Tokenization is often treated as a preprocessing step, yet in data-limited settings it directly shapes what a model can learn. We compare four segmentation strategies in the BabyLM Challenge: frequency-based BPE, morphology-aware MorPiece and ParadigmFinder, and syllable-based SylliTok. Evaluation combines two perspectives. First, an intrinsic test on the SIGMORPHON 2022 segmentation benchmark, adapted to English, measures how closely each tokenizer aligns with morpheme boundaries. Second, extrinsic tests train GPT-2 on the 10M BabyLM corpus and evaluate on the 2025 benchmark. No single tokenizer dominates. BPE remains strong on syntax-heavy tasks. ParadigmFinder excels in semantic composition and age-of-acquisition alignment. MorPiece shows advantages in discourse tracking. Morphology-aware tokenizers achieve the best intrinsic segmentation scores, and these gains translate into more robust generalisation in comprehension tasks. These results highlight tokenization as a core modeling decision, with direct consequences for compression, morphology, and the path to humanlike learning.

We discuss the strategies and results of a small-sized training program based on Italian child-directed speech (less than 3M tokens) for various network architectures. The rationale behind these experiments [1] lies in the attempt to understand the effect of this naturalistic training diet on different models architecture. Preliminary findings lead us to conclude that (a) different tokenization strategies produce only numerical, but not statistically significant, improvements overall, although segmentation aligns more or less with linguistic intuitions; and (b) modified LSTM networks with a single layer and a structurally more controlled cell state perform worse in training (compared to standard one- and two-layered LSTM models) but better on linguistically critical contrasts. This suggests that standard loss/accuracy metrics in autoregressive training procedures are linguistically irrelevant and, more generally, misleading, since the best-trained models qualify as poorer “linguistic theories” ([2], pace [3]).

This work explores alternative gating systems in simple Recurrent Neural Networks (RNNs) that induce linguistically motivated biases during training, ultimately affecting models’ performance on the BLiMP task. We focus exclusively on the BabyLM 10M training corpus (Strict-Small Track). Our experiments reveal that: (i) standard RNN variants—LSTMs and GRUs—are insufficient for properly learning the relevant set of linguistic constraints; (ii) the quality or size of the training corpus has little impact on these networks, as demonstrated by the comparable performance of LSTMs trained exclusively on the child-directed speech portion of the corpus; (iii) increasing the size of the embedding and hidden layers does not significantly improve performance. In contrast, specifically gated RNNs (eMG-RNNs), inspired by certain Minimalist Grammar intuitions, exhibit advantages in both training loss and BLiMP accuracy.

The availability of annotated legal corpora is crucial for a number of tasks, such as legal search, legal information retrieval, and predictive justice. Annotation is mostly assumed to be a straightforward task: as long as the annotation scheme is well defined and the guidelines are clear, annotators are expected to agree on the labels. This is not always the case, especially in legal annotation, which can be extremely difficult even for expert annotators. We propose a legal annotation procedure that takes into account annotator certainty and improves it through negotiation. We also collect annotator feedback and show that our approach contributes to a positive annotation environment. Our work invites reflection on often neglected ethical concerns regarding legal annotation.

We report the results of the SemEval 2022 Task 3, PreTENS, on evaluation the acceptability of simple sentences containing constructions whose two arguments are presupposed to be or not to be in an ordered taxonomic relation. The task featured two sub-tasks articulated as: (i) binary prediction task and (ii) regression task, predicting the acceptability in a continuous scale. The sentences were artificially generated in three languages (English, Italian and French). 21 systems, with 8 system papers were submitted for the task, all based on various types of fine-tuned transformer systems, often with ensemble methods and various data augmentation techniques. The best systems reached an F1-macro score of 94.49 (sub-task1) and a Spearman correlation coefficient of 0.80 (sub-task2), with interesting variations in specific constructions and/or languages.

This paper describes the efforts for the construction of Language Resources and NLP tools for Mirandese, a minority language spoken in North-eastern Portugal, now available on a community-led portal, Casa de la Lhéngua. The resources were developed in the context of a collaborative citizenship project led by Microsoft, in the context of the creation of the first TTS system for Mirandese. Development efforts encompassed the compilation of a corpus with over 1M tokens, the construction of a GTP system, syllable-division, inflection and a Part-of-Speech (POS) tagger modules, leading to the creation of an inflected lexicon of about 200.000 entries with phonetic transcription, detailed POS tagging, syllable division, and stress mark-up. Alongside these tasks, which were made easier through the adaptation and reuse of existing tools for closely related languages, a casting for voice talents among the speaking community was conducted and the first speech database for speech synthesis was recorded for Mirandese. These resources were combined to fulfil the requirements of a well-tested statistical parameter synthesis model, leading to an intelligible voice font. These language resources are available freely at Casa de la Lhéngua, aiming at promoting the development of real-life applications and fostering linguistic research on Mirandese.