DS-to-PS conversion Fei Xia University of Washington July 29, 2011 1
Main steps in building the treebank • DS treebank: – Tokenization – Morphological analysis, voice, etc. – POS tagging – DS • Propbank: adding Predicate-argument info • Automatic DS-to-PS conversion • Some manual check to ensure the conversion works well 2
Outline • Important concepts • Compatibility and consistency • Handling inconsistency 3
Important concepts • Linguistic phenomena • Representation type • Linguistic theory – Theoretical framework – Linguistic analyses • Annotation guidelines 4
Linguistic phenomena • They are what we want to present, including – General concepts: e.g., which words form a phrase? What types of phrases does a language have? – Types of relations between words or phrases (e.g., subjecthood, temporal modification) – Specific constructions (e.g., small clause) – Finer-grained distinctions (e.g., unergative vs. unaccusative) 5
Representation type • It is the type of mathematical object that is used to represent syntactic facts • Examples: DS, PS • Each representation type can decide what more specific representation devices to employ – Labels on the arcs of a tree – Use of empty nodes or coindexation between nodes 6
Linguistic theory • It explains how linguistic phenomena are represented in the chose representation type • It has two components: – Theoretical framework: it provides vocabulary and constraints in which linguistic theories can be formulated: e.g., GB, LFG, LTAG, HPSG – Linguistic analyses 7
Small clause 8
“Exceptional case - marking” analysis 9
“Raising -to- object” analysis 10
Annotation guidelines • Guideline designers need to choose the following – Linguistic phenomena to represent – Representation type – Theoretical framework – Linguistic analyses – Descriptions – Examples: sentences with DS or PS trees 11
Outline • Important concepts • Compatibility and consistency • Handling inconsistency 12
“Exceptional case - marking” analysis 13
“Raising -to- object” analysis 14
Implicit vs. explicit information • Certain aspects of information has to be expressed explicitly in DS, but not PS, or vice versa – Head in DS – Syntactic categories of phrases in PS • Not explicitly providing info does not mean that corresponding concepts does not exist in DS/PS 15
Syntactic consistency • We assume each phrase in a PS has a special word, head word, which represents the property of the phrase. • A (DS, PS) pair is called consistent if there is a way to assign a head word to each internal node in the PS so that the resulting DS is identical to the given DS. 16
Consistent pairs 17
Inconsistent pairs 18
A real example 19
Consistency assumption 20
Definition of consistency • A DS and a PS are consistent iff there exists a flattened version of the PS that is identical to the DS. • If the input DS and the desired PS are consistent, the PS can be created by stretching the DS and adding syntactic labels. 21
Checking consistency • For each (dep, head) pair in the DS – find their location in the PS and their closest antecedent – add heads to the nodes on the path between the leaf nodes and the antecedent • The DS and the PS are consistent iff each node in the PS has exactly one head. 22
(Vinken, join) (will, join) (board, join) (29, join) (join) (join) (Vinken) (join) (board) (29) 23
Outline • Important concepts • Compatibility and consistency • Handling inconsistency 24
wh-movement come come come come come come (who, come) come (come, think) 25
wh-movement come come | think come | think come come come (who, come) come (come, think) 26
wh-movement ?? think think ?? come come (who, come) come (come, think) (you, think) 27
Can DS and PS be inconsistent? • DS and PS can represent different aspects of the same overall pictures, and still be consistent. – Info provided in PropBank: e.g., empty subject, unaccusative – Info that is in PS only: e.g., traces • DS and PS should not choose “conflicting” analyses. – DS and PS are two images of the same underlying treebank, not two separate treebanks. – Ex: ba-construction in Chinese: verb, prep, or something else? – Ex: free relatives: empty nominal head • The inconsistency cases should be rare and well-motivated. 28
How to handle inconsistency? • Detect inconsistency in (DS, PS) pairs in the guidelines • Consult guideline designers to determine whether the inconsistency can be resolved by changing analyses • If not, introduce DS cons and ensure sufficient info is in DS for automatic conversion. 29
Two-stage conversion • DS to DS cons : by removing “inconsistency” between DS and PS. • DS cons to PS: by applying conversion rules 30
Case #1: long-distance movement DS prop: DS const: • Other examples: extraposition • Easily detectable due to non-projectivity • Create DS const by moving up the “moved element” and leaving a trace • which node is the “moved element”? The one that is apart from other nodes in the subtree. 31
Case #2: local scrambling Detectable by assuming canonical word order: k1 > k2 Need from PS/DS teams the canonical word order and what word order triggers movement 32
Case #3: small clause rule Detectable by dependency type k2s Need confirmation from IIIT that k2s is used only for small clause 33
Case 4: support verb Detectable by dependency type “pof” Need confirmation from IIIT that “pof” is used only for support verb 34
Conclusion • We define consistency between DS and PS • DS and PS can be inconsistent but such cases should be rare and well-motivated. • We will handle inconsistency with the two- stage approach 35
Conversion algorithm 36
Definition of conversion rule • A conversion rule is a (DS_pattern, PS_pattern) pair. • Ex: • Simplest case: – DS_pattern corresponds to only one dependency link – Decomposing DS becomes trivial – PS_pattern is a tree fragment (e.g., wh-movement) – Learning rules from (PS, DS) pairs is easy 37
Extracting rules 38
Rules extracted from the example 39
Input DS 40
41
Gluing PS segments together 42
c c c 43
44
Recommend
More recommend
