This work shop focuses on end-to-end implementation s of vector-based compositional architectures. Thi s means not only the elementary word embeddings ar e obtained from data, but also the categories/type s and their internal composition so that neural me thods can then be applied to learn how the structu re of syntactic derivations can be systematically mapped to operations on the data-driven word repre sentations. For this last step, the workshop invit es approaches that do not require the semantic ope rations to be linear maps since restricting the me aning algebra to finite dimensional vector spaces and linear maps means that vital information encod ed in syntactic derivations may be lost in transla tion.
\n\nOn the evaluation side, we welco me work on modern NLP tasks for evaluating sentenc e embeddings such as Natural Language Inference, s entence-level classification, and sentence disambi guation tasks. Special interest goes out to work t hat uses compositionality to investigate the synta ctic sensitivity of large-scale language models.\n
The workshop welcomes but is
not limited to contributions addressing the follo
wing topics:
\n - End-to-end models of composi
tional vector-based semantics
\n - Supervised
and unsupervised models for wide-coverage supertag
ging and parsing
\n - Approaches to learning w
ord/sentence representations
\n - Tasks and da
tasets requiring or benefiting from syntax
\n
- Analysis of model performance on syntactically m
otivated tasks
\n - Multi-task learning/joint
training of syntactic and semantic representations
\n - Using compositional methods to assess ne
ural network behaviour
\n - Explainable models
of sentence representation