Morphological Segmentation InsideOut Ryan Cotterell Arun Kumar Hinrich

Morphological Segmentation Inside-Out Ryan Cotterell, Arun Kumar, Hinrich Schütze

Old Idea: Surface Morphological Segmentation

unachievability Segment un achiev abil ity PREFIX STEM SUFFIX

Semi-New Idea: Canonical Morphological Segmentation

unachievability Restore unachieveableity Segment un achieve able ity PREFIX STEM SUFFIX

Why is canonicalization useful?

unachievability achievement underachiever achieves

un achiev abil ity achieve ment under achieve s

Are they the same morpheme? ? ? un achiev abil ity achieve ment under achieve s

unachievability achievement underachiever achieves

unachieveableity achievement underachieveer achieves

un achieve able ity achieve ment under achieve s

Canonical segmentations are standardized across words un achieve able ity achieve ment under achieve s

unachievability thinkable accessible untouchable

unachieveableity thinkable accessable untouchable

un achieve able ity think able access able un touch able

un achieve able ity think able access able un touch able

New Idea: Morphology as Parsing

unachievability achievement underachiever achieves

unachieveableity achievement underachieveer achieves

un achieve able ity achieve ment under achieve s

un achieve able ity achieve ment under achieve s

under achieve er

under achieve er

under achieve er

under achieve er

PREFIX STEM SUFFIX under achieve er

Why are trees useful?

Reason 1: Words are ambiguous! Tree Captures Ambiguity! PREFIX STEM PREFIX SUFFIX un lock able “incapable of being locked” “capable of being unlocked” PREFIX STEM SUFFIX un lock able “? ? ? ” Flat Segmentation Doesn’t!

Reason 2: Model Order of Affixation Path of Derivation achieve underachieve c n E d o d e A underachiever e e r T s PREFIX STEM SUFFIX under achieve er

New Resource

Morphological Tree Bank • English • Size

A Joint Model

un achieve able ity unachieveableity unachievability unachieveableity ununachievability achieve able ity Canonical Segmentation Parse Tree Underlying Word (Surface. Form)

How good is the treeunderlying form pair? (s=un achieve able ity, u=unachieveableity) How good is the underlying form-word pair? (u=unachieveableity, w=unachievability)

Inference and Learning • Inference is intractable! • Approximate inference with importance sampling • Decoding also with importance sampling • Learning Ada. Grad (Duchi et al. 2011)

Experimental Results • Key Point: Do trees help segmentation accuracy? – Baseline: flat segmentation model • New Task:

Results

Fin. Thank You!