Recent advances in large language models (LLMs) have enabled impressive performance in various tasks. However, standard prompting often struggles to produce structurally valid and accurate outputs, especially in dependency parsing. We propose a novel step-by-step instruction strategy, where universal part-of-speech tagging precedes the prediction of syntactic heads and dependency labels, and a simplified CoNLL-U like output format, our method achieves state-of-the-art accuracy on Universal Dependencies datasets across 17 languages without hallucination or contamination. We further show that multilingual fine-tuning simultaneously improves cross-language generalization performance. Our results highlight the effectiveness of explicit reasoning steps in LLM-based parsing and offer a scalable, format-consistent alternative to bracket-based approaches.

We present an open-source web service for Czech morphosyntactic analysis. The system combines a deep learning model with rescoring by a high-precision morphological dictionary at inference time. We show that our hybrid method surpasses two competitive baselines: While the deep learning model ensures generalization for out-of-vocabulary words and better disambiguation, an improvement over an existing morphological analyser MorphoDiTa, at the same time, the deep learning model benefits from inference-time guidance of a manually curated morphological dictionary. We achieve 50% error reduction in lemmatization and 58% error reduction in POS tagging over MorphoDiTa, while also offering dependency parsing. The model is trained on one of the currently largest Czech morphosyntactic corpora, the PDT-C 1.0, with the trained models available at https://hdl.handle.net/11234/1-5293. We provide the tool as a web service deployed at https://lindat.mff.cuni.

Here we take a standard method found in physics used to remove known background functions from data, for example removing the spectra associated with amorphous radiators from those associated with lattice-structure radiators to obtain enhanced spectra, that is without noise (Timm 1969). Here we consider the variations associated with covariants as similar background data to be removed, so as to observe if there is any variation associated with EDV. Similar to partial correlations, removing the background signal of a potential covariant allows us to visually evaluate the specific impact a variable of interest has on the target variable. This involves fitting the control data and the target (e.g., the size of training data and LAS) and then dividing the target variable by the predicted values from this fit. This normalized data is then used to fit a second potential covariant which too is used to divide the normalized target variable values. This can be repeated for any number of covariants. Ultimately a normalized version of the target variable is left and the control target of interest (e.g., EDV) is evaluated against these values and if a trend is still observed, it is evidence that this variable has an impact on the target variable even with the variance associated with these covariants removed. This technique ultimately acts as a way of tempering correlations we calculate and gives us a means of disentangling contributions that might not be caught by partial correlation calculations.

S{\o}gaard (2020) obtained results suggesting the fraction of trees occurring in the test data isomorphic to trees in the training set accounts for a non-trivial variation in parser performance. Similar to other statistical analyses in NLP, the results were based on evaluating linear regressions. However, the study had methodological issues and was undertaken using a small sample size leading to unreliable results. We present a replication study in which we also bin sentences by length and find that only a small subset of sentences vary in performance with respect to graph isomorphism. Further, the correlation observed between parser performance and graph isomorphism in the wild disappears when controlling for covariants. However, in a controlled experiment, where covariants are kept fixed, we do observe a strong correlation. We suggest that conclusions drawn from statistical analyses like this need to be tempered and that controlled experiments can complement them by more readily teasing factors apart.