Sharing Matters: Analysing Neurons Across Languages and Tasks in LLMs

Multilingual large language models (LLMs) have greatly increased the ceiling of performance on non-English tasks. However the mechanisms behind multilingualism in these LLMs are poorly understood. Of particular interest is the degree to which internal representations are shared between languages. Recent work on neuron analysis of LLMs has focused on the monolingual case, and the limited work on the multilingual case has not considered the interaction between tasks and linguistic representations. In our work, we investigate how neuron activation is shared across languages by categorizing neurons into four distinct groups according to their responses across different languages for a particular input: all-shared, partial-shared, specific, and non-activated. This categorization is combined with a study of neuron attribution, i.e. the importance of a neuron w.r.t an output. Our analysis reveals the following insights: (i) the linguistic sharing patterns are strongly affected by the type of task, but neuron behaviour changes across different inputs even for the same task; (ii) all-shared neurons play a key role in generating correct responses; (iii) boosting multilingual alignment by increasing all-shared neurons can enhance accuracy on multilingual tasks. The code is available at https://github.com/weixuan-wang123/multilingual-neurons.

Optimising Calls to Large Language Models with Uncertainty-Based Two-Tier Selection

Researchers and practitioners operating on a limited budget face the cost-performance trade-off dilemma. The challenging decision often centers on whether to use a large LLM with better performance or a smaller one with reduced costs. This has motivated recent research in the optimisation of LLM calls. Either a cascading strategy is used, where a smaller LLM or both are called sequentially, or a routing strategy is used, where only one model is ever called. Both scenarios are dependent on a decision criterion which is typically implemented by an extra neural model. In this work, we propose a simpler solution; we use only the uncertainty of the generations of the small LLM as the decision criterion. We compare our approach with both cascading and routing strategies using three different pairs of pre-trained small and large LLMs, on nine different tasks and against approaches that require an additional neural model. Our experiments reveal this simple solution optimally balances cost and performance, outperforming existing methods on 25 out of 27 experimental setups.

The Power of Question Translation Training in Multilingual Reasoning: Broadened Scope and Deepened Insights

Bridging the significant gap between large language model's English and non-English performance presents a great challenge. While some previous studies attempt to mitigate this gap with translated training data, the recently proposed question alignment approach leverages the model's English expertise to improve multilingual performance with minimum usage of expensive, error-prone translation. In this paper, we explore how broadly this method can be applied by examining its effects in reasoning with executable code and reasoning with common sense. We also explore how to apply this approach efficiently to extremely large language models using proxy-tuning. Experiment results on multilingual reasoning benchmarks mGSM, mSVAMP and xCSQA demonstrate that the question alignment approach can be used to boost multilingual performance across diverse reasoning scenarios, model families, and sizes. For instance, when applied to the LLaMA2 models, our method brings an average accuracy improvements of 12.2% on mGSM even with the 70B model. To understand the mechanism of its success, we analyze representation space, chain-of-thought and translation data scales, which reveals how question translation training strengthens language alignment within LLMs and shapes their working patterns.

No Train but Gain: Language Arithmetic for training-free Language Adapters enhancement

Modular deep learning is the state-of-the-art solution for lifting the curse of multilinguality, preventing the impact of negative interference and enabling cross-lingual performance in Multilingual Pre-trained Language Models. However, a trade-off of this approach is the reduction in positive transfer learning from closely related languages. In response, we introduce a novel method called language arithmetic, which enables training-free post-processing to address this limitation. Inspired by the task arithmetic framework, we apply learning via addition to the language adapters, transitioning the framework from a multi-task to a multilingual setup. The effectiveness of the proposed solution is demonstrated on three downstream tasks in a MAD-X-based set of cross-lingual schemes, acting as a post-processing procedure. Language arithmetic consistently improves the baselines with significant gains in the most challenging cases of zero-shot and low-resource applications. Our code and models are available at https://github.com/mklimasz/language-arithmetic .

Is Modularity Transferable? A Case Study through the Lens of Knowledge Distillation

The rise of Modular Deep Learning showcases its potential in various Natural Language Processing applications. Parameter-efficient fine-tuning (PEFT) modularity has been shown to work for various use cases, from domain adaptation to multilingual setups. However, all this work covers the case where the modular components are trained and deployed within one single Pre-trained Language Model (PLM). This model-specific setup is a substantial limitation on the very modularity that modular architectures are trying to achieve. We ask whether current modular approaches are transferable between models and whether we can transfer the modules from more robust and larger PLMs to smaller ones. In this work, we aim to fill this gap via a lens of Knowledge Distillation, commonly used for model compression, and present an extremely straightforward approach to transferring pre-trained, task-specific PEFT modules between same-family PLMs. Moreover, we propose a method that allows the transfer of modules between incompatible PLMs without any change in the inference complexity. The experiments on Named Entity Recognition, Natural Language Inference, and Paraphrase Identification tasks over multiple languages and PEFT methods showcase the initial potential of transferable modularity.

Compact Speech Translation Models via Discrete Speech Units Pretraining

Using Self-Supervised Learning (SSL) as model initialization is now common to obtain strong results in Speech Translation (ST). However, they also impose a large memory footprint, hindering on-device deployment. In this paper, we leverage the SSL models by pretraining smaller models on their Discrete Speech Units (DSU). We pretrain encoder-decoder models on 1) Filterbank-to-DSU and 2) DSU-to-Translation data, and take the encoder from 1) and the decoder from 2) to initialise a new model, finetuning this on limited speech-translation data. The final model becomes compact by using the DSU pretraining to distil the knowledge of the SSL model. Our method has several benefits over using DSU as model inputs, such as shorter inference pipeline and robustness over (DSU) tokenization. In contrast to ASR pretraining, it does not require transcripts, making it applicable to low-resource settings. Evaluation on CoVoST-2 X-En shows that our method is >$0.5$ BLEU better than a ST model that directly finetune the SSL model, given only half the model size, and on a par with ASR pretraining.

Code-Switched Language Identification is Harder Than You Think

Code switching (CS) is a very common phenomenon in written and spoken communication but one that is handled poorly by many natural language processing applications. Looking to the application of building CS corpora, we explore CS language identification (LID) for corpus building. We make the task more realistic by scaling it to more languages and considering models with simpler architectures for faster inference. We also reformulate the task as a sentence-level multi-label tagging problem to make it more tractable. Having defined the task, we investigate three reasonable models for this task and define metrics which better reflect desired performance. We present empirical evidence that no current approach is adequate and finally provide recommendations for future work in this area.

Prosody in Cascade and Direct Speech-to-Text Translation: a case study on Korean Wh-Phrases

Speech-to-Text Translation (S2TT) has typically been addressed with cascade systems, where speech recognition systems generate a transcription that is subsequently passed to a translation model. While there has been a growing interest in developing direct speech translation systems to avoid propagating errors and losing non-verbal content, prior work in direct S2TT has struggled to conclusively establish the advantages of integrating the acoustic signal directly into the translation process. This work proposes using contrastive evaluation to quantitatively measure the ability of direct S2TT systems to disambiguate utterances where prosody plays a crucial role. Specifically, we evaluated Korean-English translation systems on a test set containing wh-phrases, for which prosodic features are necessary to produce translations with the correct intent, whether it's a statement, a yes/no question, a wh-question, and more. Our results clearly demonstrate the value of direct translation systems over cascade translation models, with a notable 12.9% improvement in overall accuracy in ambiguous cases, along with up to a 15.6% increase in F1 scores for one of the major intent categories. To the best of our knowledge, this work stands as the first to provide quantitative evidence that direct S2TT models can effectively leverage prosody. The code for our evaluation is openly accessible and freely available for review and utilisation.

Machine Translation Meta Evaluation through Translation Accuracy Challenge Sets

Recent machine translation (MT) metrics calibrate their effectiveness by correlating with human judgement but without any insights about their behaviour across different error types. Challenge sets are used to probe specific dimensions of metric behaviour but there are very few such datasets and they either focus on a limited number of phenomena or a limited number of language pairs. We introduce ACES, a contrastive challenge set spanning 146 language pairs, aimed at discovering whether metrics can identify 68 translation accuracy errors. These phenomena range from simple alterations at the word/character level to more complex errors based on discourse and real-world knowledge. We conduct a large-scale study by benchmarking ACES on 50 metrics submitted to the WMT 2022 and 2023 metrics shared tasks. We benchmark metric performance, assess their incremental performance over successive campaigns, and measure their sensitivity to a range of linguistic phenomena. We also investigate claims that Large Language Models (LLMs) are effective as MT evaluators by evaluating on ACES. Our results demonstrate that different metric families struggle with different phenomena and that LLM-based methods fail to demonstrate reliable performance. Our analyses indicate that most metrics ignore the source sentence, tend to prefer surface-level overlap and end up incorporating properties of base models which are not always beneficial. We expand ACES to include error span annotations, denoted as SPAN-ACES and we use this dataset to evaluate span-based error metrics showing these metrics also need considerable improvement. Finally, we provide a set of recommendations for building better MT metrics, including focusing on error labels instead of scores, ensembling, designing strategies to explicitly focus on the source sentence, focusing on semantic content and choosing the right base model for representations.

Can GPT-3.5 Generate and Code Discharge Summaries?

Objective: To investigate GPT-3.5 in generating and coding medical documents with ICD-10 codes for data augmentation on low-resources labels. Materials and Methods: Employing GPT-3.5 we generated and coded 9,606 discharge summaries based on lists of ICD-10 code descriptions of patients with infrequent (generation) codes within the MIMIC-IV dataset. Combined with the baseline training set, this formed an augmented training set. Neural coding models were trained on baseline and augmented data and evaluated on a MIMIC-IV test set. We report micro- and macro-F1 scores on the full codeset, generation codes, and their families. Weak Hierarchical Confusion Matrices were employed to determine within-family and outside-of-family coding errors in the latter codesets. The coding performance of GPT-3.5 was evaluated both on prompt-guided self-generated data and real MIMIC-IV data. Clinical professionals evaluated the clinical acceptability of the generated documents. Results: Augmentation slightly hinders the overall performance of the models but improves performance for the generation candidate codes and their families, including one unseen in the baseline training data. Augmented models display lower out-of-family error rates. GPT-3.5 can identify ICD-10 codes by the prompted descriptions, but performs poorly on real data. Evaluators note the correctness of generated concepts while suffering in variety, supporting information, and narrative. Discussion and Conclusion: GPT-3.5 alone is unsuitable for ICD-10 coding. Augmentation positively affects generation code families but mainly benefits codes with existing examples. Augmentation reduces out-of-family errors. Discharge summaries generated by GPT-3.5 state prompted concepts correctly but lack variety, and authenticity in narratives. They are unsuitable for clinical practice.