MEXMA: Token-level objectives improve sentence representations

ACL ARR 2025 February Submission1122 Authors

Abstract: Cross-lingual sentence encoders (CLSE) create fixed-size sentence representations with aligned translations. Current pre-trained CLSE approaches use sentence-level objectives only. This can lead to loss of information, especially for tokens, which then degrades the sentence representation. We propose MEXMA, a novel approach that integrates both sentence-level and token-level objectives. The sentence representation in one language is used to predict masked tokens in another language, with both the sentence representation and $\textit{all tokens directly updating the encoder}$. We show that adding token-level objectives greatly improves the sentence representation quality across several tasks. Our approach outperforms current pre-trained cross-lingual sentence encoders on bitext mining as well as several downstream tasks. We also analyse the information encoded in our tokens, and how the sentence representation is built from them.

Paper Type: Long

Research Area: Semantics: Lexical and Sentence-Level

Research Area Keywords: contrastive learning,fine-tuning,mixed language,multilingualism,cross-lingual transfer,multilingual representations,multilingual pre-training,multilingual benchmarks,multilingual evaluation,phrase/sentence embedding,word/phrase alignment

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Publicly available software and/or pre-trained models

Languages Studied: acm_Arab, aeb_Arab, afr_Latn, amh_Ethi, ary_Arab, arz_Arab, asm_Beng, azb_Arab, azj_Latn, bel_Cyrl, ben_Beng, bos_Latn, bul_Cyrl, cat_Latn, ces_Latn, ckb_Arab, cym_Latn, dan_Latn, deu_Latn, ell_Grek, epo_Latn, est_Latn, eus_Latn, fin_Latn, fra_Latn, gla_Latn, gle_Latn, glg_Latn, guj_Gujr, hau_Latn, heb_Hebr, hin_Deva, hrv_Latn, hun_Latn, hye_Armn, ind_Latn, isl_Latn, ita_Latn, jav_Latn, jpn_Jpan, kan_Knda, kat_Geor, kaz_Cyrl, khm_Khmr, kir_Cyrl, kor_Hang, lao_Laoo, mal_Mlym, mar_Deva, mkd_Cyrl, mya_Mymr, nld_Latn, nno_Latn, nob_Latn, npi_Deva, pol_Latn, por_Latn, ron_Latn, rus_Cyrl, san_Deva, sin_Sinh, slk_Latn, slv_Latn, snd_Arab, som_Latn, spa_Latn, srp_Cyrl, sun_Latn, swe_Latn, swh_Latn, tam_Taml, tel_Telu, tha_Thai, tur_Latn, uig_Arab, ukr_Cyrl, urd_Arab, vie_Latn, xho_Latn, zho_Hant

Submission Number: 1122