Few-shot learning via in-context learning (ICL) is widely used in NLP, but its effectiveness is highly sensitive to example selection, often leading to unstable performance. To address this, we introduce BacKGen, a framework for generating structured Background Knowledge (BK) as an alternative to instance-based prompting. Our approach leverages Frame Semantics to uncover recurring conceptual patterns across data instances, clustering examples based on shared event structures and semantic roles. These patterns are then synthesized into generalized knowledge statements using a large language model (LLM) and injected into prompts to support contextual reasoning beyond surface-level cues. We apply BacKGen to Sentiment Phrase Classification (SPC), a task where polarity judgments frequently depend on implicit commonsense knowledge. In this setting, BK serves as an abstract representation of prototypical scenarios, enabling schematic generalization to help the model perform analogical reasoning by mapping new inputs onto generalized event structures. Experimental results with Mistral-7B and Llama3-8B demonstrate that BK-based prompting consistently outperforms standard few-shot approaches, achieving up to 29.94% error reduction.

Grounded natural language understanding in Human-Robot Interaction (HRI) requires integrating linguistic, visual, and world knowledge to ensure effective task execution. We propose an approach that enhances Multi-Modal Large Language Models (MLLMs) with a novel explicit dialogue planning phase, allowing robotic agents to systematically refine their understanding of ambiguous commands through structured clarification steps. This reduces hallucinations and improves task feasibility.To evaluate this approach, we introduce a novel dataset of over 1,100 annotated dialogues in English and Italian, designed for fine-tuning and assessing Multi-Modal models in HRI scenarios. Experimental results show that dialogue planning improves response accuracy and quality, and contributes to cross-lingual generalisation, enabling models trained in one language to transfer effectively to another. To the best of our knowledge, this is the first application of structured, goal-driven, and explicit dialogue planning in Multi-Modal LLMs for grounded interaction.

Sustainability reports are key instruments for communicating corporate impact, but their unstructured format and varied content pose challenges for large-scale analysis. This paper presents an unsupervised method to annotate paragraphs from sustainability reports against both the Global Reporting Initiative (GRI) and Sustainable Development Goals (SDG) standards. The approach combines structured metadata from GRI content indexes, official GRI–SDG mappings, and text semantic similarity models to produce weakly supervised annotations at scale. To evaluate the quality of these annotations, we train a multi-label classifier on the automatically labeled data and evaluate it on the trusted OSDG Community Dataset. The results show that our method yields meaningful labels and improves classification performance when combined with human-annotated data. Although preliminary, this work offers a foundation for scalable sustainability analysis and opens future directions toward assessing the credibility and depth of corporate sustainability claims.

This paper investigates the ability of Large Language Models (LLMs) to differentiate between canonical and non-canonical sentences in Italian, employing advanced neural architectures like LLaMA and its adaptations. Canonical sentences adhere to the standard Subject-Verb-Object (SVO) structure. We hypothesize that recent generative LLMs are influenced heavily by the English language, where non-canonical structures are very rare. Using the in-context learning technique, we probe these models and further fine-tune them for this specific task. Initial results indicate that these models continue to struggle with this task even after fine-tuning. Additionally, we introduce a new dataset comprising several hundred sentences from the poetry domain, which presents significant challenges for the canonical structure task.

This paper explores Interactive Grounded Language Understanding (IGLU) challenges within Human-Robot Interaction (HRI). In this setting, a robot interprets user commands related to its environment, aiming to discern whether a specific command can be executed. If faced with ambiguities or incomplete data, the robot poses relevant clarification questions. Drawing from the NeurIPS 2022 IGLU competition, we enrich the dataset by introducing our multi-modal data and natural language descriptions in MM-IGLU: Multi-Modal Interactive Grounded Language Understanding. Utilizing a BART-based model that integrates the user’s statement with the environment’s description, and a cutting-edge Multi-Modal Large Language Model that merges both visual and textual data, we offer a valuable resource for ongoing research in the domain. Additionally, we discuss the evaluation methods for such tasks, highlighting potential limitations imposed by traditional string-match-based evaluations on this intricate multi-modal challenge. Moreover, we provide an evaluation benchmark based on human judgment to address the limits and capabilities of such baseline models. This resource is released on a dedicated GitHub repository at https://github.com/crux82/MM-IGLU.

This paper introduces a novel framework to harness Large Language Models (LLMs) for Epidemic Intelligence, focusing on identifying and categorizing emergent socio-political phenomena within health crises, with a spotlight on the COVID-19 pandemic. Our approach diverges from traditional methods, such as Topic Models, by providing explicit support to analysts through the identification of distinct thematic areas and the generation of clear, actionable statements for each topic. This supports a Zero-shot Classification mechanism, enabling effective matching of news articles to fine-grain topics without the need for model fine-tuning. The framework is designed to be as transparent as possible, producing linguistically informed insights to make the analysis more accessible to analysts who may not be familiar with every subject matter of inherently emerging phenomena. This process not only enhances the precision and relevance of the extracted Epidemic Intelligence but also fosters a collaborative environment where system linguistic abilities and the analyst’s domain expertise are integrated.

Even if recent Transformer-based architectures, such as BERT, achieved impressive results in semantic processing tasks, their fine-tuning stage still requires large scale training resources. Usually, Data Augmentation (DA) techniques can help to deal with low resource settings. In Text Classification tasks, the objective of DA is the generation of well-formed sentences that i) represent the desired task category and ii) are novel with respect to existing sentences. In this paper, we propose a neural approach to automatically learn to generate new examples using a pre-trained sequence-to-sequence model. We first learn a task-oriented similarity function that we use to pair similar examples. Then, we use these example pairs to train a model to generate examples. Experiments in low resource settings show that augmenting the training material with the proposed strategy systematically improves the results on text classification and natural language inference tasks by up to 10% accuracy, outperforming existing DA approaches.

In real scenarios, a multilingual model trained to solve NLP tasks on a set of languages can be required to support new languages over time. Unfortunately, the straightforward retraining on a dataset containing annotated examples for all the languages is both expensive and time-consuming, especially when the number of target languages grows. Moreover, the original annotated material may no longer be available due to storage or business constraints. Re-training only with the new language data will inevitably result in Catastrophic Forgetting of previously acquired knowledge. We propose a Continual Learning strategy that updates a model to support new languages over time, while maintaining consistent results on previously learned languages. We define a Teacher-Student framework where the existing model “teaches” to a student model its knowledge about the languages it supports, while the student is also trained on a new language. We report an experimental evaluation in several tasks including Sentence Classification, Relational Learning and Sequence Labeling.

Recent Transformer-based architectures, e.g., BERT, provide impressive results in many Natural Language Processing tasks. However, most of the adopted benchmarks are made of (sometimes hundreds of) thousands of examples. In many real scenarios, obtaining high- quality annotated data is expensive and time consuming; in contrast, unlabeled examples characterizing the target task can be, in general, easily collected. One promising method to enable semi-supervised learning has been proposed in image processing, based on Semi- Supervised Generative Adversarial Networks. In this paper, we propose GAN-BERT that ex- tends the fine-tuning of BERT-like architectures with unlabeled data in a generative adversarial setting. Experimental results show that the requirement for annotated examples can be drastically reduced (up to only 50-100 annotated examples), still obtaining good performances in several sentence classification tasks.

While NLP systems become more pervasive, their accountability gains value as a focal point of effort. Epistemological opaqueness of nonlinear learning methods, such as deep learning models, can be a major drawback for their adoptions. In this paper, we discuss the application of Layerwise Relevance Propagation over a linguistically motivated neural architecture, the Kernel-based Deep Architecture, in order to trace back connections between linguistic properties of input instances and system decisions. Such connections then guide the construction of argumentations on network’s inferences, i.e., explanations based on real examples, semantically related to the input. We propose here a methodology to evaluate the transparency and coherence of analogy-based explanations modeling an audit stage for the system. Quantitative analysis on two semantic tasks, i.e., question classification and semantic role labeling, show that the explanatory capabilities (native in KDAs) are effective and they pave the way to more complex argumentation methods.

Nonlinear methods such as deep neural networks achieve state-of-the-art performances in several semantic NLP tasks. However epistemologically transparent decisions are not provided as for the limited interpretability of the underlying acquired neural models. In neural-based semantic inference tasks epistemological transparency corresponds to the ability of tracing back causal connections between the linguistic properties of a input instance and the produced classification output. In this paper, we propose the use of a methodology, called Layerwise Relevance Propagation, over linguistically motivated neural architectures, namely Kernel-based Deep Architectures (KDA), to guide argumentations and explanation inferences. In such a way, each decision provided by a KDA can be linked to real examples, linguistically related to the input instance: these can be used to motivate the network output. Quantitative analysis shows that richer explanations about the semantic and syntagmatic structures of the examples characterize more convincing arguments in two tasks, i.e. question classification and semantic role labeling.

Kernel methods enable the direct usage of structured representations of textual data during language learning and inference tasks. Expressive kernels, such as Tree Kernels, achieve excellent performance in NLP. On the other side, deep neural networks have been demonstrated effective in automatically learning feature representations during training. However, their input is tensor data, i.e., they can not manage rich structured information. In this paper, we show that expressive kernels and deep neural networks can be combined in a common framework in order to (i) explicitly model structured information and (ii) learn non-linear decision functions. We show that the input layer of a deep architecture can be pre-trained through the application of the Nystrom low-rank approximation of kernel spaces. The resulting “kernelized” neural network achieves state-of-the-art accuracy in three different tasks.

Service robots are expected to operate in specific environments, where the presence of humans plays a key role. A major feature of such robotics platforms is thus the ability to react to spoken commands. This requires the understanding of the user utterance with an accuracy able to trigger the robot reaction. Such correct interpretation of linguistic exchanges depends on physical, cognitive and language-dependent aspects related to the environment. In this work, we present the empirical evaluation of an adaptive Spoken Language Understanding chain for robotic commands, that explicitly depends on the operational environment during both the learning and recognition stages. The effectiveness of such a context-sensitive command interpretation is tested against an extension of an already existing corpus of commands, that introduced explicit perceptual knowledge: this enabled deeper measures proving that more accurate disambiguation capabilities can be actually obtained.

Sentiment Analysis systems aims at detecting opinions and sentiments that are expressed in texts. Many approaches in literature are based on resources that model the prior polarity of words or multi-word expressions, i.e. a polarity lexicon. Such resources are defined by teams of annotators, i.e. a manual annotation is provided to associate emotional or sentiment facets to the lexicon entries. The development of such lexicons is an expensive and language dependent process, making them often not covering all the linguistic sentiment phenomena. Moreover, once a lexicon is defined it can hardly be adopted in a different language or even a different domain. In this paper, we present several Distributional Polarity Lexicons (DPLs), i.e. large-scale polarity lexicons acquired with an unsupervised methodology based on Distributional Models of Lexical Semantics. Given a set of heuristically annotated sentences from Twitter, we transfer the sentiment information from sentences to words. The approach is mostly unsupervised, and experimental evaluations on Sentiment Analysis tasks in two languages show the benefits of the generated resources. The generated DPLs are publicly available in English and Italian.

Recent years show the development of large scale resources (e.g. FrameNet for the Frame Semantics) that supported the definition of several state-of-the-art approaches in Natural Language Processing. However, the reuse of existing resources in heterogeneous domains such as Human Robot Interaction is not straightforward. The generalization offered by many data driven methods is strongly biased by the employed data, whose performance in out-of-domain conditions exhibit large drops. In this paper, we present the Human Robot Interaction Corpus (HuRIC). It is made of audio files paired with their transcriptions referring to commands for a robot, e.g. in a home environment. The recorded sentences are annotated with different kinds of linguistic information, ranging from morphological and syntactic information to rich semantic information, according to the Frame Semantics, to characterize robot actions, and Spatial Semantics, to capture the robot environment. All texts are represented through the Abstract Meaning Representation, to adopt a simple but expressive representation of commands, that can be easily translated into the internal representation of the robot.

Lexical resources are basic components of many text processing system devoted to information extraction, question answering or dialogue. In paste years many resources have been developed such as FrameNet and WordNet. FrameNet describes prototypical situations (i.e. Frames) while WordNet defines lexical meaning (senses) for the majority of English nouns, verbs, adjectives and adverbs. A major difference between FrameNet and WordNet refers to their coverage. Due of this lack of coverage, in recent years some approaches have been studied to make a bridge between this two resources, so a resource is used to extend the coverage of the other one. The nature of these approaches leave from supervised to supervised methods. The major problem is that there is not a standard in evaluation of the mapping. Each different work have tested own approach with a custom gold standard. This work give an extensive evaluation of the model proposed in (De Cao et al., 2008) using gold standard proposed in other works. Moreover this work give an empirical comparison between other available resources. As outcome of this work we also release the full mapping resource made according to the model proposed in (De Cao et al., 2008).

In this paper, we present an original framework to model frame semantic resources (namely, FrameNet) using minimal supervision. This framework can be leveraged both to expand an existing FrameNet with new knowledge, and to induce a FrameNet in a new language. Our hypothesis is that a frame semantic resource can be modeled and represented by a suitable semantic space model. The intuition is that semantic spaces are an effective model of the notion of “being characteristic of a frame” for both lexical elements and full sentences. The paper gives two main contributions. First, it shows that our hypothesis is valid and can be successfully implemented. Second, it explores different types of semantic VSMs, outlining which one is more suitable for representing a frame semantic resource. In the paper, VSMs are used for modeling the linguistic core of a frame, the lexical units. Indeed, if the hypothesis is verified for these units, the proposed framework has a much wider application.

A critical step in Question Answering design is the definition of the models for question focus identification and answer extraction. In case of factoid questions, we can use a question classifier (trained according to a target taxonomy) and a named entity recognizer. Unfortunately, this latter cannot be applied to generate answers related to non-factoid questions. In this paper, we tackle such problem by designing classifiers of non-factoid answers. As the feature design for this learning task is very complex, we take advantage of tree kernels to generate large feature set from the syntactic parse trees of passages relevant to the target question. Such kernels encode syntactic and lexical information in Support Vector Machines which can decide if a sentence focuses on a target taxonomy subject. The experiments with SVMs on the TREC 10 dataset show that our approach is an interesting future research.

Different NLP applications have different efficiency constraints (i.e. quality of the results and throughput) that reflect on each core linguistic component. Syntactic processors are basic modules in some NLP application. A customization that permits the performance control of these components enables their reuse in different application scenarios. Throughput has been commonly improved using partial syntactic processors. On the other hand, specialized lexicons are generally employed to improve the quality of the syntactic material produced by specific parsing (sub)process (e.g. verb argument detection or PP attachment disambiguation) . Building upon the idea of grammar stratification, in this paper a method to push modularity and lexical sensitivity, in parsing, in view of customizable syntactic analysers is presented. A framework for modular parser design is proposed and its main properties are discussed. Parsers (i.e. different parsing module chains) are then presented and their performances are analyzed in an application-driven scenarios.