Multiple 'REL_OP' options?

[ninjalu]

I am using dependency matcher to match close questions, like
Are you....? etc.

For it to work, I need to make sure that 'are' is a parent of 'you' and also precedes 'you'.

At the moment I can use one 'LEFT_ID' and 'RIGHT_ID' pair to define the 'parenthood', as follows:

        {
            'LEFT_ID': 'anchor_aux',
            'REL_OP': '>',
            'RIGHT_ID': 'nsubj',
            'RIGHT_ATTRS': {
                'DEP': 'nsubj'
            }
        },

And another pair to define the order, as follows:

        {
            'LEFT_ID': 'anchor_aux',
            'REL_OP': '.*',
            'RIGHT_ID': 'nsubj2',
            'RIGHT_ATTRS': {
                'DEP': 'nsubj'
            }
        },

However this will become a problem if somewhere else in the sentence there is another nsubj.

Is it possible to define the relationships between these two nodes BOTH '>' and '.*' at the same time?

[polm]

You can't pass multiple REL_OPs to the DependencyParser for a single token.

In this case I think .* does what you want? It's true it doesn't define which token is the parent of the other, but given your left token is a verb and your right token is nsubj I don't think those relations can be reversed meaningfully. If you define the verb as being the ROOT reversal is also impossible.

Another thing is that if you are actually looking just for Are you type patterns, honestly maybe you don't need the dependency matcher? I can imagine variations ("Are all of you"), but it seems like just matching the literal tokens will probably cover the vast majority of cases - you might want to try that if you haven't.

[ninjalu]

Maybe the easiest thing would be adding a new REL_OP?

That's not a bad idea! Thanks I will read the source code and try to figure out how to do that.

[ninjalu]

I have added four additional "REL_OP": >.* for right child, >;* left child, <.* right parent, and <;* left parent to dependencymatcher.pyx. It seems to work as a treat for me,
Would you consider merging the changes, if I push it up? I could have a local version of spaCy, but I don't want to miss out on future improvement from your side.

Let me know what's the best way forward. Many thanks!

the changes are, in DependencyMatcher:

        self._ops = {
            "<": self._dep,
            ">": self._gov,
            "<<": self._dep_chain,
            ">>": self._gov_chain,
            ".": self._imm_precede,
            ".*": self._precede,
            ";": self._imm_follow,
            ";*": self._follow,
            "$+": self._imm_right_sib,
            "$-": self._imm_left_sib,
            "$++": self._right_sib,
            "$--": self._left_sib,
# New REL_OP as below:
            ">.*": self._right_child,
            ">;*": self._left_child,
            "<.*": self._right_parent,
            "<;*": self._left_parent,
        }

And additional of these

    def _right_child(self, doc, node):
        return [doc[child.i] for child in doc[node].children if child.i > node]
    
    def _left_child(self, doc, node):
        return [doc[child.i] for child in doc[node].children if child.i < node]

    def _right_parent(self, doc, node):
        if doc[node].head == doc[node]:
            return []
        elif doc[node].head.i > node:
            return [doc[node].head]
        return []
    
    def _left_parent(self, doc, node):
        if doc[node].head == doc[node]:
            return []
        elif doc[node].head.i < node:
            return [doc[node].head]
        return []

[polm]

Oh wow, thanks! That actually does look pretty simple.

If you could go ahead an open a PR that'd make it easy for the whole team to take a look at it.

[ninjalu]

Thanks a lot! I have created PR #10371 with the above changes.

Never attempted to contribute to open source before! It's a very small change, and not sure whether it will get merged. Still I am buzzing :D

[polm]

Thanks for making the PR, and congrats on your first! The team will take a look at it and get back to you.