Symbolic-neural learning involves deep learning methods in combination with symbolic structures. A “deep learning method” is taken to be a learning process based on gradient descent on real-valued model parameters. A “symbolic structure” is a data structure involving symbols drawn from a large vocabulary; for example, sentences of natural language, parse trees over such sentences, databases (with entities viewed as symbols), and the symbolic expressions of mathematical logic or computer programs.
Symbolic-neural learning has an innovative feature that allows to model interactions between different modals: speech, vision, and language. Such multimodal information processing is crucial for realizing research outcomes in real-word.
For growing needs and attention to multimodal research, SNL workshop this year features researches on “Beyond modality: Researches across speech, vision, and language boundaries.”
Topics of interests include, but are not limited to, the following areas:
- Speech, vision, and natural language interactions in robotics
- Multimodal and grounded language processing
- Multimodal QA and translation
- Dialogue systems
- Language as a mechanism to structure and reason about visual perception
- Image caption generation and image generation from text
- General knowledge question answering
- Reading comprehension
- Textual entailment
Deep learning systems across these areas share various architectural ideas. These include word and phrase embeddings, self-attention neural networks, recurrent neural networks (LSTMs and GRUs), and various memory mechanisms. Certain linguistic and semantic resources may also be relevant across these applications. For example, dictionaries, thesauri, WordNet, FrameNet, FreeBase, DBPedia, parsers, named entity recognizers, coreference systems, knowledge graphs and encyclopedias.