Morphological Inflection in Context

This direction is no longer present in the current build. Its last version stays readable and citable here.

Version 1 of 115 papers18 conceptsfrozen 2026-10-04

In traditional natural language processing, morphological inflection is often treated as a dictionary problem: given a base word (a lemma) and a set of grammatical features (like "plural" or "past tense"), the system generates the correct surface form. However, in real-world applications like machine translation or OCR, these grammatical labels are rarely provided. Instead, the system must look at the word's neighbors to decide its shape.

This research direction focuses on contextual morphological inflection—the ability to predict the correct form of a word by leveraging the syntactic and semantic signals embedded in the surrounding text.

The Neighborhood as a Signal

The core intuition behind this work is that a word’s grammatical role is not hidden; it is reflected in its environment. When a human reads a sentence with a missing word, they can often guess not just the meaning, but the exact grammatical form required.

In the realm of machine translation for morphologically rich languages, Arianna Bisazza and colleagues demonstrated in Word Translation Prediction for Morphologically Rich Languages with Bilingual Neural Networks that the source-side context is a powerful proxy for target-side morphology. By looking at the neighborhood of the source text, models can infer case and gender in the target language without needing explicit part-of-speech taggers.

This concept extends even to degraded text. Tao Hong showed in Integration of Visual Inter-Word Constraints and Linguistic Knowledge in Degraded Text Recognition that visual similarity between words across a document can be combined with statistical language models to resolve ambiguities. If a "noisy" image of a word could be two different things, the surrounding grammatical context acts as a tie-breaker.

Beyond Isolated Word-Tag Pairs

For a long time, morphological models operated on isolated word-tag pairs. The challenge, as identified by Yova Kementchedjhieva and the "Team Copenhagen" researchers in Copenhagen at CoNLL-SIGMORPHON 2018, is that surface forms are dictated by sentence-level constraints.

To solve this, researchers have moved toward "latent" morphological reasoning. In Contextualization of Morphological Inflection, Ekaterina Vylomova and Ryan Cotterell proposed models that treat grammatical tags as latent variables. Instead of being told a word is a "plural noun," the model "thinks" about the likely grammar of the sentence first, then uses that internal representation to generate the text. This mimics the human process of ensuring a sentence "sounds right" based on agreement rules.

Joint Realization and Linearization

One of the most significant shifts in this field is the move from modular pipelines to joint systems. In traditional Natural Language Generation (NLG), one system might decide the word order (linearization) and another might decide the word forms (inflection).

However, Ratish Puduppully and Yue Zhang argue in Transition-Based Deep Input Linearization that these tasks are inseparable. By integrating morphological inflection directly into a transition-based shift-reduce process, the model can select the correct word form as it builds the sentence. As a lemma moves from a buffer to a stack, the model looks at the words it has already placed to determine the correct tense or case. This prevents "error propagation," where a mistake in word order leads to a nonsensical morphological choice.

Decoding Latent Features

In languages like Arabic, the relationship between a word's appearance and its grammar is particularly complex. Nizar Habash and Sarah Alkuhlani explored this in Identifying Broken Plurals, Irregular Gender, and Rationality in Arabic Text. They found that when morphological cues are orthographically absent (e.g., a word that looks singular but acts plural), syntactic dependency relationships serve as a robust proxy. By analyzing how a word relates to its head or its modifiers, the system can "see" hidden properties like rationality or irregular gender that are not visible in the word's spelling alone.

Why It Works

Contextual morphological inflection works because language is redundant. The information required to inflect a word is often distributed across the entire sentence:

  1. Syntactic Agreement: A subject's number dictates the verb's ending.
  2. Semantic Constraints: The "rationality" of a noun (human vs. non-human) affects the adjectives that describe it.
  3. Positional Cues: In transition-based systems, the current state of the sentence construction provides a local environment that constrains the possible forms of the next word.

By treating morphology as a contextual problem rather than a lookup problem, these researchers have enabled models to generate more fluent, grammatically accurate text across diverse and morphologically complex languages.

Papers in this version

History & citationVersion 1 · 2026-10-04

No longer present in the current build · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_morphological_inflection_in_context_2243c817,
  title = {Morphological Inflection in Context},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/morphological-inflection-in-context/rdr_2243c8176af84bfc8fc344bc6f2f74f7},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_morphological_inflection_in_context_2243c817_v1,
  title = {Morphological Inflection in Context},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/morphological-inflection-in-context/rdr_2243c8176af84bfc8fc344bc6f2f74f7/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}