The Limitations of Corpus-Based Methods in NLP

Nowadays, most Natural Language Processing software applications use empirical “black box” methods associated with training corpora to analyze texts written in natural languages. To analyze a sequence of text, they look for similar sequences in a corpus, select among them the most similar one according to some statistical measurement or some neural-network-based optimization state, and then bring forth its analysis as the new sequence analysis. Here, I first show that the limited size of the corpora used and their questionable quality explain why most NLP applications produce unreliable results. Next, I examine the principles which are at the basis of corpus-based methods and uncover their linguistic naiveté. I finally dispute the scientific validity of empirical approaches. I propose solutions to various problems that are based on the use of carefully handcrafted linguistic methods and resources.