Word Segmentation as Unsupervised Constituency Parsing

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 4103 - 4112 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Word Segmentation as Unsupervised Constituency Parsing Raquel G. Alhama Tilburg University, Warandelaan 2, 5037AB Tilburg, the Netherlands rgalhama@uvt.nl Abstract Word identification from continuous input is typically viewed as a segmentation task. Ex- periments with human adults suggest that fa- miliarity with syntactic structures in their na- tive language also influences word identifica- tion in artificial languages; however, the re- lation between syntactic processing and word identification is yet unclear. This work takes one step forward by exploring a radically different approach of word identification, in which segmentation of a continuous input is viewed as a process isomorphic to unsuper- vised constituency parsing. Besides formal- izing the approach, this study reports simu- lations of human experiments with DIORA (Drozdov et al., 2019), a neural unsupervised constituency parser. Results show that this model can reproduce human behavior in word identification experiments, suggesting that this is a viable approach to study word identifica- tion and its relation to syntactic processing. 1 Introduction When exposed to speech in an unknown language, humans are faced with the task of finding out what are the basic combinatorial units of the language, such as phonemes, syllables, words and phrases. Since speech is continuous, humans need to rely on implicit cues –such as statistical information– to find out the building blocks of the language. One approach that studies which statistical cues can be used by humans in this task is Artificial Grammar Learning (AGL). Experiments in AGL are char- acterized by the use of artificial languages with carefully controlled statistical properties. To in- vestigate word identification with this paradigm, participants in a typical AGL experiment are first exposed to a speech-like sample of the artificial language (usually recorded with synthetic voice). Then, they participate in a test that has been de- signed to show whether participants identified the words in the artificial language. To formalize theories of how humans identify words in AGL tasks, a range of computational mod- els have been proposed over the last two decades. These models have explained a wide arrange of phenomena, using a variety of algorithms such as Bayesian inference (Frank et al., 2010), normative statistics (Swingley, 2005), cognitively inspired processes implementing recognition or memoriza- tion (Alhama and Zuidema, 2017; Perruchet and Vinter, 1998), and neural networks (French et al., 2011; Endress and Johnson, 2021). There is, however, one phenomenon that has not been addressed in the computational literature in AGL: the fact that participant’s knowledge of their native language influences performance in this type of AGL experiments. In particular, results seem to be influenced by co-occurrence statistics of sublexical units (Onnis et al., 2005; Siegelman et al., 2018; Elazar et al., 2022), and interestingly, also by the presence of left- or right-branching syntactic structures in the native language, which predict the statistics that subjects use to identify words (Onnis and Thiessen, 2013). One …

Content truncated to 500 words