DMR 2023

While deep learning methods have led to many breakthroughs in practical natural language applications, most notably in Machine Translation, Machine Reading, Question Answering, Recognizing Textual Entailment, and so on, there is still a sense among many NLP researchers that we have a long way to go before we can develop systems that can actually “understand” human language and explain the decisions they make. Indeed, “understanding” natural language entails many different human-like capabilities, and they include but are not limited to the ability to track entities in a text, understand the relations between these entities, track events and their participants described in a text, understand how events unfold in time, and distinguish events that have actually happened from events that are planned or intended, are uncertain, or did not happen at all. We believe a critical step in achieving natural language understanding is to design meaning representations for text that have the necessary meaning “ingredients” that help us achieve these capabilities. Such meaning representations can also potentially be used to evaluate the compositional generalization capacity of deep learning models.

There has been a growing body of research devoted to the design, annotation, and parsing of meaning representations in recent years. The meaning representations that have been used for semantic parsing research are developed with different linguistic perspectives and practical goals in mind and have different formal properties. Formal meaning representation frameworks such as Minimal Recursion Semantics (MRS) and Discourse Representation Theory (as exemplified in the Parallel Meaning Bank) are developed with the goal of supporting logical inference in reasoning-based AI systems and are therefore easily translatable into first-order logic, requiring proper representation of semantic components such as quantification, negation, tense, and modality. Other meaning representation frameworks such as Abstract Meaning Representation (AMR), Tecto-grammatical Representation (TR) in Prague Dependency Treebanks and the Universal Conceptual Cognitive Annotation (UCCA), put more emphasis on the representation of core predicate-argument structure, lexical semantic information such as semantic roles and word senses, or named entities and relations. There is also a more recent effort in developing a Uniform Meaning Representation (UMR) that is based on AMR but extends it to cross-linguistic settings and enhances it to represent document-level semantic content. The automatic parsing of natural language text into these meaning representations and the generation of natural language text from these meaning representations are also very active areas of research, and a wide range of technical approaches and learning methods have been applied to these problems.

This workshop will bring together researchers who are producers and consumers of meaning representations, and through their interaction develop a deeper understanding of the key elements of meaning representations that are the most valuable to the NLP community. The workshop will also provide an opportunity for meaning representation researchers to critically examine existing frameworks with the goal of using their findings to inform the design of next-generation meaning representations. A third goal of the workshop is to explore opportunities and identify challenges in the design and use of meaning representations in multilingual settings. A final goal of the workshop is to understand the relationship between distributed meaning representations trained on large data sets using network models, and the symbolic meaning representations that are carefully designed and annotated by NLP researchers and gain a deeper understanding of areas where each type of meaning representation is the most effective.

The workshop solicits papers that address one or more of the following topics:

• Design and annotation of meaning representations;
• Cross-framework comparison of meaning representations;
• Challenges and techniques in automatic parsing of meaning representations;
• Challenges and techniques in automatically generating text from meaning representations;
• Meaning representation evaluation metrics;
• Lexical resources, ontologies, and grounding in relation to meaning representations;
• Real-world applications of meaning representations;
• Issues in applying meaning representations to multilingual settings and lower-resourced languages;
• The relationship between symbolic meaning representations and distributed semantic representations;
• Formal properties of meaning representations;
• Any other topics that address the design, processing, and use of meaning representations.