Séminaire Alpage : Héctor Martinez Alonso

Héctor Martinez Alonso (Inria, Alpage)
Annotation of Regular Polysemy

The ability of certain classes of words to switch between two readings in a systematic fashion is called regular polysemy, e.g. regular polysemy is the phenomenon that describes why Locations can systematically mean Organizations, or Processes can be interpreted as Results. 

Theory in lexical semantics (cf. Pustejovsky, 95) has postulated that words that experience regular polysemy can present an underspecified sense were both readings are equally active, such as "England is rainy and organized", where each adjective selects for a certain reading, and the whole sentence presents both. However, the examples offered in the theory are often short, synthetic examples detached from actual language use.

The main goal of this work is to empirically assess whether the underspecified sense needs to be incorporated in sense lexica when trying to annotate or recognize regular polysemy. We deal with the human and automatic recognition of the underspecified sense for a series of nominal classes in English, Danish and Spanish. We find enough support for the unsupervised sense for human annotation, provided the annotators are not crowdsourced. However, automatic recognition of the underspecified sense fares poorly.

Moreover, we address the issues of annotation bias of when conducting research in lexical semantics, as different kinds of annotators show different behavior when annotating the same examples. Finally, we propose an alternative, continuous representation for regular polysemy.

In addition, the talk presents the VerDi project, dealing with veracity and omission in text.