31 N1 Relations

Mismatch between the number of N-1 relations in LAMA and in our paper

Although the LAMA paper reports only 25 N-1 relations, we treat 31 relations as N-1. We explain this mismatch here:

The reason for the mismatch is that we looked at the pattern definition, as well as the data itself, and found that some of the relations that were classified as N-M, were actually N-1. This is typically a result of how most of the data is used in practice, as an N-1 relation (subset).

For instance, the P178, developer relation is classified as N-M in LAMA. This is a correct classification in the general case, however, looking closely at the data at hand reveals that it is practically being used as an N-1 relation.

This will not necessarily hold in other applications (e.g. classifying for a relation type), but considering only the candidate set - all potential objects from a specific relation - I believe our updated mapping is accurate.

Patterns used in our study

The full list of N-1 relations we used is the following:

['P937', 'P1412', 'P127', 'P103', 'P276', 'P159', 'P140', 'P136', 'P495', 'P17', 'P361', 'P36', 'P740', 'P264', 'P407', 'P138', 'P30', 'P131', 'P176', 'P449', 'P279', 'P19', 'P101', 'P364', 'P106', 'P1376', 'P178', 'P37', 'P413', 'P27', 'P20']

The N-M relations we used are:

['P190', 'P1303', 'P39', 'P108', 'P463', 'P530', 'P47']