ABSTRACT
Natural language processing (NLP) is a particular subfield of artificial intelligence (AI) with the aim of developing language systems which can process input and generate human-like linguistic behaviour. Its results are pervasive: from simple chatbots and large language models (LLMs), to explainable AI and generative tools such as text-to-image or video. Given their dominant advances and wide range of applications, the ubiquity of these phenomena will only continue to increase (Shanahan 1). Stemming from the conceptual tendency towards the stabilisation, compression, and discretisation of representations, “practical” NLP solutions often promote the elimination of semantic noise, as an obstacle impeding “accurate” representation. In the case of LLMs – the configurations of which are historically based on a lot of NLP solutions – statistically predictable linguistic coherence is the guiding parameter structuring their effects (ibid.). Because of this, conceptually, NLP and related language-formalising fields are ultimately geared towards the statistical schematisation of natural language (NL): the production of a fini-unlimited 2 combinatorial universe of linguistic structures through the amassment of sufficient examples. This tendency not only ignores foundational observations in linguistics and the philosophy of language, such as the fundamental social plasticity of language (which we will refer to as dialogical, following no particular thinker), but it also promotes conceptual stagnation because of the closure of NL it proposes. Diagnostically, this stagnant closure suffers the fate of one-dimensionality which characterised much of Herbert Marcuse’s critique of operational and functional visions of linguistic representation. For Marcuse, the concreteness of “[t]his language, which constantly imposes images, militates against the development and expression of concepts. In its immediacy and directness, it impedes conceptual thinking; thus, it impedes thinking” (98). We will not charge NLP with the impediment of thought altogether, but it is necessary to point out that the “image of language” 3 that it circulates is certainly one where meanings are statistically given, and not debatable and dynamic (i.e., thought). As will be argued later, in very general terms: statistical closure (e.g., frequentist retrieval from the past) is a different phenomenon than probabilistic (and possibilistic) consideration (e.g., hypothesis-based speculation). As proposed by Prado Casanova: “in the resolution of uncertainty that information entails, there must be the chance of a result completely ‘perturbed’ by noise” (8).
