Hoorcollege 6 en 7 semantic parsing

Question 1

Semantic parsing

Answer 1

How do the meaning of words go together to make the meaning of a setence
* Semantic parsing as ML problem
Not good at compositionality, which is what semantics is all about

Question 2

AM dependency trees

Answer 2

*turns problem into parsing + supertagging
* supertagging: use NN to guess best graph fragments for each word

To get correct AMR:
* Need functor-argument structure
* Don’t need order*

BILSTM
* 2 recursive neural networks (one in each direction)
LSTM
* Long Short-Term Memory” Recursive Neural Network: a kind of RNN that’s good at remembering info for a long time

Question 3

Dependency parsing with a BiLSTM

Answer 3

Use word encodings from the supertagging BiLSTM to train another BiLSTM to predict AM-algebra operations between words

What we get from the NN:
* Supertagger: For a word in the sentence:
a prob distribution over graph constants (plus none)
* Edge model: For 2 words in the sentence:
a prob distribution over AM operations (plus none)

Question 4

Type system

Answer 4

graphs have types, which consist of:
● open slots
● the type that each open slot wants the graph that fills it to have

• Intransitiveverb: [S]
• No open slots: []
• Transitive verb: [S, O]
• Control verb: [S, O[S]]

Question 5

AM Dependency parsing

Answer 5

Evaluation by finding best well-typed AM dependency tree and evaluating it to a graph is
slow because type system makes parsing NP-complete and two approximate unpruned
chart parsers still slow

Question 6

Transition-based parsing

Answer 6

Transitioned-based parsing is faster than chart parsing: draw all outgoing edges of node,
determine graph constant, recursively continue with all children
* top down parsing
* no projectivity constraints
* Run-time is O(n^2)

Question 7

Lexical types and term types

Answer 7

• Lexical type: type of the graph constant.
• Term type:
  type of graph that the subtree evaluates to.

Question 8

Theoretical guarantees

Answer 8

• Soundness: Every derived AM dependency tree is well-typed.
• Completeness: All well-typed AM dependency trees can be derived.
• No dead ends: Every parser configuration can be completed.