Dependency parsing with sparse binary hypervectors: seeking feedback

[Zhonghao (Kevin) Yang]

Hi VSA community,

I’d like to share an update on my work with sparse binary hypervectors
and invite early feedback from the community.

I maintain a Python module for sparse binary hypervectors:

https://github.com/yangzh/hv

Over the summer, I have been developing a proof-of-concept language
parser with VSA as its core representational and learning mechanism:

https://yangzh.github.io/hv/nlp/index.html

The parser is still an early prototype, so my immediate goal is not to
claim state-of-the-art parsing results. Instead, I am using dependency
parsing as a test case for a broader question:

Can we build a generic, robust learning system around VSA
operations/substrate that remains efficient, interpretable, and
compositional?

The current learning formulation is roughly related to a hierarchical
HMM-style framework implemented with sparse binary hypervectors. My
working hypothesis is that this approach could extend beyond parsing
to other structured-learning problems, particularly where
transparency, efficiency, and controllable representations matter more
than maximizing benchmark scores.

I would especially appreciate feedback on:
1. Whether the overall learning formulation seems promising or
fundamentally limited.
2. Which evaluation criteria and baselines would make the work most
informative to the VSA community.
3. Which problem domains beyond dependency parsing might be a stronger
or more compelling demonstration of a general VSA-based learning
system.
4. Existing VSA, symbolic, probabilistic, or structured-prediction
work that I should study or compare against.

The project is still evolving, but I hope the implementation and
write-up are concrete enough to start a useful discussion.
I would be grateful for any comments, critiques, relevant references,
or suggestions.

Thank you,
Kevin Yang