The Tradeoff Between Morphological Normalization and Semantic Precision

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 258 papers79 conceptsfrozen 2026-10-04

In natural language processing, we often face a "data sparsity" problem: a single concept can appear in dozens of different forms. In Hungarian, a single noun can take over 1,400 forms (Tordai and de Rijke), and in many languages, the base form of a verb represents only a tiny fraction of its total occurrences (Kuznetsov and Gurevych).

To solve this, researchers traditionally use normalization—stemming or lemmatization—to collapse these variations into a single root. The intuition is simple: by treating "running," "runs," and "ran" as the same entity, we increase the statistical strength of our models. However, a vast body of research suggests that this simplification comes at a steep cost. When we strip away the "surface" of a word, we often discard the very information the model needs to understand the sentence.

The Loss of Grammatical Logic

The primary danger of aggressive normalization is the destruction of functional markers. Affixes (prefixes and suffixes) are not just decorative; they encode critical information like tense, number, gender, and case.

When Djamé Seddah and colleagues investigated why parsers struggle with morphologically rich languages like French, they found that reducing words to base forms strips away the markers required for syntactic disambiguation. Without these cues, a parser cannot tell who is doing what to whom.

This problem is particularly acute in Coreference Resolution. Haixia Chai and Michael Strube demonstrated that lemmatization hinders a model's ability to link pronouns to nouns. If a model sees only the base lemma of a noun, it loses the gender and number information necessary to match that noun with the correct pronoun later in the text.

Semantic Conflation and Over-Stemming

Beyond grammar, normalization can lead to "over-conflation"—the accidental grouping of semantically unrelated words that happen to share a linguistic root.

In Information Retrieval (IR), Chris Buckley and Gerard Salton observed that while stemming improves the breadth of matches (recall), it often fails to improve overall effectiveness because it destroys specificity. Tomek Strzalkowski and Barbara Vauthey noted that standard algorithms often over-reduce words to roots that are linguistically invalid or lack semantic precision, leading to incorrect matches between queries and documents.

This "stemming dilemma" is a major hurdle in web search. Ahmed Ali and colleagues found that aggressive stemming can be "unsafe" in massive collections where the cost of a false positive (a semantically unrelated match) outweighs the benefit of finding more documents. For example, a user searching for "jobs" (employment) might be frustrated if a system stems the query to "job" and returns results about the biblical figure.

Why Contextualized Models Suffer

One might assume that modern, "intelligent" models like BERT would be immune to these issues. In fact, the opposite is often true.

Alexandre Allauzen and Syrielle Montariol found that contextualized language models experience significant performance degradation when applied to lemmatized text. These models are trained to leverage the full sentence context; when we remove morphological variations, we limit the model's ability to use the surrounding grammatical "landscape" for disambiguation. By normalizing the text, we effectively "blind" the model to the subtle cues it was trained to see.

Moving Toward Morphological Awareness

The research suggests that the solution isn't to abandon morphology, but to treat it as a first-class citizen rather than a noise-reduction problem.

  1. Two-Tier Architectures: Instead of flattening words into tokens, models like KinyaBERT (Antoine Nzeyimana and Andre Niyongabo Rubungo) use architectures that explicitly encode both stems and affixes. This allows the model to benefit from the shared meaning of the root while still "seeing" the functional role of the affix.
  2. Paraphrase-Based Approaches: Rather than forcing a word into a single root, Preslav Nakov and Hwee Tou Ng suggest treating morphologically related words as potential paraphrases. This maintains the semantic distinction between an actor and an action (derivational morphology) while still allowing the model to recognize their relationship.
  3. Selective Normalization: In tasks like word alignment for translation, Hervé Déjean and colleagues found that selective lemmatization—applying it only where it reduces parameter space without losing critical cues—outperforms the "blanket" approach of normalizing every word in a corpus.

Ultimately, the trend in the field is moving away from "stripping" language and toward "enriching" it. As Antoine Nzeyimana argues, when we treat words as flat sequences, we force models to expend massive computational effort to reconstruct basic grammatical logic that was already explicitly encoded in the language's morphology. By preserving that structure, we create models that are both more efficient and more precise.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 71 papers
  2. Version 1 · 2026-10-04 · 58 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_morphological_information_loss_39a94feb,
  title = {morphological information loss},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/morphological-information-loss/rdr_39a94feb315e48f4a45a8198df00fd09},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_the_tradeoff_between_morphological_norma_39a94feb_v1,
  title = {The Tradeoff Between Morphological Normalization and Semantic Precision},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/the-tradeoff-between-morphological-normalization-and-semantic-precision/rdr_39a94feb315e48f4a45a8198df00fd09/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}