Unsupervised Chunking Based on Graph Propagation from Bilingual Corpus

Algorithm 1

Graph-based unsupervised chunking approach.

Bilingual Chunking Induction.

Require: Parallel English and foreign language data 𝐷^𝑒 and

𝐷^𝑐; unlabeled foreign training data Γ^𝑐; English tagger

Ensure: 𝜃^𝑐, a set of parameters learned using a constrained

unsupervised model.

(1) 𝐷^𝑒↔𝑐← word-align-bitext (𝐷^𝑒, 𝐷^𝑐)

(2) 𝐷^𝑒← chunk-tag-supervised (𝐷^𝑒)

(3) 𝐴 (𝐷^𝑒↔𝑐, 𝐷^𝑒)

(4) 𝐺← construct-graph (Γ^𝑐, 𝐷^𝑐, 𝐴)

(5) 𝐺← graph-propagate (𝐺)

(6) Δ← extract-word-constraints (𝐺)

(7) 𝜃^𝑓← chunk-induce-constraints (Γ^𝑐, Δ)

(8) return 𝜃^𝑓