Semantics and Context in Natural Language Processing - ICRI-CI

Semantics and Context in Natural
Language Processing (NLP)
Ari Rappoport
The Hebrew University
Form vs. Meaning
• Solid NLP progress with statistical learning
• OK for classification, search, prediction
– Prediction: language models in speech to text
• But all this is just form (lexicon, syntax)
• Language is for expressing meanings
– (i) Lexical, (ii) sentence, (iii) interaction
(i) Lexical Semantics
• Relatively context independent
• Sensory feature conjunction: “house”,
“hungry”, “guitar”
– Non-linguistic “semantic” machine learning: face
identification in photographs
• Categories: is-a: “a smartphone is a (type of)
product”, “iPhone is a (type of) smartphone”
• Configurations: part-of: “engine:car”, places
Generic Relationships
• Medicine : Illness
– Hunger : Thirst
– Love : Treason
– Law : Anarchy
– Stimulant : Sensitivity
• “You use X in a way W, to do V to some Z, at a
time T and a place P, for the purpose S,
because of B, causing C, …”
Flexible Patterns
• X w v Y : “countries such as France”
– Davidov & Rappoport (ACL, EMNLP, COLING, etc)
Content words, High frequency words
Meta-patterns: CHHC, CHCHC, HHC, etc.
Fully unsupervised, general
Efficient hardware filtering, clustering
Categories, SAT exams, geographic maps,
numerical questions, etc.
• Relative context independence does not solve
• Apple: fruit, company
• Madrid: Spain, New Mexico
• Which one is more relevant?
• Context must be taken into account
– Language use is always contextual
(ii) Sentence Semantics
• The basic meaning expressed by (all)
languages: argument structure “scenes”
• Dynamic or static relations between
participants; elaborators; connectors;
– “John kicked the red ball”
– “Paul and Anne walked slowly in the park”
– “She remembered John’s singing”
Several Scenes
• Linkers: cause, purpose, time, conditionality
– “He went there to buy fruits”, “Before they
arrived, the party was very quiet”, “If X then Y”
• Ground: referring to the speech situation
– “In my opinion, machine learning is the greatest
development in computer science since FFT”
[and neither were done by computer scientists]
• “Career”, “Peace”, “Democracy”
Sentence Semantics in NLP
• Mostly manual: FrameNet, PropBank
• Unsupervised algorithms
– Arg. identification, Abend & Rappoport (ACL 2010)
• Question Answering
– Bag of words (lexical semantics)
• Machine Translation
– Rewriting of forms (alignment, candidates, target
language model)
Extreme Semantic Application
• Tweet Sentiment Analysis
– Schwartz (Davidov, Tsur) & Rappoport 2010, 2013
• Coarse semantics: 2 categories (40)
• Short texts, no words lists; fixed patterns
(iii) Interaction Semantics
• ‘Understanding’ means having enough
information to DO something
– The brain’s main cycle
• Example: human-system interaction
• Full context dependence
– Relevance to your current situation
Interaction Examples
• Searching “Argo”, did you mean
– The plot? Reviews? Where and/or when to watch?
• “Chinese restaurant”
– The best in the country? In town? The nearest to
you? The best deal?
• There are hints:
– Location (regular, irregular); time (lunch?)
Interaction Directions
• Extending flexible patterns:
– Include Text-Action H and C items (words, actions)
• Action:
– represented as User Interface operations
• Shortcut: bag of words (lexical semantics) +
“current context”. Ignore sentence semantics
• Noise, failure (Siri, maps,…)
• Lexical, sentence, and interaction semantics
• Applications are possible using all levels
• As relevance to life grows, so do requirements
from algorithms
• Both sentence and interaction semantics
necessary for future smart applications
• Current focus: sentence semantics

similar documents