NLP semantics Points Semantic analysis Semantic markers Case

NLP — semantics Points Semantic analysis • Semantic markers Case analysis • Syntactic patterns • Case lists • An algorithm Quantifier scope A taste of discourse analysis A look at pragmatics CSI 4106, Winter 2005 NLP — semantics, page 1

Semantic analysis may follow parsing: map a parse tree (a syntactic structure) into a representation of meaning (a knowledge structure). Semantics resides at both sides of parsing, and elements of meaning come from words. Lexical knowledge lives in dictionaries. It has two forms. • Morphological and syntactic information about the word: part-of-speech (class), number, case, gender, tense, requirements (for verbs), and so on. • Semantic information about the word, for example, a semantic marker that locates — in a hierarchy of concepts — the concept that the word denotes. CSI 4106, Winter 2005 NLP — semantics, page 2

Semantic markers Suppose that a dictionary entry contains both syntactic and semantic information. (Verb patterns would be unused for other classes. ) For example: lexicon( Word, Class, Synt. Categories, Root, Verb. Pattern, Semantics). The word “ball” could have at least these two entries: lexicon( ball, verb, [inf, pres], ball, trans, [make. Ball]). lexicon( ball, noun, [sg], ball, _unused, [sports. Equipment, dance]). CSI 4106, Winter 2005 NLP — semantics, page 3

Semantic markers (2) Here is a place for the noun senses of “ball” (semantic markers) in a possible hierarchy: physical object. . . artifact natural object. . . equipment tool. . . sports equipment CSI 4106, Winter 2005 event. . . social natural event. . . charity entertainment. . . dance NLP — semantics, page 4

Case analysis It is one of many methods of semantic analysis, based on familiar ideas: recognize a general situation (denoted by a verb) and roles in this situation (denoted usually by noun phrases). Examples of syntactic verb patterns intransitive (subject, verb) Jim laughed. transitive (subject, verb, object) Jim found a penny. bitransitive (subject, verb, indirect object, object) Jim gave a penny to Jill. to-inf (subject, verb, infinitive clause) Jim wanted to laugh. object + to-inf (subject, verb, object, infinitive clause) Jim wanted Jill to laugh. for-object + to-inf (subject, verb, for object, infinitive clause) Jim waited for Jill to laugh. CSI 4106, Winter 2005 NLP — semantics, page 5

Case analysis (2) Examples of semantic verb patterns Agent (if the subject is animate then the subject → Agent) laughed( Jim ) Agent + Object (if the subject is animate then the subject → Agent, the object → Object) found( Jim, penny ) Agent + Object + Beneficiary if the subject is animate then the subject → Agent; if the indirect object is animate then the indirect object → Beneficiary, the object → Object) gave( Jim, penny, Jill ) CSI 4106, Winter 2005 NLP — semantics, page 6

Case analysis (3) Another example of semantic verb patterns Agent + Content (if subject is animate then subject → Agent; subordinate sentence → Content) wanted( Jim, “Jim laugh” ) wanted( Jim, “Jill laugh” ) We need some form of a pointer to the semantic structure for the embedded sentence. [Recall the “boxed” propositions in the conceptual graph notation. ] CSI 4106, Winter 2005 NLP — semantics, page 7

Case analysis (4) Lists of cases used in NLP systems have usually more than a few elements. Here is an example. Participant cases Accompaniment, Agent, Beneficiary, Exclusion, Experiencer, Instrument, Object, Recipient Causality cases Cause, Effect, Opposition, Purpose Spatial cases Direction, Location. At, Location. From, Location. To, Location. Through, Orientation, Order Temporal cases Frequency, Time. At, Time. From, Time. Through, Time. To Quality cases Content, Manner, Material, Measure CSI 4106, Winter 2005 NLP — semantics, page 8

Case analysis (5) A case marker is a syntactic element that signals the presence of a case. A preposition (in, at, from, of, for, . . . ) may mark cases. A position of a noun phrase (subject, direct object, indirect object) also marks case. Subject Agent Jim hit the ball. Experiencer Jim grew hungry as time passed. Instrument The ball broke the window. Cause The wind broke the window with a branch. Indirect object Recipient I threw the dog a ball. Beneficiary I wrote her a reference letter to her boss. Direct object Object John hits the ball. CSI 4106, Winter 2005 NLP — semantics, page 9

Case analysis (6) A few examples of markers that mark exactly one case. Location. Through Manner Location. At Exclusion Opposition Time. From Time. To CSI 4106, Winter 2005 We walked around the courtyard. She acted as my agent last year. Sit beside me. Everyone was pleased except her. They persisted despite my warning. He has been sick since the accident. We worked till dawn. NLP — semantics, page 10

Case analysis (7) Examples of markers that mark many cases: at, for. Direction The deer ran right at the hunters. Location. At I stood at the door. Time. At The case will be heard at noon. Manner The car moves at high speed. Content She is good at arts. Measure It stopped at fifty. Cause She was amazed at his insolence. Location. To Aim for the heart. Direction Run for the train. Content I stand for social responsibility. Time. Through They worked for three hours. Beneficiary I'd walk a mile for them. Purpose This drug is for people with a flu. Measure Sell it for fifty dollars. Cause He received a medal for courage. Recipient This mail is for everyone. Time. At Call him for ten o'clock. CSI 4106, Winter 2005 NLP — semantics, page 11

Case analysis (8) An algorithm of case analysis 1. In the parse tree, identify all case markers. 2. Find case patterns of the main verb (assume a knowledge base of patterns!). 3. Apply rules — based on lexical, syntactic and semantic features — to match case markers with cases. Examples of rules [see slides 6 -7 for more] active sentence, animate subject: subject → Agent Jim laughed. passive sentence, inanimate subject: subject → Object The window was broken. passive sentence, animate subject: subject → Experiencer Jim was detained. CSI 4106, Winter 2005 NLP — semantics, page 12

Quantifier scoping Every author wrote a book. a b author(a) book(b) wrote(a, b) skolemize: a author(a) book(s(a)) wrote(a, s(a)) b a author(a) book(b) wrote(a, b) skolemize: a author(a) book(B 0) wrote(a, B 0) Only one scoping is correct: which one? The man picked up all papers. THE m p man(m) paper(p) picked. Up(m, p) p THE m man(m) paper(p) picked. Up(m, p) A simple algorithm: fixed precedence, for example, the > each > what, whom > every, all, some, a But: there is no universally approved, objective ordering. CSI 4106, Winter 2005 NLP — semantics, page 13

A taste of discourse analysis Text units beyond sentences — examples • A story (such as a fairy tale, a drama, . . . ). • A news item. • Dialogue. • Technical text (manual, textbook, documentation). • A document in a document base (abstract, patent description, . . . ). Links between sentences/phrases in a larger text • Textual ordering. • Temporal link (for example, an event precedes another event). Jim saw the bus. He ran to catch it. “saw” precedes “ran” CSI 4106, Winter 2005 NLP — semantics, page 14

. . . discourse analysis (2) • Causal link (for example, reason, effect, prerequisite). Jim saw the bus pull away. He waved to the driver. “waved” could be an effect of “saw” • Coreference: linking references to the same entity. Jim bought a book. He liked it a lot. “he” = Jim, 'it' = book (and “bought” precedes “liked”) Jim bought a book. The price was good. price is a property of books (and it enables buying) Jim bought a book. He paid $10. paying is an element of (is included in) buying Jim bought a book. The dust-jacket was red. dust-jackets are parts of books CSI 4106, Winter 2005 NLP — semantics, page 15

A look at pragmatics Focus Here is one tiny example from a hypothetical NLP interface to an airline reservation system: I want to fly to Vancouver tomorrow night. There is a flight at 6. When does it arrive? At 8 local time. Is it West. Jet? No, Air Canada. Show me others. ← shift of focus Modelling beliefs: who knows what, who believes what. This can be done formally, in advanced forms of logic, for example in autoepistemic logic (check it out). CSI 4106, Winter 2005 NLP — semantics, page 16

. . . pragmatics (2) Plan-based understanding We can use scripts (see textbook, section 7. 1. 4). Jim was hungry. He stopped at Nate’s deli. A possible line of reasoning: hungry needs food buys food at at restaurant market Scripts (and other similar representations of plans) help fill gaps in the story. CSI 4106, Winter 2005 fast food deli formal burgers NLP — semantics, page 17

. . . pragmatics (3) Speech acts assert—inform—explain; ask if—ask what; order—request. Indirect speech acts The form disagrees with the intention: a question (interrogative) or a statement (declarative) really means something different. Could you pass the salt? a request Do you know that it’s raining? information Honey, Fido needs a shower. a command . . . time out — and there is still so much to tell. . . CSI 4106, Winter 2005 NLP — semantics, page 18