Week 7: dependency syntax

Question 1

What type of constituents are difficult for constituency syntax?

Answer 1

Discontinuous constituents

Question 2

What does dependency syntax limit analysis to?

Answer 2

Relationships between words

Question 3

What is the standard assumption of dependency syntax?

Answer 3

Each word has a single parent word

Question 4

Reasons dependencies are easier for computational linguistics (three)

Answer 4

1. Conceptually simple
2. Good cross-linguistically
3. Very little theoretical baggage

Question 5

UD fundamentals 1:
UDs are …

Answer 5

…word-based, where words are whatever the tokeniser gives you

Question 6

UD fundamentals 2: _______ are more important than ________

Answer 6

Content words are more important than function words

Question 7

UD fundamentals 3:
In symmetric cases, …

Answer 7

…draw edges from left to right

Question 8

UD fundamentals 4:
If a word is elided, …

Answer 8

…promote its child to the head position. If the result is “unnatural and misleading”, use the orphan relation.

Question 9

What is the main predicate of the sentence labelled as?

Answer 9

root

Question 10

What word class is normally the root?

Answer 10

Verb (not always though!)

Question 11

What are non-projective trees?

Answer 11

Trees with crossing edges

Question 12

Why are non-projective trees problematic? (Two reasons)

Answer 12

1. Not enough training data - many treebanks were converted from earlier constituency treebanks without non-projective trees
2. Some parsing algorithms cannot produce non-projective trees

Question 13

How do you go from a constituent parse to dependencies?

Answer 13

1. Replace the label of each constituent with its head word
2. Attach all word in the constituent to the head word

Question 14

How do you go from dependencies to constituents?

Answer 14

1. Identify left and right boundaries for the subtree, add parentheses and choose a label
2. Proceed recursively with each child

Question 15

Limitations of dependencies (three)

Answer 15

1. Constituency testing is impossible and competing analysis are harder to justify
2. We cannot maintain binary branching - all tree nodes are now surface words so there can be no hidden nodes
3. Dependency trees are unordered (equivalent under a permutation of indices) whereas constituents are naturally ordered.

Question 16

UD specific problems (three)

Answer 16

1. Tokenisation approaches are not uniform
2. Guidelines are not uniform
3. Annotation quality is not held to the same standard everywhere