INF5820 Natural Language Processing - NLP. H2009 Jan Tore Lønning [email protected]

Size: px

Start display at page:

Download "INF5820 Natural Language Processing - NLP. H2009 Jan Tore Lønning [email protected]"

Meagan Watson
10 years ago
Views:

1 INF5820 Natural Language Processing - NLP H2009 Jan Tore Lønning [email protected]

2 Semantic Role Labeling INF5830 Lecture 13 Nov 4, 2009

3 Today Some words about semantics Thematic/semantic roles PropBank & FrameNet Role labeling

4 What is the goal of NLP? Applications: (semantic) search Summarization Translation Man-machine interaction, e.g. GPS Semantics Grammars and parsing only a step on the way

5 Computational semantics Choose adequate semantic representations for utterances Compute representations from utterances Process representations Generate sentences from representations

6 Semantikk Logikkbasert: x( flyavgang (x) fra(x, oslo, t1) til(x,bodø, t2) tirsdag(t1)) Rammebasert FLY: AVGANG: ANKOMST: BY: BY: oslo DATO: UKEDAG: tirsdag bodø

7 Alternative representations

8 More representations

9 Core Married(adam,eve) Predicate and arguments Logic: Core = atomic formulas The atomic formulas are unstructured In addition: connectives and quantifiers: x (Student(x) Live_in(x, oslo) Happy(x))

10 In addition to first-order logic Extended logic: Adjectives: small elephant, former president Adverbs: ran fast Propositions as arguments: believes the earth is flat etc. Time and change: built a house, was president Events Co-reference: The foreign minister met the president. He told her..

11 Alternative representations Classic logic: Married(adam,eve) Davidsonian: e(married(e,adam,eve) Neo-davidsonian, alternative role levels: 1. e(married(e) & SUBJ(e, adam) & OBJ(e, eve)) 2. e(married(e) & ARG0(e, adam) & ARG1(e, eve)) 3. e(married(e) & AGENT(e, adam) & THEME(e, eve)) 4. e(married(e) & Marrier(e, adam) & Marriee(e, eve))

12 Today Some words about semantics Thematic/semantic roles PropBank & FrameNet Role labeling

13 Thematic/semantic roles Fine-structure of the core: predicate-argument Deep syntax/shallow semantics Theta roles for syntactic roles Thematic roles for semantic counterpart

14 Thematic roles Kari ga Ola en bil AGENT BEN THEME Does not correspond to syntactic function Kari ga en bil til Ola AGENT THEME BEN En bil ble gitt Ola av Kari THEME BEN AGENT Ola ble gitt en bil av Kari BEN THEME AGENT

15 Common roles

16 Role examples

17 Good for what? Linguistics: Generalizations: classes of verbs with similar patterns Alternations, e.g. dative shift Hierarchy of roles: Relationship to syntactic functions NLP: Simple inferences Representations for machine translation

18 Problems Problems: Which roles are there? No agreement How to decide on the particular roles? Fixes: Role types are not firm classes but prototypical: more and less clear-cut instances Two levels: Proto-roles: proto-agent, proto-patient Finer roles

19 Levin s verb classes In which construction types can a particular verb occur? Kim broke the window The window broke Glass breaks easily Similarly: shatter, smash Not: cut Verbs with same patterns classified together Tried to classify (all) English verbs

20 Today Some words about semantics Thematic/semantic roles PropBank & FrameNet Role labeling

21 PropBank Shallow semantic annotation of the Penn treebank Focus on semantic roles Not: quantifiers, co-reference etc.

22 PropBank cont. Uses simple roles: Arg0, Arg1, Arg2, etc. Relates to Levin s classification Roles consistent across a frameset

23 FrameNet Fillmore, Berkeley Deeper roles Semantic network, hierarchy

24 Today Some words about semantics Thematic/semantic roles PropBank & FrameNet Role labeling

25 Role labeling 1. Finding the constituents that are arguments to a predicate in a sentence 2. Determining their role Supervised learning PropBank or FrameNet or

26 Gildea & Jurafsky, 2000, 2002 Path-feature NP S VP VBD

27 Features Predicate: issued + Jurafsky and Martin Phrase-type: NP (or NP-SBJ) + + Headword: Examiner + + Headword POS-tag: NNP + Path: NP S VP VBD + + Voice: active + + Position: before + + Subcategorization: VP NP PP + Palmer et al.

28 Smoothing

29 Results (Palmer et al)

30 Alternative strategies String Chunking Role labeling Role Struct. Tagging PCFGparsing Tree Role labeling Role Struct. Dependencyparsing Dep. Struct. Deep parsing Semantic structures Ranking

CINTIL-PropBank. CINTIL-PropBank Sub-corpus id Sentences Tokens Domain Sentences for regression atsts 779 5,654 Test

CINTIL-PropBank. CINTIL-PropBank Sub-corpus id Sentences Tokens Domain Sentences for regression atsts 779 5,654 Test CINTIL-PropBank I. Basic Information 1.1. Corpus information The CINTIL-PropBank (Branco et al., 2012) is a set of sentences annotated with their constituency structure and semantic role tags, composed