Large Language Models (LLMs) undergo extensive evaluation against various benchmarks collected in established leaderboards to assess their performance across multiple tasks. However, to the best of our knowledge, there is a lack of comprehensive studies evaluating these models’ linguistic abilities independent of specific tasks. In this paper, we introduce a novel evaluation methodology designed to test LLMs’ sentence generation abilities under specific linguistic constraints. Drawing on the ‘linguistic profiling’ approach, we rigorously investigate the extent to which five LLMs of varying sizes, tested in both zero- and few-shot scenarios, effectively adhere to (morpho)syntactic constraints. Our findings shed light on the linguistic proficiency of LLMs, revealing both their capabilities and limitations in generating linguistically-constrained sentences.

Automatic methods for generating and gathering linguistic data have proven effective for fine-tuning Language Models (LMs) in languages less resourced than English. Still, while there has been emphasis on data quantity, less attention has been given to its quality. In this work, we investigate the impact of human intervention on machine-generated data when fine-tuning dialogical models. In particular, we study (1) whether post-edited dialogues exhibit higher perceived quality compared to the originals that were automatically generated; (2) whether fine-tuning with post-edited dialogues results in noticeable differences in the generated outputs; and (3) whether post-edited dialogues influence the outcomes when considering the parameter size of the LMs. To this end we created HED-IT, a large-scale dataset where machine-generated dialogues are paired with the version post-edited by humans. Using both the edited and unedited portions of HED-IT, we fine-tuned three different sizes of an LM. Results from both human and automatic evaluation show that the different quality of training data is clearly perceived and it has an impact also on the models trained on such data. Additionally, our findings indicate that larger models are less sensitive to data quality, whereas this has a crucial impact on smaller models. These results enhance our comprehension of the impact of human intervention on training data in the development of high-quality LMs.

Large Language Models (LLMs) are increasingly used as ‘content farm’ models (CFMs), to generate synthetic text that could pass for real news articles. This is already happening even for languages that do not have high-quality monolingual LLMs. We show that fine-tuning Llama (v1), mostly trained on English, on as little as 40K Italian news articles, is sufficient for producing news-like texts that native speakers of Italian struggle to identify as synthetic.We investigate three LLMs and three methods of detecting synthetic texts (log-likelihood, DetectGPT, and supervised classification), finding that they all perform better than human raters, but they are all impractical in the real world (requiring either access to token likelihood information or a large dataset of CFM texts). We also explore the possibility of creating a proxy CFM: an LLM fine-tuned on a similar dataset to one used by the real ‘content farm’. We find that even a small amount of fine-tuning data suffices for creating a successful detector, but we need to know which base LLM is used, which is a major challenge.Our results suggest that there are currently no practical methods for detecting synthetic news-like texts ‘in the wild’, while generating them is too easy. We highlight the urgency of more NLP research on this problem.

In this paper, we explore the impact of augmenting pre-trained Encoder-Decoder models, specifically T5, with linguistic knowledge for the prediction of a target task. In particular, we investigate whether fine-tuning a T5 model on an intermediate task that predicts structural linguistic properties of sentences modifies its performance in the target task of predicting sentence-level complexity. Our study encompasses diverse experiments conducted on Italian and English datasets, employing both monolingual and multilingual T5 models at various sizes. Results obtained for both languages and in cross-lingual configurations show that linguistically motivated intermediate fine-tuning has generally a positive impact on target task performance, especially when applied to smaller models and in scenarios with limited data availability.

In this study, we investigate the capability of a Neural Language Model (NLM) to distinguish between coherent and incoherent text, where the latter has been artificially created to gradually undermine local coherence within text. While previous research on coherence assessment using NLMs has primarily focused on English, we extend our investigation to multiple languages. We employ a consistent evaluation framework to compare the performance of monolingual and multilingual models in both in-domain and out-domain settings. Additionally, we explore the model’s performance in a cross-language scenario.

While Transformer-based language models are generally very robust to pruning, there is the recently discovered outlier phenomenon: disabling only 48 out of 110M parameters in BERT-base drops its performance by nearly 30% on MNLI. We replicate the original evidence for the outlier phenomenon and we link it to the geometry of the embedding space. We find that in both BERT and RoBERTa the magnitude of hidden state coefficients corresponding to outlier dimensions correlate with the frequencies of encoded tokens in pre-training data, and they also contribute to the “vertical” self-attention pattern enabling the model to focus on the special tokens. This explains the drop in performance from disabling the outliers, and it suggests that to decrease anisotopicity in future models we need pre-training schemas that would better take into account the skewed token distributions.

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.

Several studies in the literature on the interpretation of Neural Language Models (NLM) focus on the linguistic generalization abilities of pre-trained models. However, little attention is paid to how the linguistic knowledge of the models changes during the fine-tuning steps. In this paper, we contribute to this line of research by showing to what extent a wide range of linguistic phenomena are forgotten across 50 epochs of fine-tuning, and how the preserved linguistic knowledge is correlated with the resolution of the fine-tuning task. To this end, we considered a quite understudied task where linguistic information plays the main role, i.e. the prediction of the evolution of written language competence of native language learners. In addition, we investigate whether it is possible to predict the fine-tuned NLM accuracy across the 50 epochs solely relying on the assessed linguistic competence. Our results are encouraging and show a high relationship between the model’s linguistic competence and its ability to solve a linguistically-based downstream task.

This paper presents a comprehensive set of probing experiments using a multilingual language model, XLM-R, for temporal relation classification between events in four languages. Results show an advantage of contextualized embeddings over static ones and a detrimen- tal role of sentence level embeddings. While obtaining competitive results against state-of-the-art systems, our probes indicate a lack of suitable encoded information to properly address this task.

This paper presents an investigation aimed at studying how the linguistic structure of a sentence affects the perplexity of two of the most popular Neural Language Models (NLMs), BERT and GPT-2. We first compare the sentence-level likelihood computed with BERT and the GPT-2’s perplexity showing that the two metrics are correlated. In addition, we exploit linguistic features capturing a wide set of morpho-syntactic and syntactic phenomena showing how they contribute to predict the perplexity of the two NLMs.

Several studies investigated the linguistic information implicitly encoded in Neural Language Models. Most of these works focused on quantifying the amount and type of information available within their internal representations and across their layers. In line with this scenario, we proposed a different study, based on Lasso regression, aimed at understanding how the information encoded by BERT sentence-level representations is arrange within its hidden units. Using a suite of several probing tasks, we showed the existence of a relationship between the implicit knowledge learned by the model and the number of individual units involved in the encodings of this competence. Moreover, we found that it is possible to identify groups of hidden units more relevant for specific linguistic properties.

This paper investigates the relationship between two complementary perspectives in the human assessment of sentence complexity and how they are modeled in a neural language model (NLM). The first perspective takes into account multiple online behavioral metrics obtained from eye-tracking recordings. The second one concerns the offline perception of complexity measured by explicit human judgments. Using a broad spectrum of linguistic features modeling lexical, morpho-syntactic, and syntactic properties of sentences, we perform a comprehensive analysis of linguistic phenomena associated with the two complexity viewpoints and report similarities and differences. We then show the effectiveness of linguistic features when explicitly leveraged by a regression model for predicting sentence complexity and compare its results with the ones obtained by a fine-tuned neural language model. We finally probe the NLM’s linguistic competence before and after fine-tuning, highlighting how linguistic information encoded in representations changes when the model learns to predict complexity.

We study the influence of context on how humans evaluate the complexity of a sentence in English. We collect a new dataset of sentences, where each sentence is rated for perceived complexity within different contextual windows. We carry out an in-depth analysis to detect which linguistic features correlate more with complexity judgments and with the degree of agreement among annotators. We train several regression models, using either explicit linguistic features or contextualized word embeddings, to predict the mean complexity values assigned to sentences in the different contextual windows, as well as their standard deviation. Results show that models leveraging explicit features capturing morphosyntactic and syntactic phenomena perform always better, especially when they have access to features extracted from all contextual sentences.

We ask subjects whether they perceive as human-produced a bunch of texts, some of which are actually human-written, while others are automatically generated. We use this data to fine-tune a GPT-2 model to push it to generate more human-like texts, and observe that this fine-tuned model produces texts that are indeed perceived more human-like than the original model. Contextually, we show that our automatic evaluation strategy well correlates with human judgements. We also run a linguistic analysis to unveil the characteristics of human- vs machine-perceived language.

In this paper we investigate the linguistic knowledge learned by a Neural Language Model (NLM) before and after a fine-tuning process and how this knowledge affects its predictions during several classification problems. We use a wide set of probing tasks, each of which corresponds to a distinct sentence-level feature extracted from different levels of linguistic annotation. We show that BERT is able to encode a wide range of linguistic characteristics, but it tends to lose this information when trained on specific downstream tasks. We also find that BERT’s capacity to encode different kind of linguistic properties has a positive influence on its predictions: the more it stores readable linguistic information of a sentence, the higher will be its capacity of predicting the expected label assigned to that sentence.

In this paper we present a comparison between the linguistic knowledge encoded in the internal representations of a contextual Language Model (BERT) and a contextual-independent one (Word2vec). We use a wide set of probing tasks, each of which corresponds to a distinct sentence-level feature extracted from different levels of linguistic annotation. We show that, although BERT is capable of understanding the full context of each word in an input sequence, the implicit knowledge encoded in its aggregated sentence representations is still comparable to that of a contextual-independent model. We also find that BERT is able to encode sentence-level properties even within single-word embeddings, obtaining comparable or even superior results than those obtained with sentence representations.

An ongoing debate in the NLG community concerns the best way to evaluate systems, with human evaluation often being considered the most reliable method, compared to corpus-based metrics. However, tasks involving subtle textual differences, such as style transfer, tend to be hard for humans to perform. In this paper, we propose an evaluation method for this task based on purposely-trained classifiers, showing that it better reflects system differences than traditional metrics such as BLEU.

In this paper we present an NLP-based approach for tracking the evolution of written language competence in L2 Spanish learners using a wide range of linguistic features automatically extracted from students’ written productions. Beyond reporting classification results for different scenarios, we explore the connection between the most predictive features and the teaching curriculum, finding that our set of linguistic features often reflect the explicit instructions that students receive during each course.

“Voices of the Great War” is the first large corpus of Italian historical texts dating back to the period of First World War. This corpus differs from other existing resources in several respects. First, from the linguistic point of view it gives account of the wide range of varieties in which Italian was articulated in that period, namely from a diastratic (educated vs. uneducated writers), diaphasic (low/informal vs. high/formal registers) and diatopic (regional varieties, dialects) points of view. From the historical perspective, through a collection of texts belonging to different genres it represents different views on the war and the various styles of narrating war events and experiences. The final corpus is balanced along various dimensions, corresponding to the textual genre, the language variety used, the author type and the typology of conveyed contents. The corpus is fully annotated with lemmas, part-of-speech, terminology, and named entities. Significant corpus samples representative of the different “voices” have also been enriched with meta-linguistic and syntactic information. The layer of syntactic annotation forms the first nucleus of an Italian historical treebank complying with the Universal Dependencies standard. The paper illustrates the final resource, the methodology and tools used to build it, and the Web Interface for navigating it.

We automatically generate headlines that are expected to comply with the specific styles of two different Italian newspapers. Through a data alignment strategy and different training/testing settings, we aim at decoupling content from style and preserve the latter in generation. In order to evaluate the generated headlines’ quality in terms of their specific newspaper-compliance, we devise a fine-grained evaluation strategy based on automatic classification. We observe that our models do indeed learn newspaper-specific style. Importantly, we also observe that humans aren’t reliable judges for this task, since although familiar with the newspapers, they are not able to discern their specific styles even in the original human-written headlines. The utility of automatic evaluation goes therefore beyond saving the costs and hurdles of manual annotation, and deserves particular care in its design.

In this paper, we introduce Profiling–UD, a new text analysis tool inspired to the principles of linguistic profiling that can support language variation research from different perspectives. It allows the extraction of more than 130 features, spanning across different levels of linguistic description. Beyond the large number of features that can be monitored, a main novelty of Profiling–UD is that it has been specifically devised to be multilingual since it is based on the Universal Dependencies framework. In the second part of the paper, we demonstrate the effectiveness of these features in a number of theoretical and applicative studies in which they were successfully used for text and author profiling.

We present a new concept prerequisite learning method for Learning Object (LO) ordering that exploits only linguistic features extracted from textual educational resources. The method was tested in a cross- and in- domain scenario both for Italian and English. Additionally, we performed experiments based on a incremental training strategy to study the impact of the training set size on the classifier performances. The paper also introduces ITA-PREREQ, to the best of our knowledge the first Italian dataset annotated with prerequisite relations between pairs of educational concepts, and describe the automatic strategy devised to build it.

Detection and correction of errors and inconsistencies in “gold treebanks” are becoming more and more central topics of corpus annotation. The paper illustrates a new incremental method for enhancing treebanks, with particular emphasis on the extension of error patterns across different textual genres and registers. Impact and role of corrections have been assessed in a dependency parsing experiment carried out with four different parsers, whose results are promising. For both evaluation datasets, the performance of parsers increases, in terms of the standard LAS and UAS measures and of a more focused measure taking into account only relations involved in error patterns, and at the level of individual dependencies.

In this paper, we present a crowdsourcing-based approach to model the human perception of sentence complexity. We collect a large corpus of sentences rated with judgments of complexity for two typologically-different languages, Italian and English. We test our approach in two experimental scenarios aimed to investigate the contribution of a wide set of lexical, morpho-syntactic and syntactic phenomena in predicting i) the degree of agreement among annotators independently from the assigned judgment and ii) the perception of sentence complexity.

In this paper, we describe the approach of the ItaliaNLP Lab team to native language identification and discuss the results we submitted as participants to the essay track of NLI Shared Task 2017. We introduce for the first time a 2-stacked sentence-document architecture for native language identification that is able to exploit both local sentence information and a wide set of general-purpose features qualifying the lexical and grammatical structure of the whole document. When evaluated on the official test set, our sentence-document stacked architecture obtained the best result among all the participants of the essay track with an F1 score of 0.8818.

In this paper, we present the CItA corpus (Corpus Italiano di Apprendenti L1), a collection of essays written by Italian L1 learners collected during the first and second year of lower secondary school. The corpus was built in the framework of an interdisciplinary study jointly carried out by computational linguistics and experimental pedagogists and aimed at tracking the development of written language competence over the years and students’ background information.

In this paper, we present T2K^2, a suite of tools for automatically extracting domain―specific knowledge from collections of Italian and English texts. T2K^2 (Text―To―Knowledge v2) relies on a battery of tools for Natural Language Processing (NLP), statistical text analysis and machine learning which are dynamically integrated to provide an accurate and incremental representation of the content of vast repositories of unstructured documents. Extracted knowledge ranges from domain―specific entities and named entities to the relations connecting them and can be used for indexing document collections with respect to different information types. T2K^2 also includes “linguistic profiling” functionalities aimed at supporting the user in constructing the acquisition corpus, e.g. in selecting texts belonging to the same genre or characterized by the same degree of specialization or in monitoring the “added value” of newly inserted documents. T2K^2 is a web application which can be accessed from any browser through a personal account which has been tested in a wide range of domains.

The creation of language resources for less-resourced languages like the historical ones benefits from the exploitation of language-independent tools and methods developed over the years by many projects for modern languages. Along these lines, a number of treebanks for historical languages started recently to arise, including treebanks for Latin. Among the Latin treebanks, the Index Thomisticus Treebank is a 68,000 token dependency treebank based on the Index Thomisticus by Roberto Busa SJ, which contains the opera omnia of Thomas Aquinas (118 texts) as well as 61 texts by other authors related to Thomas, for a total of approximately 11 million tokens. In this paper, we describe a number of modifications that we applied to the dependency parser DeSR, in order to improve the parsing accuracy rates on the Index Thomisticus Treebank. First, we adapted the parser to the specific processing of Medieval Latin, defining an ad-hoc configuration of its features. Then, in order to improve the accuracy rates provided by DeSR, we applied a revision parsing method and we combined the outputs produced by different algorithms. This allowed us to improve accuracy rates substantially, reaching results that are well beyond the state of the art of parsing for Latin.

As the interest of the NLP community grows to develop several treebanks also for languages other than English, we observe efforts towards evaluating the impact of different annotation strategies used to represent particular languages or with reference to particular tasks. This paper contributes to the debate on the influence of resources used for the training and development on the performance of parsing systems. It presents a comparative analysis of the results achieved by three different dependency parsers developed and tested with respect to two treebanks for the Italian language, namely TUT and ISST--TANL, which differ significantly at the level of both corpus composition and adopted dependency representations.

In this paper, we present a novel approach to multi-word terminology extraction combining a well-known automatic term recognition approach, the C--NC value method, with a contrastive ranking technique, aimed at refining obtained results either by filtering noise due to common words or by discerning between semantically different types of terms within heterogeneous terminologies. Differently from other contrastive methods proposed in the literature that focus on single terms to overcome the multi-word terms' sparsity problem, the proposed contrastive function is able to handle variation in low frequency events by directly operating on pre-selected multi-word terms. This methodology has been tested in two case studies carried out in the History of Art and Legal domains. Evaluation of achieved results showed that the proposed two--stage approach improves significantly multi--word term extraction results. In particular, for what concerns the legal domain it provides an answer to a well-known problem in the semi--automatic construction of legal ontologies, namely that of singling out law terms from terms of the specific domain being regulated.

The paper reports on a detailed quantitative analysis of distributional language data of both Italian and Czech, highlighting the relative contribution of a number of distributed grammatical factors to sentence-based identification of subjects and direct objects. The work is based on a Maximum Entropy model of stochastic resolution of grammatical conflicting constraints, and is demonstrably capable of putting explanatory theoretical accounts to the challenging test of an extensive, usage-based empirical verification.