Phrase & clause chunking

YAML-driven shallow chunker: maximal munch (longer span wins; priority breaks ties), then entity locks as NP. Clause types come from clause_rules.yml via ClauseParser, not from phrase_rules.yml.

from burmesenlp import BurmeseNLP, ClauseType

nlp = BurmeseNLP()
doc = nlp.process("မိုးရွာလျှင်အိမ်မှာနေမည်။")

for sent in doc.sentence_trees:
    for clause in sent.clauses:
        print(clause.type, clause.text)
        for phr in clause.phrases:
            print(" ", phr.type, phr.text, phr.semantic_role)

Pipeline order

POS → Phrase chunking → Sentence segmentation → Clause parsing

ClauseParser consumes phrase chunks (plus markers from YAML), never raw words alone. Relative constructions nest inside NP.

Hierarchy

Document
 └── SyntaxSentence          # doc.sentence_trees
       └── Clause            # MAIN | CONDITIONAL | …
             └── Phrase      # NP / VP / PP / …
                   └── children   # RelativeModifier + HeadNoun

Flat doc.chunks remain phrase spans for backward compatibility.

Clause types (V1)

Type	Markers (see YAML)
MAIN	end markers / remainder with VP
CONDITIONAL	လျှင်, ရင်, လျှင်မူ
REASON	ကြောင့်, သဖြင့်, …
PURPOSE	ရန်, အောင်, ဖို့ (requires NP before VP)
CONTRAST	သော်လည်း, သော်, …
RELATIVE	nested in NP only (သော, တဲ့, …)

Bare VP + ရန် becomes a Purpose VP (role=PurposeVP), not a PURPOSE clause.

Grammar files

Under src/burmesenlp/corpus/grammar/:

• phrase_rules.yml / phrase_markers.yml / phrase_exceptions.yml
• clause_rules.yml — markers + phrase patterns
• postposition_roles.yml — default semantic roles for PPs (V1 assigns default only)