We are proud to host a two-day workshop on the 15th and 16th of August on End-to-End Compositional Models of Vector-Based Semantics during ESSLLI 2022!
Compositionality models the syntax-semantics interface as a structure-preserving map relating syntactic categories (types) and derivations to their counterparts in a corresponding meaning algebra. In a distributional setting, the basic building blocks are vector-based representations of word meanings (embeddings) obtained from data. These word meanings then have to be combined into meanings for larger expressions in a way that reflects the structure of their syntactic composition.
The workshop focuses on end-to-end implementations of such vector-based compositional architectures. This means not only the elementary word embeddings are obtained from data, but also the categories/types and their internal composition so that neural methods can then be applied to learn how the structure of syntactic derivations can be systematically mapped to operations on the data-driven word representations. For this last step, the workshop invites approaches that do not require the semantic operations to be linear maps since restricting the meaning algebra to finite dimensional vector spaces and linear maps means that vital information encoded in syntactic derivations may be lost in translation.
On the evaluation side, we welcome work on modern NLP tasks for evaluating sentence embeddings such as Natural Language Inference, sentence-level classification, and sentence disambiguation tasks. Special interest goes out to work that uses compositionality to investigate the syntactic sensitivity of large-scale language models.
Workshop contributions and invited talks will address the above challenges both from a theoretical and from a practical point of view.
The workshop welcomes but is not limited to contributions addressing the following topics:
- End-to-end models of compositional vector-based semantics
- Supervised and unsupervised models for wide-coverage supertagging and parsing
- Approaches to learning word/sentence representations
- Tasks and datasets requiring or benefiting from syntax
- Analysis of model performance on syntactically motivated tasks
- Multi-task learning/joint training of syntactic and semantic representations
- Using compositional methods to assess neural network behaviour
- Explainable models of sentence representation
The workshop will last for two days (9:00-15:30) and host a mix of invited and contributed talks. The schedule is as follows:
|09:30-10:30||Richard Moot (Invited Talk)
Perspectives on Neural Proof Nets
Neuro-Symbolic Proof Search for Linguistics
Challenges in Evaluating End-to-End Compositional Models: Some Test Cases in Dutch
Multi-type display calculus and algebraic semantics for modal extensions of Lambek calculus
|14:30-15:30||Mehrnoosh Sadrzadeh (Invited Talk)
Ambiguous Definite Pronoun Resolution via Lambek Calculus with Soft Sub-exponentials and Machine Learning
|09:30-10:30||Bob Coecke (Invited Talk)
A Tale of Four Disciplines for All Ages and All Languages
|11:00-11:30||Muhammad Hamza Waseem, Jonathon Liu, Vincent Wang & Bob Coecke
Language-independence of text circuits: English and Urdu (Zoom)
Grover’s Algorithm Continued: Implementation and Optimization of Word Representation Contraction
|12:00-12:30||Ido Benbaji, Omri Doron & Adèle Mortier
Word-embeddings distinguish denominal and root-derived verbs in Semitic
|14:00-14:30||Saba Nazir, Mehrnoosh Sadrzadeh & Stephen Clark
Grounding Compositional Distributional Adjective-Noun Composition in Audio Information (Zoom)
|14:30-15:00||Kin Ian Lo, Mehrnoosh Sadrzadeh & Shane Mansfield
An End-to-End Model of Anaphoric Ambiguities using Sheaf Theoretic Quantum Contextuality and BERT
|15:00-15:30||Jean-Philippe Bernardy & Shalom Lappin
Assessing the Unitary RNN as an End-to-End Compositional Model of Syntax
The workshop is funded by the research project ‘A composition calculus for vector-based semantic modelling with a localization for Dutch’ that will be in its final stage by the summer (Dutch Research Council NWO, 2017–2022). The project investigates the approach to compositionality outlined above with the objective of providing a collection of computational tools and resources for the compositional distributional study of Dutch.
Submissions consist of papers of up to 12 pages reporting on original work that has not been published or submitted elsewhere. Each submission will be refereed by at least two PC members. You can prepare your submission using LaTeX, using the EPTCS style, and upload the pdf to Easychair.
Accepted contributions have been published as a volume of Electronic Proceedings in Theoretical Computer Science, available online here!
In addition, a post-ESSLLI volume is planned with selected revised/expanded versions of workshop contributions together with reports on the results of the funding NWO project.
- 16 May 2022: Submission deadline (Extended until 23 May)
- 24 June 2022: Notification to authors
- 13 July 2022: Final copy due
- 15-16 August 2022: Workshop
- Mehrnoosh Sadrzadeh (UCL)
- Richard Moot (LIRMM)
- Bob Coecke (Cambridge Quantum Computing)
- Gemma Boleda, Universitat Pompeu Fabra
- Daisuke Bekki, Ochanomizu University
- Stergios Chatzikyriakides, University of Crete
- Stephen Clark, Cambridge Quantum
- Bob Coecke, Cambridge Quantum
- Giuseppe Greco, Vrije Universiteit Amsterdam
- Martha Lewis, Bristol University
- Michael Moortgat, Utrecht University (chair)
- Richard Moot, CNRS/LIRMM Montpellier
- Matthew Purver, Queen Mary University of Londen/Jožef Stefan Institute
- Mehrnoosh Sadrzadeh, University College of London
- Gijs Wijnholds, Utrecht University (chair)
Michael Moortgat, Gijs Wijnholds