notesum.ai

Published at October 21

Do Large Language Models Have an English Accent? Evaluating and Improving the Naturalness of Multilingual LLMs

cs.CL

cs.AI

I.2.7

Released Date: October 21, 2024

Authors: Yanzhu Guo¹, Simone Conia², Zelin Zhou³, Min Li³, Saloni Potdar³, Henry Xiao³

Aff.: ¹Inria Paris; ²Sapienza University of Rome; ³Apple

Arxiv: https://arxiv.org/abs/2410.15956v2

POS Pattern

n

-grams from Llama-3.1

Frequency

Explanation

English

Chinese

Native

English

Native

Chinese

Llama-3.1

Chinese

(ADJ, CCONJ, ADJ)

(blue, and, white)

(incoming, and, outgoing)

(eastern, and, north)

(真诚, 和, 直接)

(正常, 和, 合法)

(血腥, 和, 黑暗)

English requires conjunctions when listing

adjectives, while Chinese often omits them.

(PRON, AUX, VERB)

(They, were, married)

(he, was, named)

(it, was, built)

(他, 被, 释放)

(她, 被, 称为)

(他, 被, 任命)

175

English uses passive constructions with

auxiliary verbs more frequently than Chinese.

(ADP, DET, NOUN, ADP)

(During, the, boom, of)

(at, the, end, of)

(At, the, height, of)

(在, 多部, 电影, 中)

(在, 此, 职位, 上)

(在, 此次, 比赛, 中)

300

English depends heavily on prepositions and

determiners to structure sentences, while Chinese

tends to simplify these relationships.