Application Area: Natural Language Processing Natural communicaton Issue: producing meaning from structure More than symbol manipulations “out of sight, out of mind” “invisible idiot” “I saw the car by the house with a chimney” “I saw the car by the house with a flat tire”
NLP: A Target Behavior Task: Given a file of data containing test results, calculate the relevant statistics about each person and display and save a table of results in sorted order based on average
Command Line Solution stats < data.txt | sort –k 3,3 n > results.txt ; lpr results.txt
GUI and App Solution Select program based on icon location Select run action from hierarchical menus Specify input file in dialog box Reorder displayed output by clicking on column header once or twice Select “save as” and type file name in dialog box Select print menu option and click and type until desired format is chosen
NLP/Intelligent Agent Solution “<name>, I need stats run for the skills tests. Please sort them by average and display and save them into a file”
NLP: Context Matters What image do you have? “the prince threw a ball” “Cinderella” “caught” “a cold” “and couldn’t run fast enough to catch it”
NLP: Context and Meaning Subtlety Statement: (to in-car navigational system) “I need to get to the airport quickly” Context: You are in the New York City area 3 airports “quickly” Lowest mean Lowest standard deviation Purpose
NLP: Context and Meaning Subtlety Advance Planning Statement: (to telephone assistant for public transportation) “I need to get to the airport quickly and cheaply” “What’s a [fast|quick|short] route to the airport? “How do I get to the airport as [fast|quickly] as possible? “Can you tell me a [fast|quick|short] route to the airport.” “It would be nice to get to the airport quickly.
Traditional NLP Approach Build parse tree of structure Define algorithms to traverse tree and produce meaning structure Symbolic in nature
Traditional NLP Approach (Semantic Grammar) Blend parsing and meaning together Useful in precise and limited domain Example: Circuit Fix-it Shop: “the switch is up” assertion(position,switch,up,true) “the switch is not down” assertion(position,switch,down,false) Issue: canonical form
Statistical NLP Approaches (Normally) requires annotated corpus of data that is “correct” Use machine learning techniques for rules for identifying structures/meaning based on features in the data Human intelligence required for annotation, and feature identification
Statistical NLP: Word Sense Disambiguation Is “bass” about a fish or about music? What if a nearby word is? striped guitar piano sea violin are
Applicaton: Proverb Duke University probabilistic crossword solver Set of expert modules that each try to find a solution for a given clue Choices combined probabilistically to choose best response 85 to 95% accuracy on New York Times puzzles (1998)
Application: Watson IBM computerized Jeopardy solver (deep question answering) Lots of context Category: Rhyme Time Clue: It’s where Pele stores his ball Answer: Soccer locker Combines probabilistic and symbolic NLP
Watson (cont.) Use understanding of clues to retrieve documents that may contain the answer Expert modules probabilistically combined 2 hours CPU on a single machine
NLP References James Allen, Natural Language Understanding, 1995. Daniel Jurafsky and James Martin, Speech and Language Processing, 2009.