Semiotics for NLP Practitioners

Every now and then, when I take part in scientific conferences I seem to notice a single issue many of the speakers exhibit when presenting their work. It’s a minor thing, but it’s been increasingly common as the language models get more and more powerful. The thing I’m talking about is the lack of proper terminology when describing how an NLP model handles implicit information, or context, as described by some.

Many speakers struggle to explain that their models are somewhat capable of recognizing information that’s not explicitly stated (for example in an input sentence). That’s a pity, because there exists a very convenient theory, based in linguistics, that could help us get on a common ground when describing this elusive context thing. I’m talking about Semiotics as it was defined by Charles Morris. Semiotics is the science of relationships between the signs (e.g., letters) and the viewer. From this short explanation we can induce that Semiotics is related to linguistics, and, by extension, to NLP.

Morris introduced a very useful division of Semiotics into three dimensions:

  1. Syntax – how different signs relate to each other, how can they be arranged.
  2. Semantics – the exact meaning of signs, i.e., what concepts are represented by a sign.
  3. Pragmatics – the relationship between the signs and the user, i.e., what does the sign express for that specific user.

We have successfully solved the problems of Syntax and Semantics in NLP. We have language models that know how words should be arranged, we know what are the meaning of individual words and sentences. Now, I think that solving Pragmatics is the goal of new and upcoming NLP models. We would like to model the implicit knowledge and context carried by a sentence, while not being stated directly.

Therefore I believe that using the terminology introduced by Morris would help describing our end goal as NLP practitioners, so instead of speaking vaguely how our new models understand implicit knowledge, we can just say that the model captures Pragmatics well.