Finite-state Model of Shupamem Reduplication

Magdalena Markowska, Jeffrey Heinz, Owen RambowACL 2021

TL;DR

This paper proposes a novel finite-state machine model that can handle both full reduplication and autosegmental tone phenomena in languages like Shupamem. By synthesizing 2-way finite-state transducers (FSTs) with multi-tape FSTs, the authors demonstrate a computationally efficient way to model complex morphophonological processes that were previously challenging for standard FST approaches.

What Problem Does This Paper Address?

Many languages exhibit reduplication, a morphological process where a part or all of a word is repeated. While partial reduplication can be modeled with standard 1-way finite-state transducers (FSTs), full reduplication poses a significant challenge because the copied segment can be arbitrarily long, leading to an unbounded number of states in a 1-way FST. Furthermore, many tonal languages, like Shupamem, have tones that behave independently of the segments they attach to (autosegmental phenomena). This means a computational model needs to handle both the repetition of segments and the independent movement and interaction of tones. Existing models often struggle to integrate these two complex phenomena efficiently.

Methods

The paper proposes a synthesis of two types of FSTs:

  • 2-way FSTs: These transducers can move both forward and backward on the input tape. This capability is crucial for modeling full reduplication efficiently, as it allows the machine to "copy" the input string without an exponential increase in states.
  • Multi-tape FSTs (MT FSTs): These machines operate on multiple input and output tapes simultaneously. This is ideal for representing autosegmental phenomena, where tones and segments can be treated as independent tiers. The authors propose using separate tapes for tonal and segmental information.

The core contribution is the combination of these two approaches into a 2-way n,m Multitape FST. This allows for the simultaneous handling of segmental reduplication (via the 2-way capability) and tonal interactions (via the multi-tape representation). The paper also discusses the "linearization" of these multi-tape outputs into a single string, which is a necessary step for generating the final surface form of words.

Datasets

Not discussed in the paper. The paper focuses on analyzing the linguistic phenomena of Shupamem and proposing a computational model, rather than training on or evaluating against a specific dataset. Examples are drawn from linguistic literature.

Key Insights and Ideas

  • Unified Model for Reduplication and Tone: The most significant insight is the successful integration of 2-way FSTs for full reduplication and MT FSTs for autosegmental tone into a single, coherent framework. This addresses a long-standing challenge in computational phonology.
  • Tonal and Segmental Tiers as Independent Tapes: Representing tones and segments on separate tapes within the MT FST framework elegantly captures the linguistic insight that these elements can behave independently.
  • 2-way FSTs for Efficient Reduplication: The paper reinforces the idea that 2-way FSTs are a powerful and compact solution for modeling full reduplication, avoiding the state explosion seen with 1-way FSTs.
  • Compositionality of FSTs: The paper highlights how combining different types of FSTs (1-way, 2-way, multi-tape) through composition can lead to powerful models capable of handling complex linguistic phenomena.

Key Results

The paper does not present quantitative results in terms of accuracy or performance metrics on a dataset. Instead, it demonstrates the feasibility and elegance of the proposed finite-state model by:

  • Providing a formal definition of the 2-way n,m Multitape FST.
  • Illustrating the model's application to Shupamem nominal and verbal reduplication, including tonal alternations.
  • Showing how the model can handle specific examples like the reduplication of "ndap" (house) and "k'a" (fry).
  • Discussing how the model can be extended to other tonal languages with similar reduplication patterns.

Limitations and Future Work

  • Alternative Models: The authors acknowledge that full reduplication can also be achieved by composing 1-way MT FSTs (e.g., a 1-way 1,3 MT FST followed by a 1-way 3,1 MT FST), suggesting a need for further formal comparison between these approaches.
  • Closure Properties: Future research is planned to investigate the closure properties of classes of 2-way MT FSTs.
  • Broader Typology: The authors aim to collect more examples of full reduplication in tonal languages for the RedTyp database to enable broader empirical typological studies.
  • Generalizability: While the model is shown to work for Shupamem and is discussed in the context of other languages (Adhola, Kikerewe, Shona), further work is needed to explore its applicability to a wider range of tonal languages and reduplication types.

Summary updated

BibTeX
@inproceedings{markowska2021finite,
  author = {Magdalena Markowska and Jeffrey Heinz and Owen Rambow},
  title = {{Finite-state Model of Shupamem Reduplication}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}
Figure