The vast majority of today’s large language models (LLMs) are English-centric, having been pretrained predominantly on English text. Yet, in order to meet user expectations, models need to be able to respond appropriately in multiple languages once deployed in downstream applications. This requires strong cross-lingual transfer abilities.In this work, we investigate the minimal amount of multilinguality required during finetuning to elicit cross-lingual generalisation in English-centric LLMs. In experiments across four LLMs, we find that multilingual instruction tuning with as few as two to three languages is both necessary and sufficient to elicit effective cross-lingual generalisation, with the limiting factor being the degree to which a target language is seen during pretraining. Evaluations on five different tasks further reveal that multilingual instruction tuning is most beneficial for generative tasks that assume input/output language agreement, such as in chat settings, while being of less importance for highly structured classification-style tasks.

Some variants of self-supervised denoising objectives for pre-training encoder-decoder language models have been reported to have a negligible impact on downstream performance. Yet the design of these pre-training objectives leads to behavioural differences that can be uncovered with specific manipulations. We reproduce a recently proposed zero-shot control method and find that it is only successful on a subset of models. To understand what causes the difference in its effectiveness, we perform a set of controlled experiments, varying only the pre-training objective, and find unexpected interactions between the pre-training method and downstream controllability of models after fine-tuning. Our results show that different pre-training objectives have consequences that may not be visible in standard downstream evaluation, but which should be taken into account when developing models with controllability in mind.

We present BLESS, a comprehensive performance benchmark of the most recent state-of-the-art Large Language Models (LLMs) on the task of text simplification (TS). We examine how well off-the-shelf LLMs can solve this challenging task, assessing a total of 44 models, differing in size, architecture, pre-training methods, and accessibility, on three test sets from different domains (Wikipedia, news, and medical) under a few-shot setting. Our analysis considers a suite of automatic metrics, as well as a large-scale quantitative investigation into the types of common edit operations performed by the different models. Furthermore, we perform a manual qualitative analysis on a subset of model outputs to better gauge the quality of the generated simplifications. Our evaluation indicates that the best LLMs, despite not being trained on TS perform comparably with state-of-the-art TS baselines. Additionally, we find that certain LLMs demonstrate a greater range and diversity of edit operations. Our performance benchmark will be available as a resource for the development of future TS methods and evaluation metrics.

Automatic text simplification aims to reduce the linguistic complexity of a text in order to make it easier to understand and more accessible. However, simplified texts are consumed by a diverse array of target audiences and what might be appropriately simplified for one group of readers may differ considerably for another. In this work we investigate a novel formulation of sentence simplification as paraphrasing with controlled decoding. This approach aims to alleviate the major burden of relying on large amounts of in-domain parallel training data, while at the same time allowing for modular and adaptive simplification. According to automatic metrics, our approach performs competitively against baselines that prove more difficult to adapt to the needs of different target audiences or require significant amounts of complex-simple parallel aligned data.

Responding to online customer reviews has become an essential part of successfully managing and growing a business both in e-commerce and the hospitality and tourism sectors. Recently, neural text generation methods intended to assist authors in composing responses have been shown to deliver highly fluent and natural looking texts. However, they also tend to learn a strong, undesirable bias towards generating overly generic, one-size-fits-all outputs to a wide range of inputs. While this often results in ‘safe’, high-probability responses, there are many practical settings in which greater specificity is preferable. In this work we examine the task of generating more specific responses for online reviews in the hospitality domain by identifying generic responses in the training data, filtering them and fine-tuning the generation model. We experiment with a range of data-driven filtering methods and show through automatic and human evaluation that, despite a 60% reduction in the amount of training data, filtering helps to derive models that are capable of generating more specific, useful responses.

The task of document-level text simplification is very similar to summarization with the additional difficulty of reducing complexity. We introduce a newly collected data set of German texts, collected from the Swiss news magazine 20 Minuten (‘20 Minutes’) that consists of full articles paired with simplified summaries. Furthermore, we present experiments on automatic text simplification with the pretrained multilingual mBART and a modified version thereof that is more memory-friendly, using both our new data set and existing simplification corpora. Our modifications of mBART let us train at a lower memory cost without much loss in performance, in fact, the smaller mBART even improves over the standard model in a setting with multiple simplification levels.

Strong regional variation, together with the lack of standard orthography, makes Swiss German automatic speech recognition (ASR) particularly difficult in a multi-dialectal setting. This paper focuses on one of the many challenges, namely, the choice of the output text to represent non-standardised Swiss German. We investigate two potential options: a) dialectal writing – approximate phonemic transcriptions that provide close correspondence between grapheme labels and the acoustic signal but are highly inconsistent and b) normalised writing – transcriptions resembling standard German that are relatively consistent but distant from the acoustic signal. To find out which writing facilitates Swiss German ASR, we build several systems using the Kaldi toolkit and a dataset covering 14 regional varieties. A formal comparison shows that the system trained on the normalised transcriptions achieves better results in word error rate (WER) (29.39%) but underperforms at the character level, suggesting dialectal transcriptions offer a viable solution for downstream applications where dialectal differences are important. To better assess word-level performance for dialectal transcriptions, we use a flexible WER measure (FlexWER). When evaluated with this metric, the system trained on dialectal transcriptions outperforms that trained on the normalised writing. Besides establishing a benchmark for Swiss German multi-dialectal ASR, our findings can be helpful in designing ASR systems for other languages without standard orthography.

Online customer reviews are of growing importance for many businesses in the hospitality industry, particularly restaurants and hotels. Managerial responses to such reviews provide businesses with the opportunity to influence the public discourse and to attain improved ratings over time. However, responding to each and every review is a time-consuming endeavour. Therefore, we investigate automatic generation of review responses in the hospitality domain for two languages, English and German. We apply an existing system, originally proposed for review response generation for smartphone apps. This approach employs an extended neural network sequence-to-sequence architecture and performs well in the original domain. However, as shown through our experiments, when applied to a new domain, such as hospitality, performance drops considerably. Therefore, we analyse potential causes for the differences in performance and provide evidence to suggest that review response generation in the hospitality domain is a more challenging task and thus requires further study and additional domain adaptation techniques.

Geotagging historic and cultural texts provides valuable access to heritage data, enabling location-based searching and new geographically related discoveries. In this paper, we describe two distinct approaches to geotagging a variety of fine-grained toponyms in a diachronic corpus of alpine texts. By applying a traditional gazetteer-based approach, aided by a few simple heuristics, we attain strong high-precision annotations. Using the output of this earlier system, we adopt a state-of-the-art neural approach in order to facilitate the detection of new toponyms on the basis of context. Additionally, we present the results of preliminary experiments on integrating a small amount of crowdsourced annotations to improve overall performance of toponym recognition in our heritage corpus.