Multi-Tiered Strictly Local Functions

Phillip Burness, Kevin McMullinACL 2020

TL;DR

This paper introduces "Multi-Tiered Strictly Local" (MTSL) functions, a novel framework for modeling complex phonological phenomena that require simultaneous analysis across multiple levels of representation. By allowing functions to access and integrate information from multiple "tiers" with a strict hierarchical relationship, MTSL functions can elegantly capture interactions that were previously intractable for existing subregular models.

What Problem Does This Paper Address?

Phonological processes, like the way sounds change in a language, often involve relationships between sounds that are not directly adjacent in a spoken word. Traditional models, such as Strictly Local (SL) functions, excel at capturing local patterns but struggle with these "long-distance" dependencies. While extensions like Tier-based Strictly Local (TSL) functions allow for analysis on a single "tier" (a specific level of representation), they fall short when multiple, independent long-distance processes need to operate simultaneously or interact in complex ways. For instance, modeling languages like Tamashek Tuareg, which exhibits both sibilant harmony and labial dissimilation, requires a mechanism that can handle these concurrent rules without interference. This paper addresses this limitation by proposing a way for functions to operate across multiple tiers in a structured manner.

Methods

The core innovation is the introduction of Multi-Tiered Strictly Local (MTSL) functions. These functions extend the concept of TSL functions by allowing them to consider multiple tiers of representation simultaneously. The key restriction is that for any given input element, the associated tiers must form a strict superset-subset hierarchy. This means that if a set of tiers $T_1$ is associated with an element, and another set $T_2$ is also associated, then either $T_1 \subset T_2$ or $T_2 \subset T_1$.

Furthermore, the paper introduces the concept of target specification. This ensures that the contribution of each individual input segment to the output is determined by a specific, hierarchically structured set of tiers. This prevents "gang-up" effects where arbitrary combinations of tiers could lead to unattested phonological behaviors. The paper defines both Input MTSL (IMTSL) and Output MTSL (OMTSL) functions, analogous to their TSL counterparts, but operating across multiple tiers.

Datasets

Not discussed in the paper. The paper focuses on theoretical formalisms and uses linguistic examples from various languages (Tamashek Tuareg, Samala, Turkish, Uyghur) to illustrate the capabilities of the proposed MTSL functions.

Key Insights and Ideas

  • Multi-tier analysis is necessary for complex phonology: The paper convincingly argues that single-tier models are insufficient for languages with multiple, interacting long-distance phonological processes.
  • Hierarchical tier structure prevents overgeneration: The strict superset-subset hierarchy on tiers is crucial. It limits the combinatorial possibilities and ensures that the model adheres to attested phonological patterns, avoiding the "tier multiverse" problem.
  • Target specification localizes influence: By associating specific tiers to individual input segments, the model can precisely determine how each segment contributes to the output, even when multiple processes are at play. This is a more granular approach than considering the entire set of tiers at once.
  • MTSL functions offer elegant solutions: The paper demonstrates that MTSL functions can provide simple and intuitive analyses for phenomena that were previously challenging or impossible to model with existing formalisms.

Key Results

The paper does not present quantitative results in terms of accuracy or performance metrics on datasets. Instead, its key results are qualitative:

  • Successful modeling of complex interactions: The paper demonstrates how MTSL functions can successfully model:
  • Simultaneous independent processes (e.g., sibilant harmony and labial dissimilation in Tamashek Tuareg).
  • Conflicting processes where local rules override non-local ones (e.g., local blocking of dissimilation in Tamashek Tuareg).
  • Single targets subject to multiple harmonies across different featural dimensions (e.g., anteriority and voicing harmony in Imdlawn Tashlhiyt).
  • Last-resort triggers for harmony when preferred triggers are absent (e.g., dorsal consonants in Uyghur).
  • Formalization of a new class of functions: The paper formally defines IMTSL and OMTSL functions, providing a rigorous framework for analyzing these complex phonological patterns.

Limitations and Future Work

  • Bidirectional application: The paper acknowledges that the proposed MTSL functions do not inherently cover bidirectional phonological processes, which are typically modeled by weakly deterministic functions. The authors suggest exploring whether MTSL functions, applied in sequence and in opposite directions, could model these.
  • Learnability: While the paper proposes a powerful class of functions, its learnability is not fully explored. The authors note that learning MTSL functions, especially when tiers are unknown, presents a significant challenge. They suggest that existing work on learning multi-tiered stringsets might offer a path forward.
  • Closure properties: The paper speculates that left-reading MTSL functions might be the closure of left-reading TSL functions under composition, but this is not formally proven.
  • Complexity bounds: The authors are interested in whether the complexity bound of bidirectional processes can be lowered by using MTSL functions.

Summary updated

BibTeX
@inproceedings{burness2020multi,
  author = {Phillip Burness and Kevin McMullin},
  title = {{Multi-Tiered Strictly Local Functions}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2020},
}
Figure