From Natural Language to RDF Graphs with Pregroups

Abstract

We define an algorithm translating natural language sentences to the formal syntax of RDF, an existential conjunctive logic widely used on the Semantic Web. Our translation is based on pregroup grammars, an efficient type-logical grammatical framework with a transparent syntax-semantics interface. We introduce a restricted notion of side effects in the semantic category of finitely generated free semimodules over {0, 1} to that end. The translation gives an intensional counterpart to previous extensional models. We establish a one-to-one correspondence between extensional models and RDF models such that satisfaction is preserved. Our translation encompasses the expressivity of the target language and supports complex linguistic constructions like relative clauses and unbounded dependencies.

Towards Compositional Distributional Discourse Analysis

“…The translation from pregroup grammars to RDF triples is developed further in [10], in order to model count nouns (e.g. "a cat") and relative pronouns (e.g.…”

“…}. In Semantic Web languages, relations are always binary so we will focus on transitive verbs, for a translation of adjectives and intransitive verbs in RDF see [10]. We encode every word as a binary code with one bit up, also called a one-hot vector -both entities e i ∈ E and relations r j ∈ R correspond to one-hot row vectors, of dimension |E| and |R| respectively.…”

Autonomization of Monoidal Categories

“…The first is consists in giving logical semantics à la Montague, by translating nouns to abstract entities, adjectives to predicates on these entities, verbs to relations and sentences to logical formulas. This approach has been applied to pregroups in [Preller, 2005, Preller, 2012, Delpeuch and Preller, 2014. The second consists in representing words by real-valued vectors representing the typical context in which the words are encountered in a large corpus.…”

Complexity of Grammar Induction for Quantum Types

“…This framework is flexible, because it allows for different types of grammars, such as the Syntactic Calculus of Lambek [16] or Compact Bilinear Logic [19], also known as pregroups [15]. It also allows for different kinds of compositional semantics, which can be distributional [4], Montagovian and extensional [18], Montagovian and intensional [7], or even hybrid models [20]. But whatever the syntax or the semantics are, these approaches rely on a functor from the syntactic category to the semantic category to give meaning to a sentence.…”

“…As a finite-dimensional vector space is isomorphic to its dual, the category of finite-dimensional vector spaces is self-dual. This category has been widely used as the underlying semantic category for models of meaning, such as in [4], [20] or [7]. The objects in this category have also been used in [1] as semantic types in a learning task.…”