Linguistically-oriented formal NLI systems ensure the validity and transparency of inference. However, the combinatorial explosion of candidates, which we term the branching problem, imposes prohibitive computational overhead and a heavy cognitive burden on grammar developers. We argue that a central cause is a mismatch between the exhaustive execution paradigm and the actual workflow of grammar developers. To overcome this barrier, we propose restructuring the development workflow from exhaustive execution to interactive exploration driven by developer decisions. We realize this shift in Express, a web-based interactive development environment for lightblue, a Japanese automated inference system built upon Combinatory Categorial Grammar and Dependent Type Semantics. Express transforms branches at each stage of parsing, type checking, and proof search into explicitly selectable units, transferring control over the reasoning process to the developer. Our evaluation shows that this paradigm shift effectively reduces unnecessary computation and cognitive burden during grammar development: in a user study, we observed a 96% reduction in explored paths and improvement in the task success rate from 25% to 100%. Furthermore, a case study demonstrates a roughly 12× reduction in debugging turnaround time.

This paper proposes Neural Wani, an integration of a neural model into the automated theorem prover wani for Dependent Type Theory (DTT), aimed at accelerating proof search in natural language inference (NLI) pipelines. We implemented a lightweight LSTM-based model to predict the probability distribution of applicable inference rules and integrated it into wani’s backward inference process. Evaluation using the JSeM dataset demonstrates that Neural Wani achieves a 1.41x speedup compared to the standard non-neural baseline. Although slight overhead is observed in simpler proofs, our results indicate that neural-symbolic integration effectively guides search in complex DTT-based automated theorem proving.

In this study, we examine whether multilingual contextual embeddings encode properties of adjectives that are theoretically relevant to formal analyses of nominal modification. Using Universal Dependencies corpora for Arabic, English, and Italian, we extract contextualized adjective embeddings from the multilingual XLM-RoBERTa model and analyze them with respect to (i) semantic classes, (ii) the distinction between relational and descriptive adjectives, (iii) the distinction between gradable and non-gradable adjectives, and (iv) prenominal versus postnominal position in Italian. Our results indicate that adjective representations are organized in a shared multilingual space, but that this space is not best accounted for by a rigidly aligned universal hierarchy of semantic classes. Rather, the most salient organizing dimensions correspond to broader semantic-syntactic contrasts, in particular the relational/descriptive opposition, gradability, and, in the case of Italian, position-conditioned variation.

This paper investigates the evaluation of compositional generalization in Transformer models on the ReCOGS benchmark. The problem addressed is that ReCOGS relies on Semantic Exact Match, a binary metric that assigns the same penalty to minor local mismatches and severe structural errors, limiting diagnostic interpretation. To address this, the study introduces Compositional Graph Similarity (CGS), a graph-based metric that compares predicted and reference semantic structures through explicit edit operations, providing graded and interpretable structural evaluation. The work also uses controlled synthetic datasets to test whether low-scoring ReCOGS categories reflect true model limitations or weaknesses in dataset coverage. Empirical results show that CGS satisfies all seven quality criteria adopted for graph similarity and identifies the lowest-scoring ReCOGS categories as cp recursion (45.0%), obj pp to subj pp (65.4%), and prim to inf arg (66.7%). Follow-up experiments showed 0% Semantic Exact Match under depth extrapolation and constituent-role relocation, but 99.9% Semantic Exact Match for prim to inf arg in isolation. These findings support the conclusion that CGS is more informative than Semantic Exact Match and that Transformer limitations in ReCOGS are partly structural and partly induced by dataset distribution.

In this paper, we offer an extension of an earlier proposal for treating wh-questions within a compositional, neurally–implemented semantic framework that interfaces with memory to the other main class of questions, namely polar questions. Our proposal yields improved empirical coverage for polar questions as compared with previous formal semantic accounts. It also offers the basis for an account of the finding that understanding for wh-questions emerges in language development before that of polar questions–a finding that goes against all previous formal semantics accounts of questions where polar questions are simplest in terms of their semantic complexity.

Humans are pragmatic language users who naturally and effortlessly reason about the choice of utterances that help collaborate and engage in social interactions. In this paper, we examine whether vision-language models (VLMs) exhibit similar pragmatic reasoning effects through a validated artificial language learning paradigm. Across four experiments, we evaluate five VLMs’ sensitivity to production cost, ambiguity-driven competition effects, and the influences of visual features and model properties. We find evidence of cost effects in some VLMs. However, no model consistently exhibits competition effects driven by ambiguity risk, a hallmark of Gricean pragmatic reasoning. We also find that model scale alone does not predict pragmatic alignment; architectural choices play a larger role. Moreover, probability-based methods reveal clearer effects than prompting. Overall, current VLMs capture only a restricted subset of pragmatic effects central to Gricean reasoning, suggesting gaps in multimodal pragmatic reasoning.

Codec-based audio language models are developing, but little explainability research has been dedicated to the representation of this type of speech tokenisation. In this paper, we focus on the dictionary of 2048 tokens used in Mimi’s semantic token codebook, the neural codec of the Moshi language model (Défossez et al., 2024). We show that the ABX experiment carried out with Mimi fails to capture the mapping of the semantic tokens to phone realisations. By realigning Mimi’s representations to the TIMIT corpus transcriptions (Garofolo et al., 1993), we show that the 2048 tokens IDs of the semantic codebook map to quadphone, triphone, biphone, phone and subphone realisations. We used the TIMIT transcriptions as evidence of the validity of the allophone-based representations of these 80ms semantic token representation and examine some of the theoretical consequences for the tokenisation of speech at allophone and subphonemic level.

In this position paper, we argue that misalignments in common ground are not marginal failures of communication, but central diagnostic moments for pragmatic competence, and should therefore play a key role in the evaluation of Large Language Models (LLMs). Evaluating how models respond to such instances of mismatched or incomplete understanding moves beyond surface fluency and correctness, targeting pragmatic competence at a deeper, interactional level. At the same time, misalignments provide controlled settings for testing linguistic theories of common ground, repair, or accommodation – areas that are often difficult to investigate in human communication. We argue that this dual role makes misalignments a natural bridge between pragmatic theory and LLM evaluation.

Lemmatization is an important preprocessing step in Natural Language Processing (NLP); however, annotated resources for medieval languages such as Old Church Slavonic (OCS) are limited in scope, size, and diversity. This paper presents the annotated resources for OCS lemmatization, including annotation process, design choices and non-standard Unicode related issues. The annotated corpus is used to evaluate existing lemmatization tools (Stanza and UDPipe-2 models trained on the UD 2.12 treebank, and a dictionary-based approach) both in cross-dataset and on a corpus obtained by merging the new annotations with existing UD V2.12 OCS data. Pretrained models perform poorly (≈ 15–16%), below a dictionary baseline (≈ 38%), while retraining on the new data improves performance (up to ≈ 51%) and shows different cross-dataset generalization. Experiments in cross-dataset and on the combined corpus demonstrate that lemmatization performance depends strongly on dataset similarity, annotation conventions, and orthographic mismatch. Overall, the findings show the value of the newly annotated resources and the importance of extending OCS lemmatization benchmarks for historical Slavic NLP.

No natural language is known to have contrafactive attitude verbs, yet factives are common across natural languages. Several experiments by Strohmaier and Wimmer (2022; 2023; 2025) use transformers as model learners to investigate whether this asymmetry is due to a difference in how easy it is to learn contrafactives and factives. But they do not explore empirically-founded data distributions. We fill this gap, further improving the overall quality of training data distributions using linear programming.Our results confirm Strohmaier and Wimmer’s 2025 conclusion that there is no learnability difference in production, while establishing the impact of differences in data distributions.

Grammar engineering requires expertise in linguistic formalism and computational implementation, particularly in parallel grammar projects that balance cross-linguistic consistency with language-specific properties. This paper presents the development of Cantonese and Irish treebanks within the Parallel Grammar (ParGram) Project, where linguistic parallelism is maintained at an abstract functional level. We also investigated the methodological potential and limitations of using multilingual LLMs to support grammar engineering, focusing on Cantonese–Irish translation and the generation of formal syntactic structures using OpenAI’s gpt-oss-120b. The results showed that translation performance was generally unsatisfactory and unaffected by prompt language. For syntactic structure generation, the model produced some structurally meaningful outputs, but performed poorly on tasks requiring cross-linguistic abstraction. Nonetheless, LLM-generated outputs may still offer some reference value by suggesting alternative analyses and (partially) capturing predicate–argument relations. Overall, our findings highlight both the potential and limitations of using LLMs in collaborative grammar engineering, while underscoring the continued importance of expert-driven analysis and verification.

Negation is a useful case for studying linguistic structure because natural languages must distinguish between objects satisfying a predicate and those satisfying its complement.In emergent communication, however, a system may separate positive and negative meanings without developing a negation marker: polarity may be tied to identity, distributed across values, or reflected only in accidental correlations.We propose an information-theoretic account of negation encoding in discrete communication systems, yielding metrics for characterising lexical negation.We first study these metrics on toy languages, showing which encoding patterns they capture and which remain ambiguous.We then apply them to languages emerging in a signalling game with set-complement relations, under pressures known to favour compositionality.The results suggest that these pressures can produce high-scoring polarity-sensitive features, but not necessarily a compositional encoding of negation.More generally, we highlight both the usefulness and the limits of targeted semantic diagnostics for analysing structure in emergent languages.

Semantic type mismatch between a noun and its context is central to coercion phenomena. This paper introduces a graph-based method to examine how lexical and contextual type information is reflected in word embeddings. We select nouns from ten semantic types, annotate corpus instances for type matching (matching vs. coercion vs. other mismatch vs. unrestricted), and construct graphs using BERT and sense-enhanced embeddings. Two metrics—Neighbor Type Probability (NTP) and Neighbor Type Entropy (NTE)—are proposed to analyze neighborhood type distributions. Results show that graphs constructed with sense-enhanced embeddings reflect semantic type information better, and matching and mismatch sentences can be distinguished through the proposed metrics.