On Reversing the Generation Process in Optimality Theory

J. Eric FoslerACL 1996 · OpenReview

Abstract

Optimality Theory, a constraint-based phonology and morphology paradigm, has allowed linguists to make elegant analyses of many phenomena, including infixation and reduplication. In this work-in-progress, we build on the work of Ellison (1994) to investigate the possibility of using OT as a parsing tool that derives underlying forms from surface forms.

Introduction: The Challenge of Parsing in Optimality Theory

In linguistics, Optimality Theory (OT) describes how an underlying form (like a root word + a prefix) becomes a surface form (the word we actually say). It does this through a "survival of the fittest" process: a generator (GEN) creates a massive list of possible candidate pronunciations, and an evaluator (EVAL) uses a hierarchy of violable constraints to pick the "optimal" one.

While OT is excellent for generation, it is historically difficult to "run in reverse" for parsing—recovering the root from the spoken word. In older, rule-based derivational systems, one could simply "undo" the rules to find the root. In OT, the relationship between the input and output is often "hidden" because the evaluator only looks at the final candidates, not the process that created them.

Comparison of Derivational Phonology and Optimality Theory frameworks.

This paper addresses the computational challenge of making OT invertible so that a computer can hear a complex word and figure out its dictionary components.

Methodology: Two-Level OT and Finite State Transducers

The proposed approach builds on Ellison’s (1994) method, which represents OT constraints as Finite State Transducers (FSTs). FSTs are computational machines that map one string of symbols to another. To enable parsing, several technical shifts are introduced:

  1. From One-Level to Two-Level OT: Standard computational OT models usually only look at the output candidates. This work argues for a "two-level" approach where constraints have access to both the underlying "input" tape and the "surface" output tape simultaneously.
  2. Faithfulness Constraints as FSTs: Specific constraints—CORR (Correspondence) and MATCH (Identity)—are integrated into the FST framework. These ensure that the surface form maintains a traceable relationship to the underlying root.
  3. Multi-Tape Representation: To handle complex cases like Tagalog infixation (where a prefix like -um- jumps inside a word), the model utilizes extra "tapes" to track the underlying word and the prefix separately.
Multi-tape representation of the Tagalog word "grumadwet" including harmony marks and structural levels.

Key Insights: Constraints as Filters

The primary barrier to parsing in OT is that the generator function (GEN) typically "hides" the mapping between the input and the candidates. By making this mapping explicit within the constraints themselves, the system becomes mathematically reversible.

In this model, parsing is treated as a search through the search space of the FST. Instead of generating everything and then filtering, the FST acts as a multi-layered filter that can be traversed in either direction. Furthermore, harmony marks (the violation marks in an OT tableau) can be used as a scoring mechanism to guide a search algorithm through the massive space of possible underlying forms.

Results: Modeling Tagalog Infixation

The paper provides a theoretical proof of concept by modeling Tagalog Infixation. For example, the root gradwet ("graduate") plus the affix um becomes grumadwet. This is a difficult problem for traditional computational models because the affix is inserted into the root rather than being appended.

By including "faithfulness" tapes (Word Phonemes and Prefix Phonemes), the system can take the surface string grumadwet and "parse" it back to its underlying components um + gradwet. The results show that higher-level linguistic constraints (like NOCODA) can be seamlessly combined with alignment constraints in a single computational transducer.

Limitations and Future Work

As a work-in-progress, several challenges remain:

  • Computational Complexity: The search space for determining unknown strings during parsing is potentially massive. Efficient implementation will require smart shortcuts.
  • Heuristic Development: The exact method for using harmony marks to guide the search is still being refined. Future work envisions using these marks to prune the search tree early.
  • Implementation Gap: While the theoretical groundwork is established, full software implementation and performance benchmarks on larger lexicons were not yet completed at the time of publication.

Summary updated

Abstract

Optimality Theory, a constraint-based phonology and morphology paradigm, has allowed linguists to make elegant analyses of many phenomena, including infixation and reduplication. In this work-in-progress, we build on the work of Ellison (1994) to investigate the possibility of using OT as a parsing tool that derives underlying forms from surface forms.
BibTeX
@inproceedings{fosler1996reversing,
  author = {J. Eric Fosler},
  title = {{On Reversing the Generation Process in Optimality Theory}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1996},
  url = {https://openreview.net/forum?id=SybrohxuZH},
}
Figure