Syntactic Analysis: Parsing

Question 1

Shallow parsing, aka light parsing, aka chunking

Answer 1

• Runs from shallowest to deepest
• Shallow parsing is the middle level
• The shallowest parsing is from sentence to POS tagging only
• Deepest is a full grammar tree, complicated tree
• Shows how some chunks relate to other chunks
• Full parse tree recognizes first every phrase, phrases can be made out of other phrases, then POS’s
• Shallow parse tree breaks out (or chunks) main noun and verb phrases and prepositional phrases
• Can run a full parse tree in python
• POS tagging is a prerequisite for shallow parsing

Question 2

Why chunk?

Answer 2

• Full parse takes longer
• Full parse might not be needed
• Full parse might not be accurate on User Generated Content
• POS doesn’t group words together so sometimes we need the information about noun phrases, prepositional phrases, etc.
• POS to chunking is just a little more time so why not do it? You’re almost there anyway.
• Often will be doing NP chunking

Question 3

User Generated Content (UGC)

Answer 3

-Written by non-professionals, Twitter, FB, reviews, Reddit, etc.

Question 4

How to chunk?

Answer 4

-Create rules with RegEx for obvious tag patterns

-

Question 5

Chinking

Answer 5

• Parts we don’t want included are called chinks

- NP chunker just pulls NP, everything else is a chink

Question 6

Classifier-Based Chunkers

Answer 6

-TimBL, decision-tree based learning classifier

Question 7

IOB Standard Annotation

Answer 7

Every word is a token
I: Inside a chunk
O: Outside a chunk (a chink)
B: Begins a new chunk

For annotating, followed by type of phrase, e.g., B-NP, I-NP, B-PP, etc.

Question 8

Uses for chunking

Answer 8

-Names Entity Recognition
-Pull out NP chunks
-

Question 9

Full Grammar Parsing

Answer 9

Two types:

1) Constituency Parse, Parse, parse trees, dendrogram, penultimate is a POS tag, interior nodes, middle layer, phrases,
2) Dependency DAG - Directed Acyclic Graph

Question 10

Creating parse trees

Answer 10

Constituency parsers (90% of applications use these)

• CYK algorithm
• Stanford parser
• Link grammar parser

Dependency parser

• MST parser
• MALT parser

Question 11

Uses for parse trees

Answer 11

How to choose dependency parser vs. constituency parser?

• Constituency is often based on language. Languages with strict word order rules like English and German
• Dependency parsers are easier to engineer for languages like Czech.

Type of application

• Thematic extraction or text mining - constituency parser - titles, summaries.
• Question-answering dependency parser. Tells us the object of a verb.

Question 12

Combine lexical and syntactic analysis

Answer 12

Get themes

• Wordnet, get noun word sense, synonyms, hyponyms
• Wordnet, get verb sense, synonyms, troponyms
• Use parser to pull out noun phrases and verb phrases that has the words pulled from WordNet

Improving Sentiment Analysis
-Vocabulary based
-

Question 13

Valence

Answer 13

A number, usually -1 to 1, that indicates how emotional a word is. 0 is neutral. -1 is most negative and 1 is most positive.

Add them up to find out emotional charge.

How does a parser help with this.