The %mor Tier: Morphological Analysis

Status: Reference Last updated: 2026-10-20 (commit f7e32e22)

The %mor tier records morphological analysis. Chatter parses, represents, validates and serializes that analysis; it does not itself infer the correct analysis from speech. A successful parse or alignment check does not certify linguistic accuracy.

Format Overview

*CHI:	I want cookies .
%mor:	pron|I-Prs-Nom-S1 verb|want-Fin-Ind-Pres-S1 noun|cookie-Plur .

A morphological word has the shape POS|lemma[-Feature]*:

• POS is a category string, such as noun, verb or pron.
• | separates the category from the lemma.
• The lemma is the lexical payload.
• Each - introduces a feature, such as Plur or Number=Plur.

Items are space-separated. A MOR item contains a main morphological word and zero or more post-clitics. The tier ends with a typed terminator.

The UD MOR Format

UD-style annotation often uses lowercase category names and flat feature values, as in the example above. This is not a claim that Chatter’s model is a complete Universal Dependencies interchange format.

The parser also accepts colon-refined categories such as pro:sub and v:aux. PosCategory retains the entire tag; category() returns its head and subcategories() iterates the refinements. Do not discard those refinements merely because a particular producer emits flat UPOS-like tags.

Multi-Word Tokens (Clitics)

*CHI:	it's red .
%mor:	pron|it~aux|be-Fin-Ind-Pres-S3 adj|red .

Here the first MOR item has two morphological words: its main word and one post-clitic separated by ~. It still occupies one main-to-MOR item slot, but contributes two MOR chunks for %gra.

Terminator

MorTier requires a Terminator; its constructors cannot represent an absent MOR terminator. Missing terminators in raw input are rejected during parsing, not filled in by the model. Terminator value agreement with the main tier is checked separately from item counts.

The terminator contributes one chunk to %gra. Thus the example above has two MOR items and four chunks: pron|it, the auxiliary post-clitic, adj|red, and the period.

Feature Values and Keys

MorFeature stores an optional key and a value. These examples describe constructor behavior, not whole-document admission:

new splits at the first =. flat stores a keyless value explicitly; with_key_value stores both parts verbatim. None of these constructors consults a UD feature inventory. Empty values remain representable and are checked by MOR content validation.

Commas remain part of the value string; Chatter does not turn Int,Rel into a set of two semantic values. Likewise, it does not reconstruct missing feature keys or interpret a producer’s composite code such as S3. A consumer needs the producer’s convention for those interpretations.

Accepting keyed features syntactically does not prove that changing a producer from flat to keyed features is compatible with every consumer.

Compound Lemma Boundaries

The lemma is stored as an opaque MorStem. The current grammar permits characters such as =, !, _, and # in interior positions, with stricter restrictions on the first character. Structural delimiters such as |, -, ~, +, whitespace, period and question mark still constrain tokenization; arbitrary Unicode text is not automatically a valid MOR lemma.

Chatter does not give an interior = a separate translation-gloss field. For example, a retained perro=dog lemma is one string, not a typed pair of lemma and translation. Nor does the model interpret an interior marker as a language-specific compound boundary.

A downstream generator may normalize or remove characters before Chatter sees them. That is a separate transformation contract: inspect its exact implementation and output rather than attributing that behavior to this parser. Parsing cannot recover information already removed by a producer.

Data Model

Use the public types and accessors rather than copying their declarations:

MorTier::items() exposes item order; items_mut() exposes the existing items as a mutable slice. chunks() exposes main words, post-clitics and the terminator in their serialized order. count_chunks() includes all of them. Use the chunk projection for GRA consumers rather than counting only items.

The model’s ability to hold a value is not proof that it is syntactically valid CHAT or a linguistically valid analysis. See the validation architecture for admission boundaries.

Grammar and Parser

The grammar owner is grammar/grammar.js, particularly mor_content, mor_word, mor_pos, mor_lemma and mor_feature_value. The Tree-sitter consumer lives under crates/talkbank-parser/src/parser/tier_parsers/mor/; the model lives under crates/talkbank-model/src/model/dependent_tier/mor/.

The grammar recognizes features as surface strings. The model splits keyed features; this does not validate their UD meaning. Recovery diagnostics must be considered separately from any model produced by a parser.

Reference fixtures, spec examples and focused parser/model tests provide different correctness evidence. Merely accepting a legacy-looking string does not establish full support for a historical MOR dialect, its semantic structure, or every downstream analysis.

Validation

• E711: MOR content validation checks empty POS, lemma and feature values in main words and post-clitics. This is not a UD vocabulary check.
• E705 / E706: too few / too many MOR items for the main-tier MOR projection. That projection is not the raw content array: retraces, replacements and tag markers need the documented domain rules.
• E716: mismatched terminator values.
• E720: MOR chunk count differs from GRA relation count, including the MOR terminator and post-clitics.

E707 is a historical presence-mismatch category, not a currently emitted raw-input MOR alignment diagnostic. The retained E707 examples are subsumed by E305 because missing terminators are handled before that alignment check.

See dependent tiers for tier domains and GRA checks, and the validation error guide for diagnostic lookup.

JSON Serialization

Features serialize as strings whether flat or keyed. The terminator is an object, not a punctuation string. This selected-field example illustrates one MOR item; it is not a complete transcript:

{
  "tier_type": "Mor",
  "items": [
    {
      "main": {
        "pos": "noun",
        "lemma": "cookie",
        "features": ["Plur"]
      }
    }
  ],
  "terminator": { "type": "period" }
}

Use the revision-matched generated schema for complete shapes and the JSON integration guide for validation and import boundaries. Preserving surface strings does not promise lossless conversion of arbitrary UD or traditional MOR analyses.

This page last changed: 2026-10-20 (commit f7e32e22). The whole book last changed: 2026-10-27 (commit 5bfee2f9).