Logic Programming: from NLP to NLU?

Natural Language Processing (NLP) was one of the original motivations leading to programming in logic [1] back in the seventies, with Colmerauer’s Metamorphosis Grammars [2] and with Pereira and D.H.D Warren’s Definite Clause Grammars [3], enhanced later with mechanisms for hypothetical reasoning [4]. Montague Grammars (implemented in Prolog by D.S. Warren [5]) and work by Veronica Dahl on defining logic representations for more realistic fragments of natural languages (including long distance dependencies and anaphora resolution) [6] have all shown a penchant of Logic Programming towards the higher objectives of Natural Language Understanding (NLU). After being long delayed (and partly frozen by the AI winter) recent progress in NLU promises to bring disruptive paradigm shifts in human-computer interaction and several directly and indirectly related industries. In fact, hopes for more logic-based NLU are high again, partly due to the possibility of sharing successful technologies and tools with successful fields like deep-learning neuralnetworks, machine learning, statistical parsers and graph-based NLP. This brings us to the obvious question: what new role can logic programming play in this new context? When trying to sketch an answer, after more than a decade spent on other research topics, we became aware that one can benefit today from the extended logic programming ecosystem, consisting of classic tools like Prolog, constraint solvers, Answer-Set Programming and SAT/SMT systems, as well as machine learning techniques implemented on top of inductive and probabilistic logic programming [7]. One of the research directions we have worked on in the past, that turned out to be remarkably successful, is graph-based NLP. It has originated in a Prolog program that was using WordNet’s semantic links to improve word-sense disambiguation (WSD) [8], by building a graph connecting words, sentences with their semantic equivalence classes (synsets) and then guessing the most likely sense associated to a word, by running the PageRank algorithm [9] on the graph. A few months later, an unsupervised version of the algorithm [10] based on graphs connecting sentences to word occurrences has been shown to extract high quality summaries and keywords. It later became one of the most popular techniques for the task [11], implemented in virtually all widely used