Memory-efficient finetuning of large language models (LLMs) has recently attracted huge attention with the increasing size of LLMs, primarily due to the constraints posed by GPU memory limitations and the effectiveness of these methods compared to full finetuning. Despite the advancements, current strategies for memory-efficient finetuning, such as QLoRA, exhibit inconsistent performance across diverse bit-width quantizations and multifaceted tasks. This inconsistency largely stems from the detrimental impact of the quantization process on preserved knowledge, leading to catastrophic forgetting and undermining the utilization of pretrained models for finetuning purposes. In this work, we introduce a novel quantization framework named ApiQ, designed to restore the lost information from quantization by concurrently initializing the LoRA components and quantizing the weights of LLMs. This approach ensures the maintenance of the original LLM’s activation precision while mitigating the error propagation from shallower into deeper layers. Through comprehensive evaluations conducted on a spectrum of language tasks with various LLMs, ApiQ demonstrably minimizes activation error during quantization. Consequently, it consistently achieves superior finetuning results across various bit-widths. Notably, one can even finetune a 2-bit Llama-2-70b with ApiQ on a single NVIDIA A100-80GB GPU without any memory-saving techniques, and achieve promising results.

This paper introduces two multilingual systems, IKUN and IKUN-C, developed for the general machine translation task in WMT24. IKUN and IKUN-C represent an open system and a constrained system, respectively, built on Llama-3-8b and Mistral-7B-v0.3. Both systems are designed to handle all 11 language directions using a single model. According to automatic evaluation metrics, IKUN-C achieved 6 first-place and 3 second-place finishes among all constrained systems, while IKUN secured 1 first-place and 2 second-place finishes across both open and constrained systems. These encouraging results suggest that large language models (LLMs) are nearing the level of proficiency required for effective multilingual machine translation. The systems are based on a two-stage approach: first, continuous pre-training on monolingual data in 10 languages, followed by fine-tuning on high-quality parallel data for 11 language directions. The primary difference between IKUN and IKUN-C lies in their monolingual pre-training strategy. IKUN-C is pre-trained using constrained monolingual data, whereas IKUN leverages monolingual data from the OSCAR dataset. In the second phase, both systems are fine-tuned on parallel data sourced from NTREX, Flores, and WMT16-23 for all 11 language pairs.

The integration of language models for neural machine translation has been extensively studied in the past. It has been shown that an external language model, trained on additional target-side monolingual data, can help improve translation quality. However, there has always been the assumption that the translation model also learns an implicit target-side language model during training, which interferes with the external language model at decoding time. Recently, some works on automatic speech recognition have demonstrated that, if the implicit language model is neutralized in decoding, further improvements can be gained when integrating an external language model. In this work, we transfer this concept to the task of machine translation and compare with the most prominent way of including additional monolingual data - namely back-translation. We find that accounting for the implicit language model significantly boosts the performance of language model fusion, although this approach is still outperformed by back-translation.

Compared to sentence-level systems, document-level neural machine translation (NMT) models produce a more consistent output across a document and are able to better resolve ambiguities within the input. There are many works on document-level NMT, mostly focusing on modifying the model architecture or training strategy to better accommodate the additional context-input. On the other hand, in most works, the question on how to perform search with the trained model is scarcely discussed, sometimes not mentioned at all. In this work, we aim to answer the question how to best utilize a context-aware translation model in decoding. We start with the most popular document-level NMT approach and compare different decoding schemes, some from the literature and others proposed by us. In the comparison, we are using both, standard automatic metrics, as well as specific linguistic phenomena on three standard document-level translation benchmarks. We find that most commonly used decoding strategies perform similar to each other and that higher quality context information has the potential to further improve the translation.

Despite the known limitations, most machine translation systems today still operate on the sentence-level. One reason for this is, that most parallel training data is only sentence-level aligned, without document-level meta information available. In this work, we set out to build context-aware translation systems utilizing document-level monolingual data instead. This can be achieved by combining any existing sentence-level translation model with a document-level language model. We improve existing approaches by leveraging recent advancements in model combination. Additionally, we propose novel weighting techniques that make the system combination more flexible and significantly reduce computational overhead. In a comprehensive evaluation on four diverse translation tasks, we show that our extensions improve document-targeted scores significantly and are also computationally more efficient. However, we also find that in most scenarios, back-translation gives even better results, at the cost of having to re-train the translation system. Finally, we explore language model fusion in the light of recent advancements in large language models. Our findings suggest that there might be strong potential in utilizing large language models via model combination.

Document-level context for neural machine translation (NMT) is crucial to improve the translation consistency and cohesion, the translation of ambiguous inputs, as well as several other linguistic phenomena. Many works have been published on the topic of document-level NMT, but most restrict the system to only local context, typically including just the one or two preceding sentences as additional information. This might be enough to resolve some ambiguous inputs, but it is probably not sufficient to capture some document-level information like the topic or style of a conversation. When increasing the context size beyond just the local context, there are two challenges: (i) the memory usage increases exponentially (ii) the translation performance starts to degrade. We argue that the widely-used attention mechanism is responsible for both issues. Therefore, we propose a constrained attention variant that focuses the attention on the most relevant parts of the sequence, while simultaneously reducing the memory consumption. For evaluation, we utilize targeted test sets in combination with novel evaluation techniques to analyze the translations in regards to specific discourse-related phenomena. We find that our approach is a good compromise between sentence-level NMT vs attending to the full context, especially in low resource scenarios.

Currently, in speech translation, the straightforward approach - cascading a recognition system with a translation system - delivers state-of-the-art results.However, fundamental challenges such as error propagation from the automatic speech recognition system still remain.To mitigate these problems, recently, people turn their attention to direct data and propose various joint training methods.In this work, we seek to answer the question of whether joint training really helps cascaded speech translation.We review recent papers on the topic and also investigate a joint training criterion by marginalizing the transcription posterior probabilities.Our findings show that a strong cascaded baseline can diminish any improvements obtained using joint training, and we suggest alternatives to joint training.We hope this work can serve as a refresher of the current speech translation landscape, and motivate research in finding more efficient and creative ways to utilize the direct data for speech translation.

The filtering and/or selection of training data is one of the core aspects to be considered when building a strong machine translation system. In their influential work, Khayrallah and Koehn (2018) investigated the impact of different types of noise on the performance of machine translation systems. In the same year the WMT introduced a shared task on parallel corpus filtering, which went on to be repeated in the following years, and resulted in many different filtering approaches being proposed. In this work we aim to combine the recent achievements in data filtering with the original analysis of Khayrallah and Koehn (2018) and investigate whether state-of-the-art filtering systems are capable of removing all the suggested noise types. We observe that most of these types of noise can be detected with an accuracy of over 90% by modern filtering systems when operating in a well studied high resource setting. However, we also find that when confronted with more refined noise categories or when working with a less common language pair, the performance of the filtering systems is far from optimal, showing that there is still room for improvement in this area of research.

Checkpoint averaging is a simple and effective method to boost the performance of converged neural machine translation models. The calculation is cheap to perform and the fact that the translation improvement almost comes for free, makes it widely adopted in neural machine translation research. Despite the popularity, the method itself simply takes the mean of the model parameters from several checkpoints, the selection of which is mostly based on empirical recipes without many justifications. In this work, we revisit the concept of checkpoint averaging and consider several extensions. Specifically, we experiment with ideas such as using different checkpoint selection strategies, calculating weighted average instead of simple mean, making use of gradient information and fine-tuning the interpolation weights on development data. Our results confirm the necessity of applying checkpoint averaging for optimal performance, but also suggest that the landscape between the converged checkpoints is rather flat and not much further improvement compared to simple averaging is to be obtained.

Encoder-decoder architecture is widely adopted for sequence-to-sequence modeling tasks. For machine translation, despite the evolution from long short-term memory networks to Transformer networks, plus the introduction and development of attention mechanism, encoder-decoder is still the de facto neural network architecture for state-of-the-art models. While the motivation for decoding information from some hidden space is straightforward, the strict separation of the encoding and decoding steps into an encoder and a decoder in the model architecture is not necessarily a must. Compared to the task of autoregressive language modeling in the target language, machine translation simply has an additional source sentence as context. Given the fact that neural language models nowadays can already handle rather long contexts in the target language, it is natural to ask whether simply concatenating the source and target sentences and training a language model to do translation would work. In this work, we investigate the aforementioned concept for machine translation. Specifically, we experiment with bilingual translation, translation with additional target monolingual data, and multilingual translation. In all cases, this alternative approach performs on par with the baseline encoder-decoder Transformer, suggesting that an encoder-decoder architecture might be redundant for neural machine translation.

After its introduction the Transformer architecture quickly became the gold standard for the task of neural machine translation. A major advantage of the Transformer compared to previous architectures is the faster training speed achieved by complete parallelization across timesteps due to the use of attention over recurrent layers. However, this also leads to one of the biggest problems of the Transformer, namely the quadratic time and memory complexity with respect to the input length. In this work we adapt the locality-sensitive hashing approach of Kitaev et al. (2020) to self-attention in the Transformer, we extended it to cross-attention and apply this memory efficient framework to sentence- and document-level machine translation. Our experiments show that the LSH attention scheme for sentence-level comes at the cost of slightly reduced translation quality. For document-level NMT we are able to include much bigger context sizes than what is possible with the baseline Transformer. However, more context does neither improve translation quality nor improve scores on targeted test suites.

Data filtering for machine translation (MT) describes the task of selecting a subset of a given, possibly noisy corpus with the aim to maximize the performance of an MT system trained on this selected data. Over the years, many different filtering approaches have been proposed. However, varying task definitions and data conditions make it difficult to draw a meaningful comparison. In the present work, we aim for a more systematic approach to the task at hand. First, we analyze the performance of language identification, a tool commonly used for data filtering in the MT community and identify specific weaknesses. Based on our findings, we then propose several novel methods for data filtering, based on cross-lingual word embeddings. We compare our approaches to one of the winning methods from the WMT 2018 shared task on parallel corpus filtering on three real-life, high resource MT tasks. We find that said method, which was performing very strong in the WMT shared task, does not perform well within our more realistic task conditions. While we find that our approaches come out at the top on all three tasks, different variants perform best on different tasks. Further experiments on the WMT 2020 shared task for parallel corpus filtering show that our methods achieve comparable results to the strongest submissions of this campaign.

In this work, we conduct a comprehensive investigation on one of the centerpieces of modern machine translation systems: the encoder-decoder attention mechanism. Motivated by the concept of first-order alignments, we extend the (cross-)attention mechanism by a recurrent connection, allowing direct access to previous attention/alignment decisions. We propose several ways to include such a recurrency into the attention mechanism. Verifying their performance across different translation tasks we conclude that these extensions and dependencies are not beneficial for the translation performance of the Transformer architecture.

In order to combat overfitting and in pursuit of better generalization, label smoothing is widely applied in modern neural machine translation systems. The core idea is to penalize over-confident outputs and regularize the model so that its outputs do not diverge too much from some prior distribution. While training perplexity generally gets worse, label smoothing is found to consistently improve test performance. In this work, we aim to better understand label smoothing in the context of neural machine translation. Theoretically, we derive and explain exactly what label smoothing is optimizing for. Practically, we conduct extensive experiments by varying which tokens to smooth, tuning the probability mass to be deducted from the true targets and considering different prior distributions. We show that label smoothing is theoretically well-motivated, and by carefully choosing hyperparameters, the practical performance of strong neural machine translation systems can be further improved.

Context-aware neural machine translation (NMT) is a promising direction to improve the translation quality by making use of the additional context, e.g., document-level translation, or having meta-information. Although there exist various architectures and analyses, the effectiveness of different context-aware NMT models is not well explored yet. This paper analyzes the performance of document-level NMT models on four diverse domains with a varied amount of parallel document-level bilingual data. We conduct a comprehensive set of experiments to investigate the impact of document-level NMT. We find that there is no single best approach to document-level NMT, but rather that different architectures come out on top on different tasks. Looking at task-specific problems, such as pronoun resolution or headline translation, we find improvements in the context-aware systems, even in cases where the corpus-level metrics like BLEU show no significant improvement. We also show that document-level back-translation significantly helps to compensate for the lack of document-level bi-texts.

AppTek and RWTH Aachen University team together to participate in the offline and simultaneous speech translation tracks of IWSLT 2020. For the offline task, we create both cascaded and end-to-end speech translation systems, paying attention to careful data selection and weighting. In the cascaded approach, we combine high-quality hybrid automatic speech recognition (ASR) with the Transformer-based neural machine translation (NMT). Our end-to-end direct speech translation systems benefit from pretraining of adapted encoder and decoder components, as well as synthetic data and fine-tuning and thus are able to compete with cascaded systems in terms of MT quality. For simultaneous translation, we utilize a novel architecture that makes dynamic decisions, learned from parallel data, to determine when to continue feeding on input or generate output words. Experiments with speech and text input show that even at low latency this architecture leads to superior translation results.

Prominently used in support vector machines and logistic re-gressions, kernel functions (kernels) can implicitly map data points into high dimensional spaces and make it easier to learn complex decision boundaries. In this work, by replacing the inner product function in the softmax layer, we explore the use of kernels for contextual word classification. In order to compare the individual kernels, experiments are conducted on standard language modeling and machine translation tasks. We observe a wide range of performances across different kernel settings. Extending the results, we look at the gradient properties, investigate various mixture strategies and examine the disambiguation abilities.

This paper describes the neural machine translation systems developed at the RWTH Aachen University for the German-English, Chinese-English and Kazakh-English news translation tasks of the Fourth Conference on Machine Translation (WMT19). For all tasks, the final submitted system is based on the Transformer architecture. We focus on improving data filtering and fine-tuning as well as systematically evaluating interesting approaches like unigram language model segmentation and transfer learning. For the De-En task, none of the tested methods gave a significant improvement over last years winning system and we end up with the same performance, resulting in 39.6% BLEU on newstest2019. In the Zh-En task, we show 1.3% BLEU improvement over our last year’s submission, which we mostly attribute to the splitting of long sentences during translation. We further report results on the Kazakh-English task where we gain improvements of 11.1% BLEU over our baseline system. On the same task we present a recent transfer learning approach, which uses half of the free parameters of our submission system and performs on par with it.

Embedding and projection matrices are commonly used in neural language models (NLM) as well as in other sequence processing networks that operate on large vocabularies. We examine such matrices in fine-tuned language models and observe that a NLM learns word vectors whose norms are related to the word frequencies. We show that by initializing the weight norms with scaled log word counts, together with other techniques, lower perplexities can be obtained in early epochs of training. We also introduce a weight norm regularization loss term, whose hyperparameters are tuned via a grid search. With this method, we are able to significantly improve perplexities on two word-level language modeling tasks (without dynamic evaluation): from 54.44 to 53.16 on Penn Treebank (PTB) and from 61.45 to 60.13 on WikiText-2 (WT2).