Yun He, Wenzhe Li, Hejia Zhang, Songlin Li, Karishma Mandyam, Sopan Khosla, Yuanhao Xiong, Nanshu Wang, Xiaoliang Peng, Beibin Li, Shengjie Bi, Shishir G Patil, Qi Qi, Shengyu Feng, Julian Katz-Samuels, Richard Yuanzhe Pang, Sujan Kumar Gonugondla, Hunter Lang, Yue Yu, Yundi Qian, Maryam Fazel-Zarandi, Licheng Yu, Amine Benhalloum, Hany Hassan Awadalla, Manaal Faruqui. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Yiyao Yu, Yuxiang Zhang, Dongdong Zhang, Xiao Liang, Hengyuan Zhang, Xingxing Zhang, Mahmoud Khademi, Hany Hassan Awadalla, Junjie Wang, Yujiu Yang, Furu Wei. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Generative Large Language Models (LLMs) have achieved remarkable advancements in various NLP tasks. However, these advances have not been reflected in the translation task, especially those with moderate model sizes (i.e., 7B or 13B parameters), which still lag behind conventional supervised encoder-decoder translation models. Previous studies have attempted to improve the translation capabilities of these LLMs, but their gains have been limited. In this study, we propose a novel fine-tuning approach for LLMs that is specifically designed for the translation task, eliminating the need for the abundant parallel data that traditional translation models usually depend on. Our approach consists of two fine-tuning stages: initial fine-tuning on monolingual data followed by subsequent fine-tuning on a small set of high-quality parallel data. We introduce the LLM developed through this strategy as **A**dvanced **L**anguage **M**odel-based tr**A**nslator (**ALMA**). Based on LLaMA-2 as our underlying model, our results show that the model can achieve an average improvement of more than 12 BLEU and 12 COMET over its zero-shot performance across 10 translation directions from the WMT'21 (2 directions) and WMT'22 (8 directions) test datasets. The performance is significantly better than all prior work and even superior to the NLLB-54B model \citep{nllb} and GPT-3.5-text-davinci-003, with only 7B or 13B parameters. This method establishes the foundation for a novel training paradigm in machine translation.

Pengcheng He, Baolin Peng, Song Wang, Yang Liu, Ruochen Xu, Hany Hassan, Yu Shi, Chenguang Zhu, Wayne Xiong, Michael Zeng, Jianfeng Gao, Xuedong Huang. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Sparsely activated transformers, such as Mixture of Experts (MoE), have received great interest due to their outrageous scaling capability which enables dramatical increases in model size without significant increases in computational cost. To achieve this, MoE models replace the feedforward sub-layer with Mixture-of-Experts sub-layer in transformers and use a gating network to route each token to its assigned experts. Since the common practice for efficient training of such models requires distributing experts and tokens across different machines, this routing strategy often incurs huge cross-machine communication cost because tokens and their assigned experts likely reside in different machines. In this paper, we propose Gating Dropout, which allows tokens to ignore the gating network and stay at their local machines, thus reducing the cross-machine communication. Similar to traditional dropout, we also show that Gating Dropout has a regularization effect during training, resulting in improved generalization performance. We validate the effectiveness of Gating Dropout on multilingual machine translation tasks. Our results demonstrate that Gating Dropout improves a state-of-the-art MoE model with faster wall-clock time convergence rates and better BLEU scores for a variety of model sizes and datasets.

Multilingual Neural Machine Translation (MNMT) enables one system to translate sentences from multiple source languages to multiple target languages, greatly reducing deployment costs compared with conventional bilingual systems.The MNMT training benefit, however, is often limited to many-to-one directions.The model suffers from poor performance in one-to-many and many-to-many with zero-shot setup.To address this issue, this paper discusses how to practically build MNMT systems that serve arbitrary X-Y translation directions while leveraging multilinguality with a two-stage training strategy of pretraining and finetuning.Experimenting with the WMT'21 multilingual translation task, we demonstrate that our systems outperform the conventional baselines of direct bilingual models and pivot translation models for most directions, averagely giving +6.0 and +4.1 BLEU, without the need for architecture change or extra data collection.Moreover, we also examine our proposed approach in an extremely large-scale data setting to accommodate practical deployment scenarios.

Multilingual Neural Machine Translation (NMT) enables one model to serve all translation directions, including ones that are unseen during training, i.e. zero-shot translation.Despite being theoretically attractive, current models often produce low quality translations -commonly failing to even produce outputs in the right target language.In this work, we observe that off-target translation is dominant even in strong multilingual systems, trained on massive multilingual corpora.To address this issue, we propose a joint approach to regularize NMT models at both representation-level and gradient-level.At the representation level, we leverage an auxiliary target language prediction task to regularize decoder outputs to retain information about the target language.At the gradient level, we leverage a small amount of direct data (in thousands of sentence pairs) to regularize model gradients.Our results demonstrate that our approach is highly effective in both reducing off-target translation occurrences and improving zero-shot translation performance by +5.59 and +10.38 BLEU on WMT and OPUS datasets respectively.Moreover, experiments show that our method also works well when the small amount of direct data is not available.1

While monolingual data has been shown to be useful in improving bilingual neural machine translation (NMT), effectively and efficiently leveraging monolingual data for Multilingual NMT (MNMT) systems is a less explored area.In this work, we propose a multi-task learning (MTL) framework that jointly trains the model with the translation task on bitext data and two denoising tasks on the monolingual data.We conduct extensive empirical studies on MNMT systems with 10 language pairs from WMT datasets.We show that the proposed approach can effectively improve the translation quality for both high-resource and low-resource languages with large margin, achieving significantly better results than the individual bilingual models.We also demonstrate the efficacy of the proposed approach in the zero-shot setup for language pairs without bitext training data.Furthermore, we show the effectiveness of MTL over pre-training approaches for both NMT and cross-lingual transfer learning NLU tasks; the proposed approach outperforms massive scale models trained on single task.

Modern NLP applications have enjoyed a great boost utilizing neural networks models.Such deep neural models, however, are not applicable to most human languages due to the lack of annotated training data for various NLP tasks.Cross-lingual transfer learning (CLTL) is a viable method for building NLP models for a low-resource target language by leveraging labeled data from other (source) languages.In this work, we focus on the multilingual transfer setting where training data in multiple source languages is leveraged to further boost target language performance.Unlike most existing methods that rely only on language-invariant features for CLTL, our approach coherently utilizes both languageinvariant and language-specific features at instance level.Our model leverages adversarial networks to learn language-invariant features, and mixture-of-experts models to dynamically exploit the similarity between the target language and each individual source language 1 .This enables our model to learn effectively what to share between various languages in the multilingual setup.Moreover, when coupled with unsupervised multilingual embeddings, our model can operate in a zero-resource setting where neither target language training data nor cross-lingual resources are available.Our model achieves significant performance gains over prior art, as shown in an extensive set of experiments over multiple text classification and sequence tagging tasks including a large-scale industry dataset.

Jiatao Gu, Hany Hassan, Jacob Devlin, Victor O.K. Li. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). 2018.

Translation models often fail to generate good translations for infrequent words or phrases. Previous work attacked this problem by inducing new translation rules from monolingual data with a semi-supervised algorithm. However, this approach does not scale very well since it is very computationally expensive to generate new translation rules for only a few thousand sentences. We propose a much faster and simpler method that directly hallucinates translation rules for infrequent phrases based on phrases with similar continuous representations for which a translation is known. To speed up the retrieval of similar phrases, we investigate approximated nearest neighbor search with redundant bit vectors which we find to be three times faster and significantly more accurate than locality sensitive hashing. Our approach of learning new translation rules improves a phrase-based baseline by up to 1.6 BLEU on Arabic-English translation, it is three-orders of magnitudes faster than existing semi-supervised methods and 0.5 BLEU more accurate.

Statistical phrase-based translation learns translation rules from bilingual corpora, and has traditionally only used monolingual evidence to construct features that rescore existing translation candidates.In this work, we present a semi-supervised graph-based approach for generating new translation rules that leverages bilingual and monolingual data.The proposed technique first constructs phrase graphs using both source and target language monolingual corpora.Next, graph propagation identifies translations of phrases that were not observed in the bilingual corpus, assuming that similar phrases have similar translations.We report results on a large Arabic-English system and a medium-sized Urdu-English system.Our proposed approach significantly improves the performance of competitive phrasebased systems, leading to consistent improvements between 1 and 4 BLEU points on standard evaluation sets.Source!Target! el gato! los gatos!un gato! cat! the cat! the cats! a cat! Target!Prob.! the cat! 0.7! cat! 0.15! …! …! felino!canino!el perro!Target!Prob.! canine!0.6!dog! 0.3!…! …! Target!Prob.! the cats!0.8! cats! 0.1!…! …

In this paper we focus on the effect of on-line speech segmentation and disfluency removal methods on conversational speech translation. In a real-time conversational speech to speech translation system, on-line segmentation of speech is required to avoid latency beyond few seconds. While sentential unit segmentation and disfluency removal have been heavily studied mainly for off-line speech processing, to the best of our knowledge, the combined effect of these tasks on conversational speech translation has not been investigated. Furthermore, optimization of performance given maximum allowable system latency to enable a conversation is a newer problem for these tasks. We show that the conventional assumption of doing segmentation followed by disfluency removal is not the best practice. We propose a new approach to do simple-disfluency removal followed by segmentation and then by complex-disfluency removal. The proposed approach shows a significant gain on translation performance of up to 3 Bleu points with only 6 second latency to look ahead, using state-ofthe art machine translation and speech recognition systems. Index Terms: speech translation, disfluency removal, segmentation, sentence units, speech processing

Syntactically-enriched language models (parsers) constitute a promising component in applications such as machine translation and speech-recognition. To maintain a useful level of accuracy, existing parsers are non-incremental and must span a combinatorially growing space of possible structures as every input word is processed. This prohibits their incorporation into standard linear-time decoders. In this paper, we present an incremental, linear-time dependency parser based on Combinatory Categorial Grammar (CCG) and classification techniques. We devise a deterministic transform of CCG-bank canonical derivations into incremental ones, and train our parser on this data. We discover that a cascaded, incremental version provides an appealing balance between efficiency and accuracy.

Until quite recently, extending phrase-based statistical machine translation (PBSMT) with syntactic knowledge caused system performance to deteriorate. The most recent successful enrichments of PBSMT with hierarchical structure either employ nonlinguistically motivated syntax for capturing hierarchical reordering phenomena, or extend the phrase translation table with redundantly ambiguous syntactic structures over phrase pairs. In this paper, we present an extended, harmonized account of our previous work which showed that incorporating linguistically motivated lexical syntactic descriptions, calledsupertags, can yield significantly better PBSMT systems at insignificant extra computational cost. We describe a novel PBSMT model that integrates supertags into the target language model and the target side of the translation model. Two kinds of supertags are employed: those from lexicalized tree-adjoining grammar and combinatory categorial grammar. Despite the differences between the two sets of supertags, they give similar improvements. In addition to integrating the Markov supertagging approach in PBSMT, we explore the utility of a new surface grammaticality measure based on combinatory operators. We perform various experiments on the Arabic-to-English NIST 2005 test set addressing the issues of sparseness, scalability, and the utility of system subcomponents. We show that even when the parallel training data grows very large, the supertagged system retains a relatively stable absolute performance advantage over the unadorned PBSMT system. Arguably, this hints at a performance gap that cannot be bridged by acquiring more phrase pairs. Our best result shows a relative improvement of 6.1% over a state-of-the-art PBSMT model, which compares favorably with the leading systems on the NIST 2005 task. We also demonstrate that the advantages of a supertag-based system carry over to German-English, where improvements of up to 8.9% relative to the baseline system are observed.

Phrase-based statistical machine translation (PBSMT) systems represent the dominant approach in MT today. However, unlike systems in other paradigms, it has proven difficult to date to incorporate syntactic knowledge in order to improve translation quality. This paper improves on recent research which uses 'syntactified' target language phrases, by incorporating supertags as constraints to better resolve parse tree fragments. In addition, we do not impose any sentence-length limit, and using a log-linear decoder, we outperform a state-of-the-art PBSMT system by over 1.3 BLEU points (or 3.51% relative) on the NIST 2003 Arabic-English test corpus.

We approximate Arabic's rich morphology by a model that a word consists of a sequence of morphemes in the pattern prefix*-stem-suffix* (* denotes zero or more occurrences of a morpheme). Our method is seeded by a small manually segmented Arabic corpus and uses it to bootstrap an unsupervised algorithm to build the Arabic word segmenter from a large unsegmented Arabic corpus. The algorithm uses a trigram language model to determine the most probable morpheme sequence for a given input. The language model is initially estimated from a small manually segmented corpus of about 110,000 words. To improve the segmentation accuracy, we use an unsupervised algorithm for automatically acquiring new stems from a 155 million word unsegmented corpus, and re-estimate the model parameters with the expanded vocabulary and training corpus. The resulting Arabic word segmentation system achieves around 97% exact match accuracy on a test corpus containing 28,449 word tokens. We believe this is a state-of-the-art performance and the algorithm can be used for many highly inflected languages provided that one can create a small manually segmented corpus of the language of interest.