Sanskrit n-Retroflexion is Input-Output Tier-Based Strictly Local

Thomas Graf, Connor MayerEMNLP 2018

TL;DR

This paper introduces "Input-Output Tier-Based Strictly Local" (IO-TSL), a new formal language class that extends existing subregular phonology models. They demonstrate that IO-TSL is powerful enough to capture the complex Sanskrit n-retroflexion (nati) process, a phenomenon previously considered difficult to model within subregular frameworks. This work suggests that IO-TSL might be a suitable formalization for the entirety of segmental phonology.

What Problem Does This Paper Address?

The field of subregular phonology aims to identify formal language classes that are more restrictive than finite-state languages but still powerful enough to model natural language phenomena. This allows for tighter bounds on linguistic variation and more efficient learning algorithms. Sanskrit n-retroflexion (nati) is a well-studied phonological process known for its complexity, involving long-distance assimilation, multiple blocking effects, and specific contextual requirements. While existing subregular classes like "Strictly Local" (SL) and "Tier-Based Strictly Local" (TSL) can handle many phonological patterns, they have proven insufficient for capturing the full complexity of nati. The authors sought to find a formal language class that could account for nati, potentially pushing the boundaries of what is expressible within subregular phonology.

Methods

The paper introduces and formalizes Input-Output Tier-Based Strictly Local (IO-TSL), a generalization of TSL. The core innovation lies in the tier projection function, which in TSL simply masks out segments not belonging to a specified alphabet. In IO-TSL, this projection function becomes more sophisticated: it considers not only the local context in the input string but also the symbols already present on the tier. This "input-output" dependency allows for more complex regulation of tier construction.

The authors then apply this IO-TSL framework to model Sanskrit n-retroflexion. They define a specific IO-TSL tier projection mechanism with several rules designed to capture the various conditions of nati, such as the trigger (a retroflex continuant), the target (/n/), and the blocking conditions (intervening coronals, specific following sounds, preceding velar/labial plosives, and following retroflexes). The resulting tiers are then checked against a simple SL-3 language to determine if the string is well-formed according to the nati rules.

Datasets

The paper uses data from existing linguistic literature on Sanskrit n-retroflexion. Specifically, they cite and draw examples from:

  • Ryan (2017)
  • Müller (1886)
  • Hansson (2001)
  • Whitney (1889)
  • Emeneau (1946)

No new datasets or benchmarks were introduced. The focus is on analyzing existing linguistic data within a new formal framework.

Key Insights and Ideas

  • IO-TSL as a Natural Extension: The paper argues that IO-TSL is a natural and intuitive extension of TSL, incorporating independently motivated ideas for making tier projection more "structure-sensitive."
  • Sanskrit Nati as a Benchmark: The complexity of Sanskrit n-retroflexion serves as a strong test case for formal language classes. Successfully modeling it suggests the class has significant expressive power for phonological phenomena.
  • Tier Projection is Key: The authors highlight that the power of IO-TSL comes from its enhanced tier projection mechanism, which can look at both the input string and the partially constructed output tier. This allows for more nuanced control over which segments are preserved.
  • Low Formal Complexity for Nati: Despite the linguistic complexity of nati, the authors find that it can be captured by IO-TSL with surprisingly low bounds on the tier projection parameters (IO-TSL-(3,2,3)). This suggests that the underlying formal machinery is quite efficient.
  • Potential for Segmental Phonology: The paper conjectures that IO-TSL might be sufficient to capture all of segmental phonology, a significant claim about the formal limits of this linguistic domain.

Key Results

The primary result is the successful formalization of Sanskrit n-retroflexion using the newly proposed IO-TSL framework. The authors demonstrate that the complex conditions of nati can be encoded within an IO-TSL tier projection and an SL-3 language. They show that the resulting formal description is IO-TSL-(3,2,3), indicating a relatively low complexity for this linguistically challenging phenomenon.

Limitations and Future Work

  • Conjectures on IO-TSL Properties: The authors make several conjectures about IO-TSL, including that it is a proper subclass of star-free languages and that it is not closed under intersection, union, relative complement, or concatenation. They also conjecture its incomparability with other subregular classes like LTT, LT, PT, IBSP, and SP. These properties would mirror those of TSL, further solidifying IO-TSL's position as a natural generalization.
  • Learnability: While IO-TSL languages are finite classes (given bounds on i, j, k) and thus learnable in the limit, the paper raises questions about efficient learnability. The authors note that learners don't have direct access to the output of the tier projection function, and inferring correct contexts requires correctly built tiers, posing significant challenges for efficient learning algorithms.
  • Unattested Phenomena: The authors acknowledge that IO-TSL, like TSL extensions, might still be too powerful and could potentially allow for unattested phenomena, such as first-last harmony. Future work could involve identifying further subclasses of IO-TSL to restrict these possibilities.
  • Right-to-Left Projection: The current analysis uses left-to-right tier projection. The authors suggest that allowing right-to-left projection might also affect complexity and potentially lower thresholds.
  • Abstract Symbols: The analysis uses abstract symbols (R, S, C, P) for clarity. While they argue that the complexity remains the same when compiled out to actual segments, they note potential issues if segments overlap in their abstract roles (e.g., if [j] were both a sonorant and a coronal blocker).

Summary updated

BibTeX
@inproceedings{graf2018sanskrit,
  author = {Thomas Graf and Connor Mayer},
  title = {{Sanskrit n-Retroflexion is Input-Output Tier-Based Strictly Local}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2018},
}
Figure