities, it has largely ignored external factors such as resource availability and has primarily focused on encoder-only models. In this study, we investigate the interplay between typological features and resource-related factors in zero-shot reading comprehension tasks using decoder-only models. Specifically, we evaluate the performance of GPT-3.5 Turbo, Qwen-2.5B, and Aya-Expanse on the Belebele benchmark. We conduct a series of correlation and regression analyses to examine: (1) the influence of English similarity on transfer performance, (2) whether resource availability acts as a spurious or explanatory factor, and (3) which typological features most significantly predict multilingual model performance. Our findings offer deeper insight into the factors that drive cross-lingual generalization, with implications for improving model equity across languages.

Paper Type: Short

Research Area: Multilingualism and Cross-Lingual NLP

Research Area Keywords: typology, analysis, cross-lingual transfer, belebele, reading comprehension

Contribution Types: Model analysis & interpretability, Data analysis

Languages Studied: Afar, Amharic, Arabic, Assamese, Azerbaijani, Bambara, Bashkir, Basque, Belarusian, Bengali, Bhojpuri, Bosnian, Bulgarian, Burmese, Catalan, Cebuano, Chichewa, Chinese (Simplified), Chinese (Traditional), Croatian, Czech, Danish, Dutch, English, Estonian, Ewe, Filipino, Finnish, French, Galician, Georgian, German, Greek, Gujarati, Hausa, Hebrew, Hindi, Hungarian, Icelandic, Igbo, Ilocano, Indonesian, Irish, Italian, Japanese, Javanese, Kannada, Kazakh, Khmer, Kinyarwanda, Korean, Kurdish (Kurmanji), Kyrgyz, Lao, Latvian, Lithuanian, Luganda, Luxembourgish, Macedonian, Maithili, Malagasy, Malay, Malayalam, Maltese, Marathi, Meiteilon (Manipuri), Mongolian, Nepali, Northern Sotho, Norwegian, Nyanja, Odia, Oromo, Pashto, Persian, Polish, Portuguese, Punjabi, Quechua, Romanian, Russian, Samoan, Scots Gaelic, Serbian, Sesotho, Shona, Sindhi, Sinhala, Slovak, Slovenian, Somali, Spanish, Sundanese, Swahili, Swedish, Tagalog, Tajik, Tamil, Tatar, Telugu, Thai, Tigrinya, Turkish, Turkmen, Ukrainian, Urdu, Uzbek, Vietnamese, Welsh, Western Frisian, Xhosa, Yiddish, Yoruba, Zulu

Reassignment Request Area Chair: This is not a resubmission

Reassignment Request Reviewers: This is not a resubmission

A1 Limitations Section: This paper has a limitations section.

A2 Potential Risks: N/A

B Use Or Create Scientific Artifacts: Yes

B1 Cite Creators Of Artifacts: Yes

B1 Elaboration: Section 3

B2 Discuss The License For Artifacts: No

B2 Elaboration: All resources used are under a Public use license

B3 Artifact Use Consistent With Intended Use: Yes

B3 Elaboration: Section 3

B4 Data Contains Personally Identifying Info Or Offensive Content: N/A

B5 Documentation Of Artifacts: Yes

B5 Elaboration: Section 3

B6 Statistics For Data: No

B6 Elaboration: We referenced the original dataset paper. We did not have enough space in a short paper to included the statistics.

C Computational Experiments: Yes

C1 Model Size And Budget: No

C1 Elaboration: We used a standard library for inference-only so the computational budget and requirements are minimal.

C2 Experimental Setup And Hyperparameters: N/A

C2 Elaboration: Inference-only

C3 Descriptive Statistics: Yes

C3 Elaboration: Section 4

C4 Parameters For Packages: Yes

C4 Elaboration: Section 3

D Human Subjects Including Annotators: No

D1 Instructions Given To Participants: N/A

D2 Recruitment And Payment: N/A

D3 Data Consent: N/A

D4 Ethics Review Board Approval: N/A

D5 Characteristics Of Annotators: N/A

E Ai Assistants In Research Or Writing: No

E1 Information About Use Of Ai Assistants: N/A

Author Submission Checklist: yes

Submission Number: 328