Understanding Natural Language Understanding Systems. A Critical Analysis

Alessandro Lenci
2023-03-01
Abstract:The development of machines that «talk like us», also known as Natural Language Understanding (NLU) systems, is the Holy Grail of Artificial Intelligence (AI), since language is the quintessence of human intelligence. The brief but intense life of NLU research in AI and Natural Language Processing (NLP) is full of ups and downs, with periods of high hopes that the Grail is finally within reach, typically followed by phases of equally deep despair and disillusion. But never has the trust that we can build «talking machines» been stronger than the one engendered by the last generation of NLU systems. But is it gold all that glitters in AI? do state-of-the-art systems possess something comparable to the human knowledge of language? Are we at the dawn of a new era, in which the Grail is finally closer to us? In fact, the latest achievements of AI systems have sparkled, or better renewed, an intense scientific debate on their true language understanding capabilities. Some defend the idea that, yes, we are on the right track, despite the limits that computational models still show. Others are instead radically skeptic and even dismissal: The present limits are not just contingent and temporary problems of NLU systems, but the sign of the intrinsic inadequacy of the epistemological and technological paradigm grounding them. This paper aims at contributing to such debate by carrying out a critical analysis of the linguistic abilities of the most recent NLU systems. I contend that they incorporate important aspects of the way language is learnt and processed by humans, but at the same time they lack key interpretive and inferential skills that it is unlikely they can attain unless they are integrated with structured knowledge and the ability to exploit it for language use.
Artificial Intelligence,Computation and Language
What problem does this paper attempt to address?
The problems that this paper attempts to solve are: **Do current natural language understanding (NLU) systems truly possess human - like language understanding abilities, and in which aspects do they still have deficiencies?** Specifically, through a critical analysis of the most advanced NLU systems, the author explores the following key issues: 1. **Sources of "knowledge" in NLU systems and their representation methods**: - NLU systems obtain their "core knowledge" through self - supervised learning of large - scale texts, and this knowledge is stored in the internal state of the network in the form of continuous vectors. - This "knowledge" mainly comes from the co - occurrence statistics of language items, rather than a deep understanding of the semantic or grammatical structures of language items. 2. **Performance and limitations of NLU systems**: - Although NLU systems such as chatGPT perform impressively on multiple tasks, they still have many errors and limitations, especially in language processing and reasoning. - The paper points out that the "intelligent" behavior of NLU systems depends more on probabilistic pattern matching rather than true understanding. 3. **The gap between NLU systems and human language understanding**: - The author believes that NLU systems lack an understanding of the world outside of language, that is, they are unable to connect language with actual perception, action, and reasoning. - This view is called the "grounding argument" and emphasizes the limitations of NLU systems in acquiring and using language. 4. **Future development directions**: - The author proposes that in order to make NLU systems closer to human language understanding abilities, they need to be combined with multimodal data (such as images, videos, etc.) to make up for their deficiencies in perception and reasoning. - At the same time, the author also discusses the possibility of increasing the model scale and improving the pre - training method, but points out that this may be only a partial solution. In summary, this paper aims to explore how to further narrow the gap between machines and humans in language understanding by critically analyzing the current capabilities and limitations of NLU systems.