My invited online talk will be on “Referential search”.
Looking for a “band”? Does it matter what or which “band” you find? Traditionally, search is seen as a conceptual problem. Query words are mapped to information words, solving problems of ambiguity, variation and relevance. But language is first of all a reference system that hardly ever describes things fully. We use words and expressions assuming that we resolve these as references: mentions of instances. Conceptually “band” can mean different things and if it is a music band there can be many styles . Some of these bands you can rent for different purposes and for a specific price. Still you probably have specific band in mind and whatever instance of a band you find, it will be another specific instance of a (musical) “band”. Maybe you do not like this one for any particular reason or maybe this one is booked.
We complete our mental picture by (shared) background knowledge. Our descriptions are incomplete and we fill gaps and relations that are not mentioned. There are many pragmatic factors that determine the form of our expressions to make reference, thus resulting in variation that needs to be modelled by systems. What does that mean for a search system that matches requests to information and services described through language?
In this presentation, I will discuss how natural language search can be modeled as a referential model. In such a model, systems need to count and quantify across instances that are conceptually similar but differ at an instance level. A simple case is an e-commerce system that offers these instances for sale but describes them though conceptual categories. I will however also discuss the problem of instance-identity & reference in more complex cases such as the 2018 Semeval task on Counting events. I will conclude that referential search also provides great opportunities in a future Internet of Things, where multimodal devices without display, mouse or keyboard also represent a referential world with which we naturally interact through referential language. Interaction between people and systems in an e-commerce or service setting represents a complex abstract retrieval problem, when we can only using language to make reference without a direct representation or grounding of the instances.