Andrew Caines

Andrew Caines, Luca Benedetto, Shiva Taslimipoor et al. · cambridge

The recent release of very large language models such as PaLM and GPT-4 has made an unprecedented impact in the popular media and public consciousness, giving rise to a mixture of excitement and fear as to their capabilities and potential uses, and shining a light on natural language processing research which had not previously received so much attention. The developments offer great promise for education technology, and in this paper we look specifically at the potential for incorporating large language models in AI-driven language teaching and assessment systems. We consider several research areas and also discuss the risks and ethical considerations surrounding generative AI in education technology for language learners. Overall we find that larger language models offer improvements over previous models in text generation, opening up routes toward content generation which had not previously been plausible. For text generation they must be prompted carefully and their outputs may need to be reshaped before they are ready for use. For automated grading and grammatical error correction, tasks whose progress is checked on well-known benchmarks, early investigations indicate that large language models on their own do not improve on state-of-the-art results according to standard evaluation metrics. For grading it appears that linguistic features established in the literature should still be used for best performance, and for error correction it may be that the models can offer alternative feedback styles which are not measured sensitively with existing methods. In all cases, there is work to be done to experiment with the inclusion of large language models in education technology for language learners, in order to properly understand and report on their capacities and limitations, and to ensure that foreseeable risks such as misinformation and harmful bias are mitigated.

Richard Diehl Martinez, Zebulon Goriely, Hope McGovern et al.

We describe our team's contribution to the STRICT-SMALL track of the BabyLM Challenge. The challenge requires training a language model from scratch using only a relatively small training dataset of ten million words. We experiment with three variants of cognitively-motivated curriculum learning and analyze their effect on the performance of the model on linguistic evaluation tasks. In the vocabulary curriculum, we analyze methods for constraining the vocabulary in the early stages of training to simulate cognitively more plausible learning curves. In the data curriculum experiments, we vary the order of the training instances based on i) infant-inspired expectations and ii) the learning behavior of the model. In the objective curriculum, we explore different variations of combining the conventional masked language modeling task with a more coarse-grained word class prediction task to reinforce linguistic generalization capabilities. Our results did not yield consistent improvements over our own non-curriculum learning baseline across a range of linguistic benchmarks; however, we do find marginal gains on select tasks. Our analysis highlights key takeaways for specific combinations of tasks and settings which benefit from our proposed curricula. We moreover determine that careful selection of model architecture, and training hyper-parameters yield substantial improvements over the default baselines provided by the BabyLM challenge.

Christopher Davis, Christopher Bryant, Andrew Caines et al.

Targeted studies testing knowledge of subject-verb agreement (SVA) indicate that pre-trained language models encode syntactic information. We assert that if models robustly encode subject-verb agreement, they should be able to identify when agreement is correct and when it is incorrect. To that end, we propose grammatical error detection as a diagnostic probe to evaluate token-level contextual representations for their knowledge of SVA. We evaluate contextual representations at each layer from five pre-trained English language models: BERT, XLNet, GPT-2, RoBERTa, and ELECTRA. We leverage public annotated training data from both English second language learners and Wikipedia edits, and report results on manually crafted stimuli for subject-verb agreement. We find that masked language models linearly encode information relevant to the detection of SVA errors, while the autoregressive models perform on par with our baseline. However, we also observe a divergence in performance when probes are trained on different training sets, and when they are evaluated on different syntactic constructions, suggesting the information pertaining to SVA error detection is not robustly encoded.

Kamil Bujel, Andrew Caines, Helen Yannakoudakis et al.

Long-sequence transformers are designed to improve the representation of longer texts by language models and their performance on downstream document-level tasks. However, not much is understood about the quality of token-level predictions in long-form models. We investigate the performance of such architectures in the context of document classification with unsupervised rationale extraction. We find standard soft attention methods to perform significantly worse when combined with the Longformer language model. We propose a compositional soft attention architecture that applies RoBERTa sentence-wise to extract plausible rationales at the token-level. We find this method to significantly outperform Longformer-driven baselines on sentiment classification datasets, while also exhibiting significantly lower runtimes.

Christopher Davis, Andrew Caines, Øistein Andersen et al.

Thanks to recent advances in generative AI, we are able to prompt large language models (LLMs) to produce texts which are fluent and grammatical. In addition, it has been shown that we can elicit attempts at grammatical error correction (GEC) from LLMs when prompted with ungrammatical input sentences. We evaluate how well LLMs can perform at GEC by measuring their performance on established benchmark datasets. We go beyond previous studies, which only examined GPT* models on a selection of English GEC datasets, by evaluating seven open-source and three commercial LLMs on four established GEC benchmarks. We investigate model performance and report results against individual error types. Our results indicate that LLMs do not always outperform supervised English GEC models except in specific contexts -- namely commercial LLMs on benchmarks annotated with fluency corrections as opposed to minimal edits. We find that several open-source models outperform commercial ones on minimal edit benchmarks, and that in some settings zero-shot prompting is just as competitive as few-shot prompting.

Filip Trhlik, Andrew Caines, Paula Buttery

Pre-trained language models (LMs) have, over the last few years, grown substantially in both societal adoption and training costs. This rapid growth in size has constrained progress in understanding and mitigating their biases. Since re-training LMs is prohibitively expensive, most debiasing work has focused on post-hoc or masking-based strategies, which often fail to address the underlying causes of bias. In this work, we seek to democratise pre-model debiasing research by using low-cost proxy models. Specifically, we investigate BabyLMs, compact BERT-like models trained on small and mutable corpora that can approximate bias acquisition and learning dynamics of larger models. We show that BabyLMs display closely aligned patterns of intrinsic bias formation and performance development compared to standard BERT models, despite their drastically reduced size. Furthermore, correlations between BabyLMs and BERT hold across multiple intra-model and post-model debiasing methods. Leveraging these similarities, we conduct pre-model debiasing experiments with BabyLMs, replicating prior findings and presenting new insights regarding the influence of gender imbalance and toxicity on bias formation. Our results demonstrate that BabyLMs can serve as an effective sandbox for large-scale LMs, reducing pre-training costs from over 500 GPU-hours to under 30 GPU-hours. This provides a way to democratise pre-model debiasing research and enables faster, more accessible exploration of methods for building fairer LMs.

Zébulon Goriely, Richard Diehl Martinez, Andrew Caines et al.

Language models are typically trained on large corpora of text in their default orthographic form. However, this is not the only option; representing data as streams of phonemes can offer unique advantages, from deeper insights into phonological language acquisition to improved performance on sound-based tasks. The challenge lies in evaluating the impact of phoneme-based training, as most benchmarks are also orthographic. To address this, we develop a pipeline to convert text datasets into a continuous stream of phonemes. We apply this pipeline to the 100-million-word pre-training dataset from the BabyLM challenge, as well as to standard language and grammatical benchmarks, enabling us to pre-train and evaluate a model using phonemic input representations. Our results show that while phoneme-based training slightly reduces performance on traditional language understanding tasks, it offers valuable analytical and practical benefits.

Richard Diehl Martinez, Zebulon Goriely, Andrew Caines et al.

Language models strongly rely on frequency information because they maximize the likelihood of tokens during pre-training. As a consequence, language models tend to not generalize well to tokens that are seldom seen during training. Moreover, maximum likelihood training has been discovered to give rise to anisotropy: representations of tokens in a model tend to cluster tightly in a high-dimensional cone, rather than spreading out over their representational capacity. Our work introduces a method for quantifying the frequency bias of a language model by assessing sentence-level perplexity with respect to token-level frequency. We then present a method for reducing the frequency bias of a language model by inducing a syntactic prior over token representations during pre-training. Our Syntactic Smoothing method adjusts the maximum likelihood objective function to distribute the learning signal to syntactically similar tokens. This approach results in better performance on infrequent English tokens and a decrease in anisotropy. We empirically show that the degree of anisotropy in a model correlates with its frequency bias.

Kelvin Wey Han Chan, Christopher Bryant, Li Nguyen et al.

Code-switching (CSW) is a common phenomenon among multilingual speakers where multiple languages are used in a single discourse or utterance. Mixed language utterances may still contain grammatical errors however, yet most existing Grammar Error Correction (GEC) systems have been trained on monolingual data and not developed with CSW in mind. In this work, we conduct the first exploration into the use of GEC systems on CSW text. Through this exploration, we propose a novel method of generating synthetic CSW GEC datasets by translating different spans of text within existing GEC corpora. We then investigate different methods of selecting these spans based on CSW ratio, switch-point factor and linguistic constraints, and identify how they affect the performance of GEC systems on CSW text. Our best model achieves an average increase of 1.57 $F_{0.5}$ across 3 CSW test sets (English-Chinese, English-Korean and English-Japanese) without affecting the model's performance on a monolingual dataset. We furthermore discovered that models trained on one CSW language generalise relatively well to other typologically similar CSW languages.

Jesujoba O. Alabi, Israel Abebe Azime, Miaoran Zhang et al.

This paper introduces AFRIDOC-MT, a document-level multi-parallel translation dataset covering English and five African languages: Amharic, Hausa, Swahili, Yorùbá, and Zulu. The dataset comprises 334 health and 271 information technology news documents, all human-translated from English to these languages. We conduct document-level translation benchmark experiments by evaluating neural machine translation (NMT) models and large language models (LLMs) for translations between English and these languages, at both the sentence and pseudo-document levels. These outputs are realigned to form complete documents for evaluation. Our results indicate that NLLB-200 achieved the best average performance among the standard NMT models, while GPT-4o outperformed general-purpose LLMs. Fine-tuning selected models led to substantial performance gains, but models trained on sentences struggled to generalize effectively to longer documents. Furthermore, our analysis reveals that some LLMs exhibit issues such as under-generation, repetition of words or phrases, and off-target translations, especially for African languages.

Suchir Salhan, Hongyi Gu, Donya Rooein et al.

Multi-turn dialogues between a child and a caregiver are characterized by a property called contingency - that is, prompt, direct, and meaningful exchanges between interlocutors. We introduce ContingentChat, a teacher-student framework that benchmarks and improves multi-turn contingency in a BabyLM trained on 100M words. Using a novel alignment dataset for post-training, BabyLM generates responses that are more grammatical and cohesive. Experiments with adaptive teacher decoding strategies show limited additional gains. ContingentChat demonstrates the benefits of targeted post-training for dialogue quality and indicates that contingency remains a challenging goal for BabyLMs.

Yuan Gao, Suchir Salhan, Andrew Caines et al.

To bridge the gap between performance-oriented benchmarks and the evaluation of cognitively inspired models, we introduce BLiSS 1.0, a Benchmark of Learner Interlingual Syntactic Structure. Our benchmark operationalizes a new paradigm of selective tolerance, testing whether a model finds a naturalistic learner error more plausible than a matched, artificial error within the same sentence. Constructed from over 2.8 million naturalistic learner sentences, BLiSS provides 136,867 controlled triplets (corrected, learner, artificial) for this purpose. Experiments on a diverse suite of models demonstrate that selective tolerance is a distinct capability from standard grammaticality, with performance clustering strongly by training paradigm. This validates BLiSS as a robust tool for measuring how different training objectives impact a model's alignment with the systematic patterns of human language acquisition.

Bianca-Mihaela Ganescu, Suchir Salhan, Andrew Caines et al.

Training vision-language models on cognitively-plausible amounts of data requires rethinking how models integrate multimodal information. Within the constraints of the Vision track for the BabyLM Challenge 2025, we propose a lightweight decoder-based architecture with (1) token-wise dynamic gating for adaptive fusion of linguistic and visual cues, (2) feature modulation and channel attention to maximise the utility of limited visual information and (3) auxiliary contrastive objectives for visual grounding. Evaluation on five benchmarks (BLiMP, BLiMP Supplement, EWoK, Winoground and VQA) shows competitive or superior performance to multimodal baselines. More notably, our dynamic gate discovers interpretable patterns without explicit supervision, favouring visual cues for content words and linguistic cues for function words. While we identify limitations in the Challenge constraints, such as the information bottleneck created by global image embeddings and training instability from the dataset split, our findings establish dynamic gating as a powerful tool for efficient multimodal learning, offering both interpretability and performance even under severe constraints.

Shantanu Thorat, Andrew Caines

Existing AIG (AI-generated) text detectors struggle in real-world settings despite succeeding in internal testing, suggesting that they may not be robust enough. We rigorously examine the machine-learning procedure to build these detectors to address this. Most current AIG text detection datasets focus on zero-shot generations, but little work has been done on few-shot or one-shot generations, where LLMs are given human texts as an example. In response, we introduce the Diverse Adversarial Corpus of Texts Yielded from Language models (DACTYL), a challenging AIG text detection dataset focusing on one-shot/few-shot generations. We also include texts from domain-specific continued-pre-trained (CPT) language models, where we fully train all parameters using a memory-efficient optimization approach. Many existing AIG text detectors struggle significantly on our dataset, indicating a potential vulnerability to one-shot/few-shot and CPT-generated texts. We also train our own classifiers using two approaches: standard binary cross-entropy (BCE) optimization and a more recent approach, deep X-risk optimization (DXO). While BCE-trained classifiers marginally outperform DXO classifiers on the DACTYL test set, the latter excels on out-of-distribution (OOD) texts. In our mock deployment scenario in student essay detection with an OOD student essay dataset, the best DXO classifier outscored the best BCE-trained classifier by 50.56 macro-F1 score points at the lowest false positive rates for both. Our results indicate that DXO classifiers generalize better without overfitting to the test set. Our experiments highlight several areas of improvement for AIG text detectors.

Luna Wang, Andrew Caines, Alice Hutchings

The curation of hate speech datasets involves complex design decisions that balance competing priorities. This paper critically examines these methodological choices in a diverse range of datasets, highlighting common themes and practices, and their implications for dataset reliability. Drawing on Max Weber's notion of ideal types, we argue for a reflexive approach in dataset creation, urging researchers to acknowledge their own value judgments during dataset construction, fostering transparency and methodological rigour.

Felermino D. M. A. Ali, Andrew Caines, Jaimito L. A. Malavi

Major advancement in the performance of machine translation models has been made possible in part thanks to the availability of large-scale parallel corpora. But for most languages in the world, the existence of such corpora is rare. Emakhuwa, a language spoken in Mozambique, is like most African languages low-resource in NLP terms. It lacks both computational and linguistic resources and, to the best of our knowledge, few parallel corpora including Emakhuwa already exist. In this paper we describe the creation of the Emakhuwa-Portuguese parallel corpus, which is a collection of texts from the Jehovah's Witness website and a variety of other sources including the African Story Book website, the Universal Declaration of Human Rights and Mozambican legal documents. The dataset contains 47,415 sentence pairs, amounting to 699,976 word tokens of Emakhuwa and 877,595 word tokens in Portuguese. After normalization processes which remain to be completed, the corpus will be made freely available for research use.

Andrew Caines, Helen Yannakoudakis, Helena Edmondson et al.

The Teacher-Student Chatroom Corpus (TSCC) is a collection of written conversations captured during one-to-one lessons between teachers and learners of English. The lessons took place in an online chatroom and therefore involve more interactive, immediate and informal language than might be found in asynchronous exchanges such as email correspondence. The fact that the lessons were one-to-one means that the teacher was able to focus exclusively on the linguistic abilities and errors of the student, and to offer personalised exercises, scaffolding and correction. The TSCC contains more than one hundred lessons between two teachers and eight students, amounting to 13.5K conversational turns and 133K words: it is freely available for research use. We describe the corpus design, data collection procedure and annotations added to the text. We perform some preliminary descriptive analyses of the data and consider possible uses of the TSCC.

Ahmed Zaidi, Andrew Caines, Russell Moore et al.

The forgetting curve has been extensively explored by psychologists, educationalists and cognitive scientists alike. In the context of Intelligent Tutoring Systems, modelling the forgetting curve for each user and knowledge component (e.g. vocabulary word) should enable us to develop optimal revision strategies that counteract memory decay and ensure long-term retention. In this study we explore a variety of forgetting curve models incorporating psychological and linguistic features, and we use these models to predict the probability of word recall by learners of English as a second language. We evaluate the impact of the models and their features using data from an online vocabulary teaching platform and find that word complexity is a highly informative feature which may be successfully learned by a neural network model.