Tim Van de Cruys. A Linguist's Guide to Neural Networks

This course will provide an overview of the neural network paradigm to natural language processing, with a specific focus on linguistic applications. In recent years, neural network approaches have obtained strong performance on a wide range of different NLP tasks, using end-to-end neural architectures that do not rely on traditional, task-specific feature engineering. The course will provide an overview of the various neural architectures that exist, with a specific focus on current state of the art transformer-based architectures, and the methods that are used to train them. Moreover, the course will pay specific attention to linguistic aspects, viz. what linguistic information might be implicitly present within these models, and how can they be used for linguistic analysis. Each session will consist of a theoretical lecture, followed by a hands-on practical session.

Stephanie Evert and Gabriella Lapesa. Hands-on Distributional Semantics for Linguists using R 

Distributional semantic models (DSM) – also known as “word space”, “distributional similarity”, or more recently “word embeddings” – are based on the assumption that the meaning of a word can (at least to a certain extent) be inferred from its usage, i.e. its distribution in text. Therefore, these models dynamically build semantic representations – in the form of high-dimensional vector spaces – through a statistical analysis of the contexts in which words occur. DSMs are a promising technique for solving the lexical acquisition bottleneck by unsupervised learning, and their distributed representation provides a cognitively plausible, robust and flexible architecture for the organisation and processing of semantic information.

In this introductory course we will highlight the interdisciplinary potential of DSMs beyond standard semantic similarity tasks; our overview will put a strong focus on cognitive modeling and theoretical linguistics. This course aims to equip participants with the background knowledge and skills needed to build different kinds of DSM representations – from traditional “count” models to neural word embeddings – and apply them to a wide range of tasks. The hands-on sessions will be conducted in R.

Valerio Basile. High-quality Language Resources

Language Resources (LRs) play a key role in Natural Language Processing (NLP). Creating and maintaining corpora and lexica is a complex task that requires following several methodological steps to ensure the quality of the resources employed for training and benchmarking NLP models. The path to a successful LR is filled with critical design decisions, each impacting the final outcome, and potential pitfalls.

This course will provide an overview of the methodologies involved in designing annotation schemes, performing a robust annotation, representing and sharing the annotated data.

Throughout the course, a LR will be collectively created by the students by means of hands-on, interactive examples.

Jonathan Ginzburg and Andy Lücking. Multimodal interaction in Dialogue and its Meaning

After a brief introduction to the formal frameworks considered, we will discuss various multimodal phenomena (including pointing, laughter, headshake), motivating certain formal devices needed for their description and finally consider what this entails for the interface of semantics with theories of memory, emotion, and rapport.

Diego Frassinelli and Sabine Schulte Im Walde. Cognitive and Computational Models of Abstractness

Across disciplines, researchers are eager to gain insight into empirical features of abstract vs. concrete concepts and words. In the first part of this course we present an overview of the cognitive science literature which reports extensive analyses of how concrete concepts are processed, with however little consensus about the nature of abstract concepts. In the second part of this course we look into this dichotomy from a computational perspective, where the inclusion of information regarding the concreteness of words has been demonstrated to play a key role across NLP tasks, such as the automatic identification of figurative language. Additionally, we describe and discuss the procedures of collecting human-generated ratings of abstractness and their usage for both communities. Overall, this course thus aims at introducing and discussing cognitive and computational resources and empirical studies in order to understand the role and application of abstractness in large-scale data-driven models.

Gabriella Lapesa and Eva Maria Vecchi. Argument Mining between NLP and Social Sciences

Argument Mining is a highly interdisciplinary field in Natural Language Processing. Given a linguistic unit (a speech, an essay, forum post, or a tweet), its goal is to determine what is the position adopted by the author/speaker on a certain topic/issue (e.g., whether or not vaccinations should be enforced), and to identify the support, if any, provided by the speaker for its position. In this introductory course we discuss a selection of issues related to Argument Mining, structured along three main coordinates: the core notion of Argument Quality (How do we recognise good arguments?); the modeling challenges related to the automatic extraction of argument structures (multilingualism; evaluation of different modeling architectures); the application potential (computational social science; education). The course aims to highlight the interdisciplinary aspect of this field, ranging from the collaboration of theory and practice between NLP and social sciences, to approaching different types of linguistic structures (i.e., social media versus parliamentary texts), linguistic analysis of such structures, and the ethical issues involved (i.e., how to use Argument Mining for the social good).

James Pustejovsky and Nikhil Krishnaswamy. Multimodal Semantics for Affordances and Actions

This course introduces the requirements and challenges involved in developing a multimodal semantics for human-computer and human-robot interactions. Unlike unimodal interactive agents (e.g., text-based chatbots or voice-based personal digital assistants), multimodal HCI and HRI inherently require a notion of embodiment, or an understanding of the agent's placement within the environment and that of its interlocutor.

This requires not only the robust recognition and generation of expressions through multiple modalities (language, gesture, vision, action), but also the encoding of situated meaning: (a) the situated grounding of expressions in context; (b) an interpretation of the expression contextualized to the dynamics of the discourse; and (c) an appreciation of the actions and consequences associated with objects in the environment.

This in turn impacts how we computationally model human-human communicative interactions, with particular relevance to the shared understanding of affordances and actions over objects.

Dmitry Ustalov. Graphs, Computation, and Language

Employing properties of linguistic networks allows discovering structure and making predictions. This introductory course seeks answers to three questions: (1) how to express the linguistic phenomena as graphs, (2) how to gain knowledge based on them, and (3) how to assess the quality of this knowledge. We will start with traditional graph-based Natural Language Processing (NLP) methods like TextRank and Markov Clustering and finish with such contemporary Machine Learning techniques as DeepWalk and Graph Convolutional Networks. As the growing interest in NLP methods urges their meaningful evaluation, we pay special attention to quality assessment and human judgements. The course has five lectures on Language Graphs, Graph Clustering, Graph Embeddings, Evaluation, and Crowdsourcing. They elaborately go through the essential algorithms step-by-step, discuss case studies, and suggest insightful references and datasets. The target audience is undergraduate and graduate students, data analysts, and interdisciplinary researchers (but it is not limited to them).

Lasha Abzianidze. Natural Language Reasoning with a Natural Theorem Prover 

Natural language reasoning is a complex task that requires understanding the meanings of natural language expressions and recognizing semantic relations between them. The course is built around this highly interdisciplinary task. It introduces a computational theory of reasoning, called Natural Tableau, that combines the idea of a Natural Logic (logic using natural language as its vehicle of reasoning) with the semantic tableau method (a proof calculus that searches for certain situations).

The course not only introduces the theory of Natural Tableau but also shows its practical applications. In particular, we will show how an automated theorem prover, called LangPro, based on Natural Tableau, is used for Recognizing Textual Entailment (RTE) benchmarks. To overcome the knowledge sparsity and boost its performance, LangPro uses abductive reasoning to learn lexical relations from RTE training data.

Moreover, it will be also demonstrated how the prover can be extended for other languages than English, namely, for Dutch.

During the course, attendees will also have the opportunity to run LangPro on RTE problems and examine human-readable proofs.

Fausto Carcassi and Jakub Szymanik. Computational approaches to the explanation of universal properties of meaning

One of the most successful programs in semantics has been the identification of meaning universals. Recently, there has been a surge in research combining semantic and cognitive science to explain the origins of such universals. We will introduce this current and theoretically rich debate, taking this opportunity to teach computational methods to study language and cognition, e.g., learning models, minimal-description length, or information theory. We will start with an overview of the linguistic debates about universals. Then, we will focus on the universals that evolve from increasing learnability and reducing complexity. We will discuss iterated learning as a mechanism connecting learnability and language-level patterns. Next, we will present the pressure towards languages that are optimized for communication. Finally, we will discuss the universals emerging as a tradeoff between these pressures. Throughout, we will use convexity in nominal semantics and the universals of quantification as case studies.

Cristiano Chesi and Gregory M. Kobele. From minimal(ist) formalizations to parsing: pros and cons of a symbolic approach in a deep-learning era

A linguistic grammar formalism allows us to produce a deep analysis. The point of such an analysis is that disparate effects (word order, sentence meaning, prosodic contour, online processing cost, eye movements during reading, blood oxygenation levels in the brain, etc) are partially subsumed under a single cause (a syntactic structure). Among the many linguistic grammar formalisms on the market, in this course we will embrace a Minimalist perspective (Chomsky, 1995, 2001) because of its linguistic influence. We will start with the formalism of (Stabler, 2011, 1997) and we will show, on the one hand, how crucial empirical distinctions can be derived in terms of morphosyntactic featural manipulations, on the other, how various parsing strategies, based on Minimalist Grammars, can be formulated, which differ both in terms of complexity and cognitive plausibility. Here we will show how these different strategies make different predictions with respect to empirical data, and leverage these to reconstruct a version of MGs that better accounts for both linguistic (acceptability/grammaticality) and behavioural (fMRI/EEG/Eyetracking) data. In the end, a brief incursion on language modeling will be attempted, comparing the predictions of these models with the ones based on MGs.

Kyle Richardson and Gregor Betz. Argument and Logical Analysis in Humans and Machines

This is intended to be an advanced course that focuses on applied argumentation analysis both in humans and machines. More specifically, the goal is to cover the basics of normative models of text understanding and argumentation models as ordinarily studied in philosophy and to outline recent attempts to model argumentation dynamics in machines using state-of-the-art neural NLP (with a focus on recent pre-trained transformers).

David Traum. Computational Models of Grounding in Dialogue

Grounding is the process by which participants in a conversation establish new common ground. This process includes not just transmission of declarative utterances, but inferential and feedback processes. This process is also of critical importance to artificial dialogue systems, which include additional challenges of imperfect input recognition, and limited ontologies and inferential ability. In this course we will review models and uses for common ground in pragmatics and computational agent theories, and then examine a variety of proposals of how common ground can be established. These proposals include both descriptive analyses of behavior, as well as generative proposals that can be used by computational systems engaged in dialogue to decide on next moves. We will also look at multimodal grounding, and advanced topics, including multiparty grounding, incremental grounding and degrees of grounding, as well as how grounding models have been used for studying other social phenomena.

Gijs Wijnholds and Michael Moortgat. Compositional models of vector-based semantics: from theory to tractable implementation

Vector-based compositional architectures combine a distributional view of word meanings with a modelling of the syntax-semantics interface as a structure-preserving map relating syntactic categories (types) and derivations to their counterparts in a corresponding meaning algebra. This design is theoretically attractive, but faces challenges when it comes to large-scale practical applications. First there is the curse of dimensionality resulting from the fact that semantic spaces directly reflect the complexity of the types of the syntactic front end. Secondly, modelling of the meaning algebra in terms of finite dimensional vector spaces and linear maps means that vital information encoded in syntactic derivations is lost in translation. The course compares and evaluates methods that are being proposed to face these challenges. Participants gain a thorough understanding of theoretical and practical issues involved, and acquire hands-on experience witha set of user-friendly tools and resources.