Emergence of syntax


Katrien Beuls, Emilia Garcia, Luc Steels 


This session presents a number of case studies focusing on the emergence of syntactic structure, more specifically constituent structure. The case studies start from a language game in which agents refer to more than one object in the shared situation. They therefore have to express multiple predicates using multiple words. This raises potentially a combinatorial explosion: How can the hearer know (without some form of syntax) which words belong together, i.e. which words refer to the same object. Syntax solves this problem through formal means such as word order, hierarchical structure, and agreement marking. 


We argue that agents adopt a set of language strategies with which they individually and collectively create and maintain a shared language. A language strategy has components for invention, adoption and alignment. Depending on the nature of the strategy, we get different types of language systems to emerge. The talk presents a series of such language strategies, first focusing on agreement and then on word order. 


The discussion of agreement starts from a lexical strategy, in which agents do not signal which words are about the same object and thus end up in combinatorial explosions during parsing and semantic interpretation. Next, agents adopt a formal marking strategy in which 'stickers' are attached to the words that belong together. This strategy already works but is not realistic with respect to human language. The next strategy uses meaningful markers, i.e. markers that are also attached to words but introduce meaning on their own, which is similar to an agreement marker like "-arum" in Latin which signals not only agreement but also feminine, plural and genitive case. This strategy will be shown to work also but is still not fully in line with the historical data, which show that agreement systems arise by a grammaticalisation process: existing words are recruited as agreement markers and these words then undergo a process of change in two directions: erosion of the form on the one hand and generalisation (bleaching) and conventionalisation of the meaning on the other. We introduce additional strategies that have these grammaticalisation processes as a side effect, thus demonstrating how and why agreement systems with similar properties as found in human natural languages could have evolved. 


The discussion of word ordering starts again from a lexical strategy, which is shown to lead again to combinatorial explosions in parsing. The next strategy is to group the words that belong together in the utterance, but without introducing any internal word ordering, nor any ordering among the groups. This substantially reduces the combinatorial explosion, particularly when hearers continuously use the shared situation to cut down on possible interpretations. Further improvements are reached when the agents start using patterns. A first possible strategy in this direction is word-based. When a series of words is used to refer to an object, then those words are stored in the same order and reused. Because different agents may invent different word patterns, a process is needed so that they collectively negotiate which agreement they should use. An improvement of this strategy consists in introducing syntactic categories to make the patterns more generally applicable so that fewer patterns need to be stored, agreed upon, and learned. We introduce such a strategy and show that it is leading to a shared set of syntactic categories and patterns. Finally, we show how hierarchical structure arises when the agents start expressing relations between objects. 


Further references: Beuls, K. and L. Steels (2013) Agent-based models for the Emergence and Evolution of Grammatical Agreement. 

Plos One 18 March 2013. http://dx.plos.org/10.1371/journal.pone.0058960