Using gold POS tags to influence Dependency Parsing at inference time

[nunzianteantonio]

Hi everyone,

I’m working with spaCy v3.x on Italian product titles and I have trained a custom POS tagger and dependency parser on a domain-specific dataset.

In my use case, I sometimes need to override or correct POS tags for specific tokens (e.g., distinguishing between NOUN/ADJ and VERB in ambiguous product titles). I would like the dependency tree to be computed or "steered" taking these corrected ("gold") POS tags into account.

I am aware that in spaCy's standard architecture, the DependencyParser and Tagger are typically independent components that both sit on top of the Tok2Vec or Transformer listener.

My questions are:

1. Is there a supported way to make the DependencyParser "aware" of the token.pos_ values if I manually set them before the parser runs in the pipeline?
2. If the parser doesn't natively look at the POS attribute during inference, is there a recommended strategy (e.g., a custom architectural change or a specific training approach) to enforce consistency between corrected POS tags and the resulting parse tree?
3. Specifically for Italian, where morphological ambiguity is high, has anyone successfully implemented a "POS-aware" parser in spaCy?

I’d appreciate any guidance or confirmation on whether this is supported or if I should look into alternative architectures (like Stanza's pretagged mode).

Thank you for your time and for this amazing library!