FrontPage
- Barriere's (2009) TerminoWeb?
- Normallized MI --> frequency divided by the number of senses in wordnet
- Greenstone/FLAX Web Phrases
Keynote †
- Magnetoencephalography (MEG)
- word recognition
- visual cortices, Occipital Lobe --> abstract sign recognition
- lemma/lexical access
- recomposition of stem and suffix
Keynote2: Prof. Wen-Lian Hsu (Academia Sinica) †
- The archaelogist is in trouble ([people: [be]: [in a situation])
- The old archaelogist named John Doe is in a very serious trouble with his academia reputation.
- phrases --> classes (NP, PP, AdjP)
- Simple phrase: word, word combination
- Complex phrase: named entity frame
- Scenerio frame:
- Frame elements
- How to score an Insertion?
- Principle Pattern Analysis
- 60 patterns out of all the sentence structures will cover approximately 98%.
- Using that framework, you can do the insertion/deletion operations over the pattern.
Keynote by Tsujii Junichi (Microsoft Research, Beijing) †
- Language acquisition --> skimming huge amount of data --> wrong?
- Semantic trigger (Pinker) --> certain contexts
- text <--> knowledge (bootstrapping)
- Innate ability (Chomsky)
- grammar (linear sequence --> structure --> meaning)
- rationalism vs. empiricism
- Ken Church (2011): A pendulum swung too far toward empiricism/data
- Integrating [grammar, rule] into [data]
- Knowledge mining
- Triggering: Big knowledge to enhance NLP abilities
- entity linking
- relation extraction
- Entity linking
- MS --> Microsoft or Multiple sclerosis
- entity relation in knowledge domain (human knows all these relations)
- categories in knowledge
- How to use those structural knowledge to disambiguate the meaning of words in a certain context
- Build classifiers, looking at the context and tell which a certain word belongs to a particular entity
- Inference relationship among different cues with different strength.
- Semantic parsing
- Identify grammatical relationships
- Extract entity relationships
- Parsing: Current state of the arts
- F-value too high
- Most relations are very simple (DET-N, ADJ-N) and trivial
- Non-trivial one: PP attachment
- accuracy is around 80-85%
- semantically crucial problems are not properly solved by current parsing techniques
- Fragments of parsing are compared to each other and compared in terms of similarities.
- inferences involved --> difficult
- statistical models of inferences have become scalable and feasible.
- integration of NLP with knowledge-based inferences is becoming a practical framework of natural language understanding.
|