Outline Communication Symbolic Natural Language Processing - PDF document

Outline • Communication • Symbolic Natural Language Processing Communication • Reading: R&N Sect. 22.1-22.6 July 13, 2006 CS 486/686 University of Waterloo 2 CS486/686 Lecture Slides (c) 2006 P. Poupart Communication Turing Test • Communication: intentional exchange of • Can a computer fool a human to think information brought about by the that it is communicating with another production and perception of signs human? drawn from shared system of convention. • Language: – Enables us to communicate – Intimately tied to thinking 3 4 CS486/686 Lecture Slides (c) 2006 P. Poupart CS486/686 Lecture Slides (c) 2006 P. Poupart Speech Components of Communication • Speech: communication act • Intention – Speaker S decides that there is some proposition P – Talking worth saying to hearer H. – Writing utterances • Generation – Facial expression – Speaker plans how to turn proposition P into an – Gesture utterance (i.e. a sequence of words W) • Synthesis situation – Speaker produces the physical realization W’ of the words W (i.e., vibration in air, ink on paper) utterances Speaker Hearer 5 6 CS486/686 Lecture Slides (c) 2006 P. Poupart CS486/686 Lecture Slides (c) 2006 P. Poupart 1

Components of Communication Components of Communication • Disambiguation • Perception – Hearer perceives physical realization W’ as W 2 and – Hearer infers that speaker intended to decodes it as the words W 2 (i.e., speech convey P i (where ideally P i = P). recognition, optical character recognition) • Incorporation • Analysis – Hearer decides to believe P i (or not). – Hearer infers W 2 has possible meanings P 1 , P 2 , …, P n – Three parts: • Syntactic interpretation • Semantic interpretation • Pragmatic interpretation 7 8 CS486/686 Lecture Slides (c) 2006 P. Poupart CS486/686 Lecture Slides (c) 2006 P. Poupart Components of Communication Difficulties SPEAKER • How could communication go wrong? Synthesis: Intention: Generation: "The wumpus is dead" [thaxwahmpaxsihzdehd] Know(H, Alive(Wumpus,S )) L – Insincerity 3 – Speech recognition errors HEARER – Ambiguous utterance Perception: Analysis: Disambiguation: S (Parsing): – Different contexts "The wumpus is dead" L Alive(Wumpus,S ) NP VP 3 Article Noun Verb Adjective The wumpus is dead Incorporation: (Semantic Interpretation): L Alive(Wumpus,Now) Tired(Wumpus,Now) TELL( KB, L Alive(Wumpus,S ) (Pragmatic Interpretation): 3 Alive(Wumpus,S ) L 3 Tired(Wumpus,S ) 3 9 10 CS486/686 Lecture Slides (c) 2006 P. Poupart CS486/686 Lecture Slides (c) 2006 P. Poupart Language Grammar • Formal language • Grammar specifies the compositional structure of complex messages – Set of strings of terminal symbols (words) – Strict rules • Each string in a language can be analyzed/generated by the grammar – E.g., first order logic, Java • A grammar is a set of rewrite rules • Natural language – S � NP VP – No strict definition – Article � the | a | an | … – Chinese, Danish, English, etc. 11 12 CS486/686 Lecture Slides (c) 2006 P. Poupart CS486/686 Lecture Slides (c) 2006 P. Poupart 2

Grammar Types Grammar Types • Regular grammar: • Context sensitive grammar: – nonterminal � terminal [nonterminal] – More symbols on left-hand side – S � a S – ASB � AAaBB – S � b • Recursively enumerable grammar: • Context free grammar (CFG): – No constraints – nonterminal � anything – S � aSb 13 14 CS486/686 Lecture Slides (c) 2006 P. Poupart CS486/686 Lecture Slides (c) 2006 P. Poupart Lexicon example Grammar example • Noun � breeze | glitter | agent • S � NP VP | S Conjunction S • Verb � is | see | smell | shoot • NP � Pronoun | Name | Noun | Article Noun | NP PP | NP RelClause • Adjective � right | left | east | dead • VP � Verb | VP NP | VP Adjective | VP • Adverb � there | nearby | ahead PP | VP Adverb • Pronoun � me | you | I | it • PP � Preposition PP • Name � John | Mary | Boston • RelClause � that VP • Article � the | a | an 15 16 CS486/686 Lecture Slides (c) 2006 P. Poupart CS486/686 Lecture Slides (c) 2006 P. Poupart Grammaticality Judgements Grammaticality Judgements • Overgeneration examples: Set of strings – Me go Boston. Natural – I smell pit gold wumpus nothing east. Grammar language • Undergeneration example: agreement – I think the wumpus is smelly Goal: design grammar to match natural language 17 18 CS486/686 Lecture Slides (c) 2006 P. Poupart CS486/686 Lecture Slides (c) 2006 P. Poupart 3

Syntactic Analysis Top-down parsing • Parsing: process of finding a parse tree • Start with S and search for a tree that for a given input string has strings at leaves S S NP NP VP VP NP NP proposition proposition noun noun pronoun verb pronoun verb I the I the shoot wumpus shoot wumpus 19 20 CS486/686 Lecture Slides (c) 2006 P. Poupart CS486/686 Lecture Slides (c) 2006 P. Poupart Bottom up parsing Parsing efficiency • Start with string and search for a tree • Top-down and bottom up parsing that has S as root inefficient… – Exponential running time S • Alternative: chart parsing NP VP NP – Dynamic programming proposition noun – Cubic running time pronoun verb I shoot the wumpus 21 22 CS486/686 Lecture Slides (c) 2006 P. Poupart CS486/686 Lecture Slides (c) 2006 P. Poupart Augmented Grammars Parse ambiguity • Grammars tend to overgenerate • Some sentences have many grammatical – Ex: “me eat apple” parses • Augment grammar to require • Example: – Agreement between subject and verb – “Fall leaves fall and spring leaves spring” • Ex: “I smells” vs “I smell” – Agreement between verb subcategory and complement • Ex: “give the gold to me” • Ex: “give me the gold” 23 24 CS486/686 Lecture Slides (c) 2006 P. Poupart CS486/686 Lecture Slides (c) 2006 P. Poupart 4

Semantic Interpretation Ambiguity • Extract meaning from utterances • Possible causes: – Metonymy: figure of speech in which one • Traditional approach object is used to stand for another – Express meaning with logic – Metaphor: figure of speech in which a phrase with one literal meaning is used to • Problem suggest a different meaning by analogy – Ambiguous semantics – Vagueness – Ex: “Helicopter powered by human flies” – Unknown context 25 26 CS486/686 Lecture Slides (c) 2006 P. Poupart CS486/686 Lecture Slides (c) 2006 P. Poupart Next Class Context/Experience • Meaning often grounded in experience • Next Class: •Probabilistic Language Processing •Russell and Norvig Ch. 23 • But humans and machines have different experiences because of different sensors… • Is that a problem for natural language understanding? 27 28 CS486/686 Lecture Slides (c) 2006 P. Poupart CS486/686 Lecture Slides (c) 2006 P. Poupart 5