Morphology, Syntax, and What’s in Between: Reifying Cross-linguistic Variation through Cross-linguistic Parsing

A central part of NLP research is devoted to the development of parsing systems which analyze the way words combine to form phrases and sentences. This analysis is considered the first step towards natural language understanding, or, as it is often put, extracting “who did what to whom”. Such information is crucial for computer programs that perform tasks such as information retrieval, question answering, and machine translation, to name just a few. To date, parsing systems that were developed and applied for English show excellent performance, but the application of the same models to languages such as German, Czech, Arabic, Hebrew, Turkish, Hindi, and more, doesn’t necessarily yield comparable results. This discrepancy gives rise to a fascinating challenge, namely: Is it possible to devise a single parsing system which is abstract enough to accomodate different languages, and yet specific enough to learn from data the structure and properties of particular languages? Addressing this question is not only technologically challenging, but is also of utter importance from a scientific point of view. The number of languages in the world is estimated at around 4000-6000 and yet it is striking to realize how similar the different languages are in the principles underlying their organization. In order to formally refer to the commonalities and differences between languages the linguist Noam Chomsky used the term universal grammar (UG), referring to a set of principles which are shared by all languages, and parameters settings that differentiate languages from one another. Modern typological studies following the work of (Sapir 1921, Greenberg 1963) and others play an important part in unraveling the different dimensions of variation between languages (the ’principles’) and instantiating their possible values (the parameter ‘settings’). Typology tells us, for instance, that all languages have nouns and verbs, and that the relation between a noun and a verb can be of a subject or of an object type. This is the functional dimension of language description. Languages vary, however, in how they express these concepts. For instance, subjects in English tend to appear before the verb, but in Warlpiri, subjects can appear almost anywhere in the sentence. This is the basic word-order (or, structural) dimension of language variation (Greenberg 1963). Languages also vary in how much information is expressed within a single word. In English, for instance, the subject, verb and object of a sentence are expressed in different words. In Arabic, Hebrew, or Turkish, the subject, verb and object may appear within a single word. This is the morphological dimension of variation (Sapir 1921).