Similar Language Pairs Research Articles

Dual-perspective fusion for word translation enhancement

Most Bilingual Lexicon Induction (BLI) methods retrieve word translation pairs by finding the closest target word for a given source word based on cross-lingual word embeddings (WEs). However, we find that solely retrieving translation from the source-to-target perspective leads to some false positive translation pairs, which significantly harm the precision of BLI. To address this problem, we propose a novel and effective method to improve translation pair retrieval in cross-lingual WEs. Specifically, we apply a fusion of both source-side and target-side perspectives throughout the retrieval process to alleviate false positive word pairings that emanate from a single perspective. Moreover, in translation scenarios using Large Language Models (LLMs), we propose fusing the LLMs perspective with the BLI model perspective to enhance LLM’s translation capability. On benchmark datasets of BLI, our proposed method achieves competitive performance compared to existing state-of-the-art (SOTA) methods. It demonstrates effectiveness and robustness across six experimental languages, including similar language pairs and distant language pairs, under both supervised and unsupervised settings.

Enhancing bilingual lexicon induction via harnessing polysemous words

Bilingual Lexicon Induction (BLI) aims to induce word translation pairs across different languages. We find there is a large proportion of polysemous words in the dataset used by BLI, which impacts the quality of the induced translation pairs. In particular, the mainstream BLI models map two monolingual word embeddings (WEs) into a shared space and regard the closest cross-lingual words as translation pairs. However, when a word has multiple translations, it becomes challenging to ensure that the polysemous word is the closest neighbor to all its translations. Based on the analysis of polysemy, we propose a simple yet effective method by harnessing polysemous words to enhance BLI. We first detect polysemous words in the seed lexicons and filter their multiple translations by comparing their semantic similarities. Therefore, a refined seed lexicon mitigates the confusion caused by translations with divergent semantics for the BLI model. Moreover, in scenarios where only single translations are available in the seed lexicon, we propose to employ a Large Language Model (LLM) to acquire monolingual synonyms, thereby refining the structure of the seed lexicon more effectively. On benchmark BLI datasets, our method demonstrates superior performance compared to five state-of-the-art (SOTA) baseline systems across almost all experimental language pairs, encompassing both sides of three similar language pairs and three distant language pairs.

Enhancing Bilingual Lexicon Induction via Bi-directional Translation Pair Retrieving

Most Bilingual Lexicon Induction (BLI) methods retrieve word translation pairs by finding the closest target word for a given source word based on cross-lingual word embeddings (WEs). However, we find that solely retrieving translation from the source-to-target perspective leads to some false positive translation pairs, which significantly harm the precision of BLI. To address this problem, we propose a novel and effective method to improve translation pair retrieval in cross-lingual WEs. Specifically, we consider both source-side and target-side perspectives throughout the retrieval process to alleviate false positive word pairings that emanate from a single perspective. On a benchmark dataset of BLI, our proposed method achieves competitive performance compared to existing state-of-the-art (SOTA) methods. It demonstrates effectiveness and robustness across six experimental languages, including similar language pairs and distant language pairs, under both supervised and unsupervised settings.

Unsupervised Neural Machine Translation for Similar and Distant Language Pairs

Unsupervised neural machine translation (UNMT) has achieved remarkable results for several language pairs, such as French–English and German–English. Most previous studies have focused on modeling UNMT systems; few studies have investigated the effect of UNMT on specific languages. In this article, we first empirically investigate UNMT for four diverse language pairs (French/German/Chinese/Japanese–English). We confirm that the performance of UNMT in translation tasks for similar language pairs (French/German–English) is dramatically better than for distant language pairs (Chinese/Japanese–English). We empirically show that the lack of shared words and different word orderings are the main reasons that lead UNMT to underperform in Chinese/Japanese–English. Based on these findings, we propose several methods, including artificial shared words and pre-ordering, to improve the performance of UNMT for distant language pairs. Moreover, we propose a simple general method to improve translation performance for all these four language pairs. The existing UNMT model can generate a translation of a reasonable quality after a few training epochs owing to a denoising mechanism and shared latent representations. However, learning shared latent representations restricts the performance of translation in both directions, particularly for distant language pairs, while denoising dramatically delays convergence by continuously modifying the training data. To avoid these problems, we propose a simple, yet effective and efficient, approach that (like UNMT) relies solely on monolingual corpora: pseudo-data-based unsupervised neural machine translation. Experimental results for these four language pairs show that our proposed methods significantly outperform UNMT baselines.

On determining matrix language of code-switching between southern Min and Mandarin

This paper asks whether there is a universally applicable code-switching model. To answer this question, Myers-Scotton’s (2002) Matrix Language Frame model was tested with Southern Min/Mandarin code-switched data. The results showed that the two criteria (i.e. the morpheme order principle and the system morpheme principle) proposed to identify the matrix language in this model were found inapplicable. This is because Southern Min and Mandarin share most of their morphosyntactic structure and have a limited amount of inflectional morphology. To solve this theoretical problem, one supporting theory of the matrix language frame model (i.e. the uniform structure principle) and other possible criteria were also tested, but they were not successful. With the re-introduction of an additional criterion (i.e. the morpheme counting principle) to identify the matrix language, this study proposes a revised version of the matrix language frame model which was applied to re-analyze the Southern Min/Mandarin code-switched data. The results show that the matrix languages of 92.63% of the Southern Min/Mandarin bilingual clauses (340 in total) were unambiguously identified. This suggests that unlike the original model, the revised version of the matrix language frame model is able to cover a wider range of code-switched data (i.e. bothtypologically similar and different language pairs). With the empirical success of the original matrix language frame model reported in the previous literature and the success of additional principle proposed to particularly analyse code-switched data of a typologically similar language pair (Southern Min and Mandarin), this study argues that the revised version of the matrix language frame model is universally applicable. 提要: 本研究之目的在于寻找一个能适用于各种不同语言特性之语码转换 (code-switching)理论。为达此目的, 我们利用普通话及闽南话之语码 转换语料, 来测试 Myers-Scotton (2002)所提出的「主体语结构模式理 论 (The Matrix Language Frame Model)」。但测试结果发现, 该理论中 用来确认主体语的两项指标: 「词序原则 (the morpheme order principle)」及「系统词素原则 (the system morpheme principle)」皆无法 适用于普通话及闽南话的转换语料。由于此两种语言在句法结构上有 极高的相似性, 且在屈折型态学 (inflectional morphology)上的特性亦不 明显, 因而凸显出该理论无法处理句法、型态相似的语码转换语料的 缺陷。本研究尝试采用结构一致性原则 (the uniform structure principle) 及其他的替代方案, 但都无法有效确认普通话及闽南话转换语料的主 体语。因此, 我们在该理论的原有基础上稍做修正, 加入另一个确认 主体语的指标: 「词素计算原则 (the morpheme counting principle)」来 分析相同语料。分析结果显示: 在总数 340 个包含普通话及闽南话的 双语子句中, 有 92. 63%的子句之主体语能被清楚确认。由于现有的文 献指出, 原有的「主体语结构模式理论」已成功通过许多在句法、型 态上差异性大的语码转换语料的检测, 本研究所提出的修正版本更强 化了该理论, 使其亦能涵盖语言结构具高度相似性的语料。因此, 我 们认为本研究所提出「主体语结构模式理论」修正版本, 可适用于各 种不同语言所构成的语码转换语料。

Semantic Access in Number Word Translation

The revised hierarchical model of bilingualism (e.g., Kroll & Stewart, 1994) assumes that second language (L2) words primarily access semantics through their first language (L1) translation equivalents. Consequently, backward translation from L2 to L1 should not imply semantic access but occurs through lexical wordform associations. However, recent research with Dutch-French bilinguals showed that both backward and forward translation of number words yields a semantic number magnitude effect (Duyck & Brysbaert, 2004), providing evidence for strong form-to-meaning mappings of L2 number words. In two number-word translation experiments with Dutch-English-German trilinguals, the present study investigated whether semantic access in L1-L2 and L1-L3 number-word translation depends on lexical similarity of the languages involved. We found that backward translation from these more similar language pairs to L1 still yields a semantic magnitude effect, whereas forward translation does not, in contrast with the Dutch-French results of Duyck and Brysbaert (2004). We argue against a dual route model of word translation and suggest that the degree of semantic activation in translation depends on lexical form overlap between translation equivalents.