In languages like English, word order is the primary signal for meaning. If you swap the positions of "dog" and "cat" in "The dog chased the cat," the roles of subject and object flip. Consequently, most early Natural Language Processing (NLP) models were built on the assumption that proximity and sequence are the most important features for understanding structure.
However, for Morphologically Rich Languages (MRLs)—such as Arabic, Hebrew, Turkish, or German—this assumption fails. In these languages, grammatical relationships are often encoded inside the words themselves through inflections, rather than through their position in a sentence. Research across several decades has converged on a singular insight: to parse these languages accurately, models must treat morphological features like case, gender, and number not just as metadata, but as the "structural anchors" that hold a sentence together.
The Agreement Constraint as a Filter
The most intuitive reason why morphology matters is agreement. In many languages, a modifier (like an adjective) must match its head (like a noun) in gender, number, and person. Manny Rayner and colleagues demonstrated as early as 2001 that enforcing these constraints in speech recognition systems can reduce semantic error rates by up to 46%.
Agreement acts as a powerful filter. When a parser is deciding which noun a pronoun refers to, or which noun an adjective modifies, morphological features provide a "lock and key" mechanism. As Robert Ingria and David Stallard argued, without checking for person, number, and gender, a resolution algorithm will inevitably link words to linguistically incompatible partners. In MRLs, these features are often more reliable than word order for determining syntactic links (Lee et al., 2011).
Moving Beyond Atomic Words
Standard parsers often treat words as "atomic" units—indivisible tokens. This creates a massive data sparsity problem in MRLs, where a single root word can take hundreds of different forms. Christos Christodoulopoulos and Sharon Goldwater showed that breaking words down into sub-word features allows models to generalize across different forms of the same category.
Miguel Ballesteros and his team found that gender and number are particularly "productive" features. They serve as anchors that allow a parser to link a dependent to its head across long distances in a sentence, even when other words intervene. This is why Benoît Crabbé proposed integrating these attributes directly into the state transitions of a parser, effectively letting the model "see" the internal grammar of a word as it builds the sentence tree.
The Case for Case
Among all morphological features, grammatical case (which marks the role of a noun, such as subject or object) is perhaps the most critical for parsing. In languages with flexible word order, case markers are the primary signal for who did what to whom.
Jivnesh Sandhan and Amrith Krishna noted that case tagging provides a more potent inductive bias than any other morphological signal. Even modern neural models struggle without it; Clara Vania and Adam Lopez demonstrated that character-level models often fail to resolve "case syncretism" (when different cases look identical) unless they are explicitly provided with case information.
Joint Modeling and Feature Percolation
A recurring challenge in this field is the "chicken-and-egg" problem: you need to know the morphology to parse the syntax, but you often need the syntactic context to disambiguate the morphology. Reut Tsarfaty has been a leading proponent of jointly modeling these two steps.
One elegant solution to this is feature percolation. Khalil Sima'an and Tsarfaty introduced a "third dimension" to parsing parameters: depth. In this framework, morphological agreement features are not just kept at the word level; they are "percolated" up the parse tree to higher-level nodes. This allows the model to ensure that an entire phrase remains morphologically consistent, resolving ambiguities that word order alone cannot touch.
Morphology in the Age of LLMs
While Large Language Models (LLMs) have revolutionized NLP, the need for explicit morphological features has not disappeared. Recent work on the UniDive framework by Kyuwon Kim and colleagues shows that even instruction-tuned LLMs see significant accuracy boosts when fine-grained morphological attributes like honorifics and mood are integrated directly into the dependency tree.
As Mathieu Dehouck and Pascal Denis concluded, morphology is not just a way to handle rare words; it is a primary driver of grammatical structure. By treating these features as first-class citizens in the parsing process, researchers continue to bridge the gap between rigid, order-based models and the fluid, inflection-based reality of the world's languages.