Logic, Language and Learning

1
Logic, Language and Learning

• Chapter 12
• Logic and Language
• Wrapping up
• Luc De Raedt

2
Language

• What we have not done so far
• Linguistics aspects have been simplified
• Semantics / Reasoning
• Translation
• Disambiguation
• Probabilistic/learning approaches
• ...
• Many other points

3
Linguistic issues

• Developing proper NLP systems requires much more
  complicated grammars to model all aspects of
  Language
• An overview is given in Chapter 4 of Covington
• Part of Speech Tagging
• Grammatical Relations
• Cases
• Etc.

4
Part of speech tagging

• What ?
• Given a sentence
• Find tags (word cat) of all words
• Approach often statistical/probabilistic

5
Grammatical Relations

• Need to analyze relationships among constituents
• Grammatical relations
• Table 4.2.

6
Cases

• Cases
• Cf. German
• Nominativ, Dativ, Akkusativ

7
Semantics

• Mapping Sentence -gt Logical Form
• Some (very simple) illustrations from Flachs
  book
• Use of lambda abstraction
• Tables 7.2,3

8
Semantics

• Use Logical Form for Reasoning
• Assumes domain/world/background knowledge exist
  in the knowledge base
• NL understanding
• requires a lot of domain knowledge
• requires many artificial intelligence components
• Planning
• Learning
• Reasoning
• Knowledge representation
• Etc.
• Often the bottleneck, cause of brittleness of NLP
  system
• Dependent on goal of NLP system
• Information retrieval
• Translation
• Dialogues

9
Translation

• Map sentence in language1 to sentence in
  language2
• Two approaches
• Direct mapping
• Map to interlingua (e.g. logical form)
• Advantage of Prolog/UBG for interlingua
• Grammars are bi-directional
• s(What,every,dog,barked)
• What(all(X,dog(X),barked(X))
• s((all(X,dog(X),barked(X)),S)
• S every,dog,barked
• Use two bi-directional grammars with same Logical
  Form

• Table 8.1

10
Disambiguation

• One of the key problems in NLP
• Structural disambiguation
• Empirical approach
• Learn probabilistic CFG grammars
• VP -gt V NP PP (0.3)
• VP -gt V NP (0.4)
• VP -gt V (0.3)
• Focus on most probable parses
• Word sense disambiguation

• Fig 4.2

11
Word Sense Disambiguation

• What ?
• Words do not always have a unique meaning.
• Eg. Bachelor 58, run 54, go 38
• Three types
• Homonymy
• Two words that have no relation but that sound
  alike
• The bark of a tree versus the bark of a dog
• Polysemy
• Different senses but connected
• The kernel of an OS versus The kernel of a nut
• Argument ambiguity
• Mary is cooking (Subjectagent)
• The potatoes are cooking (Subjecttheme)

12
Empirical NLP

• Combining Machine Learning with NLP
• The linguist (Noam Chomsky 1969)
• It must be recognized that the notion of
  probability of a sentence is an entirely
  useless one, under any known interpretation of
  this term
• The statistician (Fred Jelenek 1988, head of IBM
  speech group)
• Anytime a linguist leaves the group the
  recognition rate goes up

13
Empirical NLP

• Since the mid-nineties
• Empirical NLP has become a major trend
• Shallow parsing versus deep parsing
• Importance of Corpora
• E.g. the Wall Street Journal Corpus
• Etc.