The Vector Grounding Problem

Confusingly, the notion of grounding is also used in relation to another aspect of language, which has to do with communication (Clark & Brennan 1991, Traum 1994). In this context, the ’grounding problem’ refers to the problem of establishing common ground between speakers in conversation:

In communication, common ground cannot be properly updated without a process we shall call grounding . . . In conversation, for example, the participants try to establish that what has been said has been understood. In our terminology, they try to ground what has been said – that is, make it part of their common ground. (Clark & Brennan 1991, p. 128)

From this perspective, linguistic communication can be seen as a form of coordinated action that involves collaborating to reach a common understanding of what is said. The process of communicative grounding may involve explicit clarification strategies, such as prompting an interlocutor to repeat or reformulate a corrupted or ambiguous statement. But it also involves tracking communicative intentions during the exchange, to infer what is communicated from what is said. Such inferences can be made partly on the basis of pragmatic presuppositions, namely what speakers take for granted or assume to be true when they use certain sentences in the context of the conversation (Stalnaker 2002).

The communicative notion of grounding has also been discussed by natural language processing (NLP) researchers (Brennan 1998, Di Maro 2021, Chandu et al. 2021). For example, Chandu et al. (2021) distinguish between static grounding, in which common ground in basic communication between a human user and a virtual agent is automatically established by querying a database (e.g., asking for a weather report); and dynamic grounding, in which common ground is more gradually established in back-and-forth communication between a human and a dialogue agent. Bender & Koller (2020) define meaning to be a relation between linguistic expressions and communicative intent, and characterise this relation as a form of grounding:

Without access to a means of hypothesizing and testing the underlying communicative intents, reconstructing them from the forms alone is hopeless, and [LLMs’] language use will eventually diverge from the language use of an agent who can ground their language in coherent communicative intents. (Bender & Koller 2020, p. 5189)

While communicative and referential grounding are distinct notions, they are related in important ways. Referential grounding anchors language to the world, but the same linguistic expressions may be referentially grounded in different ways for different speakers, due to differences in perception, knowledge, and conceptualisation. This means that calibrating reference in conversation requires communicative grounding—language users must work together to negotiate a common way of connecting language to the world. Without such collaboration, there is no guarantee that speakers really mean the same thing, even if they are using the same words.