Large Language Models (LLMs) can generate highly persuasive text, raising concerns about their misuse for propaganda, manipulation, and other harmful purposes.This leads us to our central question: Is LLM-generated persuasion more difficult to automatically detect than human-written persuasion?To address this, we categorize controllable generation approaches for producing persuasive content with LLMs and introduce Persuaficial, a high-quality multilingual benchmark covering six languages: English, German, Polish, Italian, French and Russian.Using this benchmark, we conduct extensive empirical evaluations comparing human-authored and LLM-generated persuasive texts.We find that although overtly persuasive LLM-generated texts can be easier to detect than human-written ones, subtle LLMgenerated persuasion consistently degrades automatic detection performance.Beyond detection performance, we provide the first comprehensive linguistic analysis contrasting human and LLM-generated persuasive texts, offering insights that may guide the development of more interpretable and robust detection tools.

Nikolaos Nikolaidis, Nicolas Stefanovitch, Purificação Silvano, Dimitar Iliyanov Dimitrov, Roman Yangarber, Nuno Guimarães, Elisa Sartori, Ion Androutsopoulos, Preslav Nakov, Giovanni Da San Martino, Jakub Piskorski. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Arkadiusz Modzelewski, Witold Sosnowski, Tiziano Labruna, Adam Wierzbicki, Giovanni Da San Martino. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

This study presents a novel corpus of 15,356 Polish web articles, including articles identified as containing disinformation.Our dataset enables a multifaceted understanding of disinformation.We present a distinctive multilayered methodology for annotating disinformation in texts.What sets our corpus apart is its focus on uncovering hidden intent and manipulation in disinformative content.A team of experts annotated each article with multiple labels indicating both disinformation creators' intents and the manipulation techniques employed.Additionally, we set new baselines for binary disinformation detection and two multiclass multilabel classification tasks: manipulation techniques and intention types classification.

Jakub Piskorski, Nicolas Stefanovitch, Nikolaos Nikolaidis, Giovanni Da San Martino, Preslav Nakov. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

The COVID-19 pandemic has been at the center of the lives of many of us for at least a couple of years, during which periods of isolation and lockdowns were common. How all that affected our mental well-being, especially the ones’ who were already in distress? To investigate the matter we analyse the online discussions on Sanctioned Suicide, a forum where users discuss suicide-related topics freely. We collected discussions starting from March 2018 (before pandemic) up to July 2022, for a total of 53K threads with 700K comments and 16K users. We investigate the impact of COVID-19 on the discussions in the forum. The data show that covid, while being present in the discussions, especially during the first lockdown, has not been the main reason why new users registered to the forum. However, covid appears to be indirectly connected to other causes of distress for the users, i.e. anxiety for the economy.

Recent years have witnessed the proliferation of offensive content online such as fake news, propaganda, misinformation, and disinformation. While initially this was mostly about textual content, over time images and videos gained popularity, as they are much easier to consume, attract more attention, and spread further than text. As a result, researchers started leveraging different modalities and combinations thereof to tackle online multimodal offensive content. In this study, we offer a survey on the state-of-the-art on multimodal disinformation detection covering various combinations of modalities: text, images, speech, video, social media network structure, and temporal information. Moreover, while some studies focused on factuality, others investigated how harmful the content is. While these two components in the definition of disinformation – (i) factuality, and (ii) harmfulness –, are equally important, they are typically studied in isolation. Thus, we argue for the need to tackle disinformation detection by taking into account multiple modalities as well as both factuality and harmfulness, in the same framework. Finally, we discuss current challenges and future research directions.

The automatic identification of harmful content online is of major concern for social media platforms, policymakers, and society. Researchers have studied textual, visual, and audio content, but typically in isolation. Yet, harmful content often combines multiple modalities, as in the case of memes. With this in mind, here we offer a comprehensive survey with a focus on harmful memes. Based on a systematic analysis of recent literature, we first propose a new typology of harmful memes, and then we highlight and summarize the relevant state of the art. One interesting finding is that many types of harmful memes are not really studied, e.g., such featuring self-harm and extremism, partly due to the lack of suitable datasets. We further find that existing datasets mostly capture multi-class scenarios, which are not inclusive of the affective spectrum that memes can represent. Another observation is that memes can propagate globally through repackaging in different languages and that they can also be multilingual, blending different cultures. We conclude by highlighting several challenges related to multimodal semiotics, technological constraints, and non-trivial social engagement, and we present several open-ended aspects such as delineating online harm and empirically examining related frameworks and assistive interventions, which we believe will motivate and drive future research.

According to recent research, geometric deep learning allows to reach unprecedented accuracy for online misinformation detection. By fully leveraging the news social context, URL propagation paths in social networks are first represented as graphs and then classified using Graph Neural Network (GNN) models. Despite these remarkable efforts, researchers are still hampered by the scarcity of high-quality benchmark datasets, and as a result, the efficacy of state-of-the-art approaches could be overestimated. So far, in order to obtain a decent number of third-party fact-checked URLs, researchers have either sampled news from notoriously reliable and unreliable sources using distant supervision, or they have gathered pre-labeled URLs from third-party fact-checking websites. In the former case, resulting datasets can be quite large, but also noisy and biased since pieces of news are labeled as true or false according to their source label, and not individually fact-checked. In the latter case, assigned labels are more reliable, but the included news articles are usually in a single language and they may reflect unknown editorial decisions. As a result, datasets of the latter type are typically small, homogeneous, and thus unrealistically easy for automatic fake news detection models. In this work, we present FbMultiLingMisinfo, a new multilingual benchmark dataset, aimed at a more realistic evaluation of state-of-the-art misinformation detection models. URLs in our dataset come from the Facebook Privacy-Protected Full URLs Data Set, which we augmented with their propagation paths on Twitter. Our experimental results show that, when GNN-based models are tested on FbMultiLingMisinfo, recent misinformation detection results are only partially confirmed. We further show that a sharp reduction in the training size significantly reduces the model accuracy on FbMultiLingMisinfo, but not on two other widely used benchmark datasets for fake news detection.

Social media have democratized content creation and have made it easy for anybody to spread information online. However, stripping traditional media from their gate-keeping role has left the public unprotected against biased, deceptive and disinformative content, which could now travel online at breaking-news speed and influence major public events. For example, during the COVID-19 pandemic, a new blending of medical and political disinformation has given rise to the first global infodemic. We offer an overview of the emerging and inter-connected research areas of fact-checking, disinformation, "fake news'', propaganda, and media bias detection. We explore the general fact-checking pipeline and important elements thereof such as check-worthiness estimation, spotting previously fact-checked claims, stance detection, source reliability estimation, detection of persuasion techniques, and detecting malicious users in social media. We also cover large-scale pre-trained language models, and the challenges and opportunities they offer for generating and for defending against neural fake news. Finally, we discuss the ongoing COVID-19 infodemic.

Dimitar Dimitrov, Bishr Bin Ali, Shaden Shaar, Firoj Alam, Fabrizio Silvestri, Hamed Firooz, Preslav Nakov, Giovanni Da San Martino. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

The rise of Internet and social media changed not only how we consume information, but it also democratized the process of content creation and dissemination, thus making it easily available to anybody. Despite the hugely positive impact, this situation has the downside that the public was left unprotected against biased, deceptive, and disinformative content, which could now travel online at breaking-news speed and allegedly influence major events such as political elections, or disturb the efforts of governments and health officials to fight the ongoing COVID-19 pandemic. The research community responded to the issue, proposing a number of inter-connected research directions such as fact-checking, disinformation, misinformation, fake news, propaganda, and media bias detection. Below, we cover the mainstream research, and we also pay attention to less popular, but emerging research directions, such as propaganda detection, check-worthiness estimation, detecting previously fact-checked claims, and multimodality, which are of interest to human fact-checkers and journalists. We further cover relevant topics such as stance detection, source reliability estimation, detection of persuasion techniques in text and memes, and detecting malicious users in social media. Moreover, we discuss large-scale pre-trained language models, and the challenges and opportunities they offer for generating and for defending against neural fake news. Finally, we explore some recent efforts aiming at flattening the curve of the COVID-19 infodemic.

The reporting and the analysis of current events around the globe has expanded from professional, editor-lead journalism all the way to citizen journalism. Nowadays, politicians and other key players enjoy direct access to their audiences through social media, bypassing the filters of official cables or traditional media. However, the multiple advantages of free speech and direct communication are dimmed by the misuse of media to spread inaccurate or misleading claims. These phenomena have led to the modern incarnation of the fact-checker --- a professional whose main aim is to examine claims using available evidence and to assess their veracity. Here, we survey the available intelligent technologies that can support the human expert in the different steps of her fact-checking endeavor. These include identifying claims worth fact-checking, detecting relevant previously fact-checked claims, retrieving relevant evidence to fact-check a claim, and actually verifying a claim. In each case, we pay attention to the challenges and the potential impact on real-world fact-checking.

The rise of social media has democratized content creation and has made it easy for anybody to share and to spread information online. On the positive side, this has given rise to citizen journalism, thus enabling much faster dissemination of information compared to what was possible with newspapers, radio, and TV. On the negative side, stripping traditional media from their gate-keeping role has left the public unprotected against the spread of disinformation, which could now travel at breaking-news speed over the same democratic channel. This situation gave rise to the proliferation of false information, specifically created to affect individual people's beliefs, and ultimately to influence major events such as political elections; it also set the dawn of the Post-Truth Era, where appeal to emotions has become more important than the truth. More recently, with the emergence of the COVID-19 pandemic, a new blending of medical and political misinformation and disinformation has given rise to the first global infodemic. Limiting the impact of these negative developments has become a major focus for journalists, social media companies, and regulatory authorities.

The recent proliferation of "fake news" has triggered a number of responses, most notably the emergence of several manual fact-checking initiatives.As a result and over time, a large number of fact-checked claims have been accumulated, which increases the likelihood that a new claim in social media or a new statement by a politician might have already been factchecked by some trusted fact-checking organization, as viral claims often come back after a while in social media, and politicians like to repeat their favorite statements, true or false, over and over again.As manual fact-checking is very time-consuming (and fully automatic fact-checking has credibility issues), it is important to try to save this effort and to avoid wasting time on claims that have already been fact-checked.Interestingly, despite the importance of the task, it has been largely ignored by the research community so far.Here, we aim to bridge this gap.In particular, we formulate the task and we discuss how it relates to, but also differs from, previous work.We further create a specialized dataset, which we release to the research community.Finally, we present learning-to-rank experiments that demonstrate sizable improvements over state-of-the-art retrieval and textual similarity approaches.

We explore the task of predicting the leading political ideology or bias of news articles.First, we collect and release a large dataset of 34,737 articles that were manually annotated for political ideology -left, center, or right-, which is well-balanced across both topics and media.We further use a challenging experimental setup where the test examples come from media that were not seen during training, which prevents the model from learning to detect the source of the target news article instead of predicting its political ideology.From a modeling perspective, we propose an adversarial media adaptation, as well as a specially adapted triplet loss.We further add background information about the source, and we show that it is quite helpful for improving article-level prediction.Our experimental results show very sizable improvements over using state-of-the-art pre-trained Transformers in this challenging setup.

Propaganda campaigns aim at influencing people's mindset with the purpose of advancing a specific agenda. They exploit the anonymity of the Internet, the micro-profiling ability of social networks, and the ease of automatically creating and managing coordinated networks of accounts, to reach millions of social network users with persuasive messages, specifically targeted to topics each individual user is sensitive to, and ultimately influencing the outcome on a targeted issue. In this survey, we review the state of the art on computational propaganda detection from the perspective of Natural Language Processing and Network Analysis, arguing about the need for combined efforts between these communities. We further discuss current challenges and future research directions.

We present proppy, the first publicly available real-world, real-time propaganda detection system for online news, which aims at raising awareness, thus potentially limiting the impact of propaganda and helping fight disinformation. The system constantly monitors a number of news sources, deduplicates and clusters the news into events, and organizes the articles about an event on the basis of the likelihood that they contain propagandistic content. The system is trained on known propaganda sources using a variety of stylistic features. The evaluation results on a standard dataset show stateof-the-art results for propaganda detection.

Giovanni Da San Martino, Seunghak Yu, Alberto Barrón-Cedeño, Rostislav Petrov, Preslav Nakov. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Learning with limited resources (processing power and memory) on a stream of data is a challenging problem. When dealing with structured data, in particular with graphs, state-of- the-art graph kernels coupled with budget-aware online learning algorithms provide an efficient and effective solution. They map the input graph in a feature space that can be represented explicitly in sparse format. In this paper, we propose a method to enhance existing graph kernels in a streaming scenario with strict memory constraints. Specifically, we combine state-of-the-art online kernel methods with the power and flexibility of the feature representation of Extreme Learning Machines (ELM). Although being in principle a simple idea, our proposal, applied to several real-world datasets, outperformed state-of-the-art algorithms on streams of graphs with respect to predictive performance.

The availability of graph data with node attributes that can be either discrete or real-valued is constantly increasing. While existing Kernel methods are effective techniques for dealing with graphs having discrete node labels, their adaptation to nondiscrete or continuous node attributes has been limited, mainly for computational issues. Recently, a few kernels especially tailored for this domain, and that trade predictive performance for computational efficiency, have been proposed. In this brief, we propose a graph kernel for complex and continuous nodes' attributes, whose features are tree structures extracted from specific graph visits. The kernel manages to keep the same complexity of the state-of-the-art kernels while implicitly using a larger feature space. We further present an approximated variant of the kernel, which reduces its complexity significantly. Experimental results obtained on six real-world data sets show that the kernel is the best performing one on most of them. Moreover, in most cases, the approximated version reaches comparable performances to the current state-of-the-art kernels in terms of classification accuracy while greatly shortening the running times.

Several approaches have been proposed to model either the explicit sequential structure of an argumentative text or its implicit hierarchical structure.So far, the adequacy of these models of overall argumentation remains unclear.This paper asks what type of structure is actually important to tackle downstream tasks in computational argumentation.We analyze patterns in the overall argumentation of texts from three corpora.Then, we adapt the idea of positional tree kernels in order to capture sequential and hierarchical argumentative structure together for the first time.In systematic experiments for three text classification tasks, we find strong evidence for the impact of both types of structure.Our results suggest that either of them is necessary while their combination may be beneficial.

In many Information Retrieval tasks, the boundary between classes is not well defined, and assigning a document to a specific class may be complicated, even for humans. For instance, a document which is not directly related to the user's query may still contain relevant information. In this scenario, an option is to define an intermediate class collecting ambiguous instances. Yet some natural questions arise. Is this annotation strategy convenient? how should the intermediate class be treated? To answer these questions, we explored two community question answering datasets whose comments were originally annotated with three classes. We re-annotated a subset of instances considering a binary good vs bad setting. Our main contribution is to show empirically that the inclusion of an intermediate class to assess Boolean relevance is not useful. Moreover, in case the data is already annotated with a 3-class strategy, the instances from the intermediate class can be safely removed at training time.

We study how to find relevant questions in community forums when the language of the new questions is different from that of the existing questions in the forum. In particular, we explore the Arabic-English language pair. We compare a kernel-based system with a feed-forward neural network in a scenario where a large parallel corpus is available for training a machine translation system, bilingual dictionaries, and cross-language word embeddings. We observe that both approaches degrade the performance of the system when working on the translated text, especially the kernel-based system, which depends heavily on a syntactic kernel. We address this issue using a cross-language tree kernel, which compares the original Arabic tree to the English trees of the related questions. We show that this kernel almost closes the performance gap with respect to the monolingual system. On the neural network side, we use the parallel corpus to train cross-language embeddings, which we then use to represent the Arabic input and the English related questions in the same space. The results also improve to close to those of the monolingual neural network. Overall, the kernel system shows a better performance compared to the neural network in all cases.

We study the impact of different types of features for question ranking in community Question Answering: bag-of-words models (BoW), syntactic tree kernels (TKs) and rank features. It should be noted that structural kernels have never been applied to the question reranking task, i.e., question to question similarity, where they have to model paraphrase relations. Additionally, the informal text, typically present in forums, poses new challenges to the use of TKs. We compare our learning to rank (L2R) algorithms against a strong baseline given by the Google rank (GR). The results show that (i) our shallow structures used in TKs are robust enough to noisy data and (ii) improving GR requires effective BoW features and TKs along with an accurate model of GR features in the used L2R algorithm.

Community question answering (cQA) websites are focused on users who query questions onto an online forum, expecting for other users to provide them answers or suggestions. Unlike other social media, the length of the posted queries has no limits and queries tend to be multi-sentence elaborations combining context, actual questions, and irrelevant information. We approach the problem of question ranking: given a user’s new question, to retrieve those previously-posted questions which could be equivalent, or highly relevant. This could prevent the posting of nearly-duplicate questions and provide the user with instantaneous answers. For the first time in cQA, we address the selection of relevant text —both at sentence- and at constituent-level— for parse tree-based representations. Our supervised models for text selection boost the performance of a tree kernel-based machine learning model, allowing it to overtake the current state of the art on a recently released cQA evaluation framework.

In real-world data, e.g., from Web forums, text is often contaminated with redundant or irrelevant content, which leads to introducing noise in machine learning algorithms. In this paper, we apply Long Short-Term Memory networks with an attention mechanism, which can select important parts of text for the task of similar question retrieval from community Question Answering (cQA) forums. In particular, we use the attention weights for both selecting entire sentences and their subparts, i.e., word/chunk, from shallow syntactic trees. More interestingly, we apply tree kernels to the filtered text representations, thus exploiting the implicit features of the subtree space for learning question reranking. Our results show that the attention-based pruning allows for achieving the top position in the cQA challenge of SemEval 2016, with a relatively large gap from the other participants while greatly decreasing running time.

In recent years there has been a growing interest in text quantification, a supervised learning task where the goal is to accurately estimate, in an unlabelled set of items, the prevalence (or "relative frequency") of each class c in a predefined set C. Text quantification has several applications, and is a dominant concern in fields such as market research, the social sciences, political science, and epidemiology. In this paper we tackle, for the first time, the problem of ordinal text quantification, defined as the task of performing text quantification when a total order is defined on the set of classes; estimating the prevalence of "five stars" reviews in a set of reviews of a given product, and monitoring this prevalence across time, is an example application. We present OQT, a novel tree-based OQ algorithm, and discuss experimental results obtained on a dataset of tweets classified according to sentiment strength.

Simone Filice, Giovanni Da San Martino, Alessandro Moschitti. Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2015.

Shafiq Joty, Alberto Barrón-Cedeño, Giovanni Da San Martino, Simone Filice, Lluís Màrquez, Alessandro Moschitti, Preslav Nakov. Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing. 2015.

Tree kernels proposed in the literature rarely use information about the relative location of the substructures within a tree. As this type of information is orthogonal to the one commonly exploited by tree kernels, the two can be combined to enhance state-of-the-art accuracy of tree kernels. In this brief, our attention is focused on subtree kernels. We describe an efficient algorithm for injecting positional information into a tree kernel and present ways to enlarge its feature space without affecting its worst case complexity. The experimental results on several benchmark datasets are presented showing that our method is able to reach state-of-the-art performances, obtaining in some cases better performance than computationally more demanding tree kernels.

In this paper, we show how learning models generated by a recently introduced state-of-the-art kernel for graphs can be optimized from the point of view of memory occupancy. After a brief description of the kernel, we introduce a novel representation of the explicit feature space of the kernel based on an hash function which allows to reduce the amount of memory needed both during the training phase and to represent the final learned model. Subsequently, we study the application of a feature selection strategy based on the F-score to further reduce the number of features in the final model. On two representative datasets involving binary classification of chemical graphs, we show that it is actually possible to sensibly reduce memory occupancy (up to one order of magnitude) for the final model with a moderate loss in classification performance.

This paper proposes a new tree-based kernel for graphs. Graphs are decomposed into multisets of ordered Directed Acyclic Graphs (DAGs) and a family of kernels computed by application of tree kernels extended to the DAG domain. We focus our attention on the efficient development of one member of this family. A technique for speeding up the computation is given, as well as theoretical bounds and practical evidence of its feasibility. State of the art results on various benchmark datasets prove the effectiveness of our approach.

BACKGROUND: The identification of robust lists of molecular biomarkers related to a disease is a fundamental step for early diagnosis and treatment. However, methodologies for biomarker discovery using microarray data often provide results with limited overlap. It has been suggested that one reason for these inconsistencies may be that in complex diseases, such as cancer, multiple genes belonging to one or more physiological pathways are associated with the outcomes. Thus, a possible approach to improve list stability is to integrate biological information from genomic databases in the learning process; however, a comprehensive assessment based on different types of biological information is still lacking in the literature. In this work we have compared the effect of using different biological information in the learning process like functional annotations, protein-protein interactions and expression correlation among genes. RESULTS: Biological knowledge has been codified by means of gene similarity matrices and expression data linearly transformed in such a way that the more similar two features are, the more closely they are mapped. Two semantic similarity matrices, based on Biological Process and Molecular Function Gene Ontology annotation, and geodesic distance applied on protein-protein interaction networks, are the best performers in improving list stability maintaining almost equal prediction accuracy. CONCLUSIONS: The performed analysis supports the idea that when some features are strongly correlated to each other, for example because are close in the protein-protein interaction network, then they might have similar importance and are equally relevant for the task at hand. Obtained results can be a starting point for additional experiments on combining similarity matrices in order to obtain even more stable lists of biomarkers. The implementation of the classification algorithm is available at the link: http://www.math.unipd.it/~dasan/biomarkers.html.

Almost all tree kernels proposed in the literature match substructures without taking into account their relative positioning with respect to one another. In this paper, we propose a novel family of kernels which explicitly focus on this type of information. Specifically, after defining a family of tree kernels based on routes between nodes, we present an efficient implementation for a member of this family. Experimental results on four different datasets show that our method is able to reach state of the art performances, obtaining in some cases performances better than computationally more demanding tree kernels.

The development of neural network (NN) models able to encode structured input, and the more recent definition of kernels for structures, makes it possible to directly apply machine learning approaches to generic structured data. However, the effectiveness of a kernel can depend on its sparsity with respect to a specific data set. In fact, the accuracy of a kernel method typically reduces as the kernel sparsity increases. The sparsity problem is particularly common in structured domains involving discrete variables which may take on many different values. In this paper, we explore this issue on two well-known kernels for trees, and propose to face it by recurring to self-organizing maps (SOMs) for structures. Specifically, we show that a suitable combination of the two approaches, obtained by defining a new class of kernels based on the activation map of a SOM for structures, can be effective in avoiding the sparsity problem and results in a system that can be significantly more accurate for categorization tasks on structured data. The effectiveness of the proposed approach is demonstrated experimentally on two relatively large corpora of XML formatted data and a data set of user sessions extracted from website logs.

Kernel methods are effective approaches to the modeling of structured objects in learning algorithms. Their major drawback is the typically high computational complexity of kernel functions. This prevents the application of computational demanding algorithms, e.g. support vector machines, on large datasets. Consequently, on-line learning approaches are required. Moreover, to facilitate the application of kernel methods on structured data, additional efficiency optimization should be carried out. In this paper, we propose direct acyclic graphs to reduce the computational burden and storage requirements by representing common structures and feature vectors. We show the benefit of our approach for the perceptron algorithm using tree and polynomial kernels. The experiments on a quite extensive dataset of about one million of instances show that our model makes the use of kernels for trees practical. From the accuracy point of view, the possibility of using large amount of data has allowed us to reach the state-of-the-art on the automatic detection of semantic role labeling as defined in the conference on natural language learning shared task

Kernel methods have been shown to be very effective for applications requiring the modeling of structured objects. However kernels for structures usually are too computational demanding to be applied to complex learning algorithms, e.g. Support Vector Machines. Consequently, in order to apply kernels to large amount of structured data, we need fast on-line algorithms along with an efficiency optimization of kernel-based computations. In this paper, we optimize this computation by representing set of trees by minimal Direct Acyclic Graphs (DAGs) allowing us i) to reduce the storage requirements and ii) to speed up the evaluation on large number of trees as it can be done 'one-shot' by computing kernels over DAGs. The experiments on predicate argument subtrees from PropBank data show that substantial computational savings can be obtained for the perceptron algorithm.