Support for link feature recommendations

[jvdzwaan]

I'm trying to create a recommender for measurements. In our data model, we have layers Measurement and Measurementpart that are linked with a MeasurementPartLink. A measurement has a feature parts, which points to an array of MeasurementPartLinks.

Example:

w.c.Measurement(
    begin=3, 
    end=12, 
    parts=u.c.FSArray(
        elements=[
            webanno_custom_MeasurementPartsLink(
                xmiID=16159, 
                role='value', 
                target=webanno_custom_Measurementpart(
                    xmiID=16038, 
                    begin=17, 
                    end=19, 
                    type=Type(name=webanno.custom.Measurementpart)
                ), 
                type=Type(name=webanno.custom.MeasurementPartsLink)
            )
        ]
    ), 
) 

Is it possible to create a recommender that makes recommendations like this? If so, do you have an example of how to do this?

I created a a hard-coded example that seems to produce the correct data structure (no errors from inception), but it doesn't show up as recommendation in inception.

[reckart]

Span recommendations are best supported and relation recommendations are somewhat supported. I'm also looking into document-level annotation recommendations atm. However, link-feature recommendations are not supported at all atm. But it is a good point - it would be good to at least open an issue for supporting these in INCEpTION (won't be easy though - so unlikely something to be added soon).

[reckart]

However, in any case, the recommendation system is currently designed so that every recommender can only predict a single feature on a single layer.

You might be able to model this using multiple recommenders - one for the spans and another one for a relation between the spans. But the spans would need to be on the same layer (no cross-layer relations atm) and the relation would need to be a proper relation layer - not a link feature.

[jvdzwaan]

Thank you for the very fast response!

If I understand you correctly, we might be able to make it work by changing the data model from link based to relation based. I'll look into that.

I could also open an issue in INCEpTION for supporting link based recommendations, but if cross layer is not possible, I'm not sure how useful that would be. I'll start with doing a bit more research.

[reckart]

Link features can go across layers - but they are tricky to support in the recommender system atm.

Relations cannot go across layers atm.

[reckart]

@jvdzwaan I have moved the issue from the external recommender module to INCEpTION itself.

[jvdzwaan]

Thank you!

[reckart]

@jvdzwaan This will be included in v31. Note though that in order for a link recommendation to appear in the UI, the annotation that has the slot feature as well as the slot filler already need to be present as proper annotations in the document. Only then the link between the two will be shown as a suggestion that can be accepted.