2. Semantics

3. Where are we in the “Big Picture” Morph Analysis Parsing WORLD of FACTS Inference Engine Semantic Interpreter ASRSpeech Text Syntactic Parse Semantic Representation

4. Semantic Representation Syntactic representation, phrases and tree structures dependency information between words Semantic representation What’s the purpose of this representation? – Interface between syntactic information and the inference engine Requirements on the semantic representation Supports inference – Every CEO is wealthy and Gates is a CEO  Gates is wealthy Normalizes syntactic variations – Delta serves NYC == NYC is served by Delta Has the capacity of representing the distinctions in language phenomena – John believes Delta serves NYC ≠ Delta serves NYC Unambiguous representation – John wants to eat someplace close to the university

5. Mechanisms for Expressing Meaning Linguistic means for expressing meaning Words: lexical semantics and word senses – Delta will serve NYC – This flight will serve peanuts – John will serve as CEO Syntactic information: predicate-argument structure – John wants to eat a cake – John wants Mary to eat a cake – John wants a cake Prosodic information in Speech – Legumes are a good source of vitamins Gesture information in multimodal communication

6. First-order Predicate Calculus: A refresher A formal system used to derive new propositions and verify their truth given a world. Syntax of FOPC Formulae: quantifiers and connectives on predicates Predicates: n-ary predications of facts and relations Terms: constants, variables and functions World: Truth assignments to formulae Inference: Modus ponens – Every CEO is wealthy : ∀ x CEO(x)  wealthy(x) – Gates is a CEO : CEO(Gates) – Derives: wealthy(Gates) Given a world, determining the truth value of a formula is a search process – backward chaining and forward chaining Much like the top-down and bottom-up parsing algorithms.

7. Logic for Language Representations for different aspects of language. Entities – Delta, Gates, AT&T Categories – restaurants, airlines, students Events – I ate lunch. I ate at my desk  I ate lunch at my desk Time (utterance time, reference time, event time) – I ate lunch when the flight arrived – I had eaten lunch when the flight arrived Aspect – Stative, activity, achievement and accomplishment Quantification – Every person loves some movie Predication – John is a teacher Modal operators – John believes Mary went to the movies

8. Linking Syntax and Semantics How to compute semantic representations from syntactic trees? We could have one function for each syntactic tree that maps it to its semantic representation. Too many such functions Not all aspects of the tree might be needed for its semantics Meaning derives from The people and activities represented (predicates and arguments, or, nouns and verbs) The way they are ordered and related: syntax of the representation, which may also reflect the syntax of the sentence Compositionality Assumption: The meaning of the whole sentence is composed of the meaning of its parts. George cooks. Dan eats. Dan is sick. Cook(George) Eat(Dan) Sick(Dan) If George cooks and Dan eats, Dan will get sick. (Cook(George) ^ eat(Dan))  Sick(Dan) The trick is to decide on what the size of the part should be. Rule-by-rule hypothesis

9. Linking Syntax and Semantics – contd. Compositionality: Augment the lexicon and the grammar (as we did with feature structures) Devise a mapping between rules of the grammar and rules of semantic representation For CFGs, this amounts to a Rule-to-Rule Hypothesis Each grammar rule is embellished with instructions on how to map the components of the rule to a semantic representation. S  NP VP {VP.sem(NP.sem)} Each semantic function is defined in terms of the semantic representation of choice.

10. Syntax-Driven Semantics S : fly(birds) NP : birds VP : fly N : birds birds V : fly fly There are still a few free parameters: a.What should the semantic representation of each component be? b.How should we combine the component representations? Depends on what the final representation we want.

11. A Simple Example McDonald’s serves burgers. Associating constants with constituents ProperNoun  McDonald’s {McDonald’s} PlNoun  burgers {burgers} Defining functions to produce these from input NP  ProperNoun {ProperNoun.sem} NP  PlNoun {PlNoun.sem} Assumption: meaning representations of children are passed up to parents for non-branching constituents Verbs are where the action is

12. V  serves { ∃ (e,x,y) Isa(e,Serving) ^ Server(e,x) ^ Served(e,y)} where e = event, x = agent, y = patient Will every verb have its own distinct representation? – McDonald’s hires students. – McDonald’s gave customers a bonus. – Predicate(Agent, Patient, Beneficiary) Once we have the semantics for each constituent, how do we combine them? VP  V NP {V.sem(NP.sem)} Goal for VP semantics: E(e,x) Isa(e,Serving) ^ Server(e,x) ^ Served(e,burgers) VP.sem must tell us – Which variables to be replaced by which arguments – How this replacement is done

13. Lambda Notation Extension to First Order Predicate Calculus x P(x) + variable(s) + FOPC expression in those variables Lambda binding Apply lambda-expression to logical terms to bind lambda- expression’s parameters to terms (lambda reduction) Simple process: substitute terms for variables in lambda expressionxP(x) (car)  P(car)

14. Lambda Abstraction and Application Abstraction: Make variable in the body available for binding. to external arguments provided by semantics of other constituents (e.g. NPs) Application: Substitute the bound variable with the value Semantic attachment for V  serves {V.sem(NP.sem)} { ∃ (e,x,y) Isa(e,Serving) ^ Server(e,y) ^ Served(e,x)} converts to the lambda expression: {x ∃ (e,y) Isa(e,Serving) ^ Server(e,y) ^ Served(e,x)} Now ‘x’ is available to be bound when V.sem is applied to NP.sem of direct object (V.sem(NP.sem)) application binds x to value of NP.sem (burgers) Value of VP.sem becomes: { ∃ (e,y) Isa(e,Serving) ^ Server(e,y) ^ Served(e,burgers)}

15. Lambda Abstraction and Application – contd. Similarly, we need a semantic attachment for S  NP VP {VP.sem(NP.sem)} to add the subject NP to our semantic representation of McDonald’s serves burgers Back to V.sem for serves We need another -abstraction in the value of VP.sem Change semantic representation of V to include another argument to be bound later V  serves {x y ∃ (e) Isa(e,Serving) ^ Server(e,y) ^ Served(e,x)} Value of VP.sem becomes: {y ∃ (e) Isa(e,Serving) ^ Server(e,y) ^ Served(e,burgers)} Value of S.sem becomes: { ∃ (e) Isa(e,Serving) ^ Server(e,McDonald’s) ^ Served(e,burgers)}

16. Several Complications For example, terms can be complex A restaurant serves burgers. ‘a restaurant’: ∃ x Isa(x,restaurant) E e Isa(e,Serving) ^ Server(e, ) ^ Served(e,burgers) Allows quantified expressions to appear where terms can by providing rules to turn them into well-formed FOPC expressions Issues of quantifier scope Every restaurant serves burgers. Every restaurant serves every burger.

17. Semantic representations for other constituents? Adjective phrases: – Happy people, cheap food, purple socks – intersective semantics Nom  Adj Nom {x Nom.sem(x) ^ Isa(x,Adj.sem)} Adj  cheap {Cheap} x Isa(x, Food) ^ Isa(x,Cheap) …works ok … But….fake gun? Local restaurant? Former friend? Would-be singer? Ex Isa(x, Gun) ^ Isa(x,Fake)

18. Doing Compositional Semantics Incorporating compositional semantics into CFG requires: Right representation for each constituent based on the parts of that constituent (e.g. Adj) Right representation for a category of constituents based on other grammar rules, making use of that constituent (e.g. V.sem) This gives us a set of function-like semantic attachments incorporated into our CFG E.g. Nom  Adj Nom {x Nom.sem(x) ^ Isa(x,Adj.sem)} A number of formalisms that extend CFGs to allow larger compositionality domains.

19. Computing the Semantic Representation Two approaches: Compute the semantic representation of each constituent as the parser progresses through the rules. – Semantic representations could be used to rule out parses – Wasted time in constructing semantics for unused constituents. Let the parser complete the syntactic parse and then recover the semantic representation. – in a bottom-up traversal. Issues of ambiguous syntactic representation Packing ambiguity Underspecified semantics.

20. Non-Compositional Language Non-compositional modifiers: fake, former, local Metaphor: You’re the cream in my coffee. She’s the cream in George’s coffee. The break-in was just the tip of the iceberg. This was only the tip of Shirley’s iceberg. Idioms: The old man finally kicked the bucket. The old man finally kicked the proverbial bucket. Deferred reference: The ham sandwich wants his check. Solutions? Mix lexical items with special grammar rules? Or???

21. Lexical Semantics

22. Thinking about Words Again Lexeme: an entry in the lexicon that includes an orthographic representation a phonological form a symbolic meaning representation or sense Some typical dictionary entries: Red (‘red) n: the color of blood or a ruby Blood (‘bluhd) n: the red liquid that circulates in the heart, arteries and veins of animals

23. Right (‘rIt) adj: located nearer the right hand esp. being on the right when facing the same direction as the observer Left (‘left) adj: located nearer to this side of the body than the right Can we get semantics directly from online dictionary entries? Some are circular All are defined in terms of other lexemes You have to know something to learn something What can we learn from dictionaries? Relations between words: – Oppositions, similarities, hierarchies

24. Homonomy Homonyms: Words with same form – orthography and pronunciation -- but different, unrelated meanings, or senses (multiple lexemes) A bank holds investments in a custodial account in the client’s name. As agriculture is burgeoning on the east bank, the river will shrink even more Word sense disambiguation: what clues? Similar phenomena homophones - read and red – same pronunciation/different orthography homographs - bass and bass – same orthography/different pronunciation

25. Ambiguity: Which applications will these cause problems for? A bass, the bank, /red/ General semantic interpretation Machine translation Spelling correction Speech recognition Text to speech Information retrieval

26. Polysemy Word with multiple but related meanings (same lexeme) They rarely serve red meat. He served as U.S. ambassador. He might have served his time in prison. What’s the difference between polysemy and homonymy? Homonymy: Distinct, unrelated meanings Different etymology? Coincidental similarity?

27. Polysemy: Distinct but related meanings idea bank, sperm bank, blood bank, bank bank How different? – Different subcategorization frames? – Domain specificity? – Can the two candidate senses be conjoined? ?He served his time and as ambassador to Norway. For either, practical task: What are its senses? (related or not) How are they related? (polysemy ‘easier’ here) How can we distinguish them?

28. Tropes, or Figures of Speech MetaphorMetaphor: one entity is given the attributes of another (tenor/vehicle/ground) Life is a bowl of cherries. Don’t take it serious…. We are the eyelids of defeated caves. ?? Metonymy: one entity used to stand for another (replacive) GM killed the Fiero. The ham sandwich wants his check. Both extend existing sense to new meaning Metaphor: completely different concept Metonymy: related concepts

29. Synonymy Substitutability: different lexemes, same meaning How big is that plane? How large is that plane? How big are you? Big brother is watching. What influences substitutability? Polysemy (large vs. old sense) register: He’s really cheap/?parsimonious. collocational constraints: roast beef, ?baked beef economy fare ?economy price

30. Finding Synonyms and Collocations Automatically from a Corpus Synonyms: Identify words appearing frequently in similar contexts Blast victims were helped by civic-minded passersby. Few passersby came to the aid of this crime victim. Collocations: Identify synonyms that don’t appear in some specific similar contexts Flu victims, flu suffers,… Crime victims, ?crime sufferers, …

31. Hyponomy General: hypernym (super…ordinate) dog is a hypernym of poodle Specific: hyponym (under..neath) poodle is a hyponym of dog Test: That is a poodle implies that is a dog Ontology: set of domain objects Taxonomy? Specification of relations between those objects Object hierarchy? Structured hierarchy that supports feature inheritance (e.g. poodle inherits some properties of dog)

32. Semantic Networks Used to represent lexical relationships e.g. WordNet (George Miller et al) Most widely used hierarchically organized lexical database for English Synset: set of synonyms, a dictionary-style definition (or gloss), and some examples of uses --> a concept Databases for nouns, verbs, and modifiers Applications can traverse network to find synonyms, antonyms, hierarchies,... Available for download or online use http://www.cogsci.princeton.edu/~wn

33. Using WN, e.g. in Question-Answering Pasca & Harabagiu ’01 results on TREC corpus Parses questions to determine question type, key words (Who invented the light bulb?) Person question; invent, light, bulb The modern world is an electrified world. It might be argued that any of a number of electrical appliances deserves a place on a list of the millennium's most significant inventions. The light bulb, in particular, profoundly changed human existence by illuminating the night and making it hospitable to a wide range of human activity. The electric light, one of the everyday conveniences that most affects our lives, was invented in 1879 simultaneously by Thomas Alva Edison in the United States and Sir Joseph Wilson Swan in England. Finding named entities is not enough

34. Compare expected answer ‘type’ to potential answers For questions of type person, expect answer is person Identify potential person names in passages retrieved by IR Check in WN to find which of these are hyponyms of person Or, Consider reformulations of question: Who invented the light bulb For key words in query, look for WN synonyms E.g. Who fabricated the light bulb? Use this query for initial IR Results: improve system accuracy by 147% (on some question types)

35. Thematic Roles ∃ w,x,y,z {Giving(x) ^ Giver(w,x) ^ Givee(z, x) ^ Given(y,x)} A set of roles for each event: Agent: volitional causer -- John hit Bill. Experiencer: experiencer of event – Bill got a headache. Force: non-volitional causer – The concrete block struck Bill on the head. Theme/patient: most affected participant – John hit Bill. Result: end product – Bill got a headache. Content: proposition of propositional event – Bill thought he should take up martial arts.

36. Instrument: instrument used -- John hit Bill with a bat Beneficiary: qui bono – John hit Bill to avenge his friend Source: origin of object of transfer event – Bill fled from New York to Timbuktu Goal: destination of object -- Bill led from New York to Timbuktu But there are a lot of verbs, with a lot of frames… FramenetFramenet encoded frames for many verb categories

37. Thematic Roles and Selectional Restrictions Selectional restrictions: semantic constraint that a word (lexeme) imposes on the concepts that go with it George hit Bill with ….John/a gun/gusto. Jim killed his philodendron/a fly/Bill. ?His philodendron killed Jim. The flu/Misery killed Jim.

38. Thematic Roles/Selectional Restrictions In practical use: Given e.g. a verb and a corpus (plus FrameNet) What conceptual roles are likely to accompany it? What lexemes are likely to fill those roles? Assassinate Give Imagine Fall Serve

39. Schank's Conceptual Dependency Eleven predicate primitives represent all predicates Objects decomposed into primitive categories and modifiers But few predicates result in very complex representations of simple things ∃ x,y Atrans(x) ^ Actor(x,John) ^ Object(x,Book) ^ To(x,Mary) ^ Ptrans(y) ^ Actor(y,John) ^ Object(y,Book) ^ To(y,Mary) John caused Mary to die vs. John killed Mary

40. Robust Semantics, Information Extraction, and Information Retrieval

41. Problems with Syntax-Driven Semantics Compositionality: Expects correspondence between syntactic and semantic structures. – Mismatch between syntactic structures and semantic structures: certainly not rule-to-rule. (inadequacy of CFGs) I like soup. Soup is what I like. Constituent trees contain many structural elements not clearly important to making semantic distinctions – Resort to dependency trees. Too abstract: Syntax driven semantic representations are sometimes very abstract. – Nominal  Adjective Nominal λ x Nominal.sem(x) AM(x,Adj.sem) – Cheap restaurant, Italian restaurant, local restaurant Robust Semantic processing: Trade-off Portability Expressivity

42. Semantic Grammars Before: CFG with syntactic categories with semantic representation composition overlaid. Now: CFG with domain-specific semantic categories Domain specific: Rules correspond directly to entities and activities in the domain I want to go from Boston to Baltimore on Thursday, September 24 th Greeting  {Hello|Hi|Um…} TripRequest  Need-spec travel-verb from City to City on Date Note: Semantic grammars are still CFGs.

43. Pros and Cons of Semantic Grammars Semantic grammars encode task knowledge and constrains the range of possible user input. I want to go to Boston on Thursday. I want to leave from there on Friday for Baltimore. TripRequest  Need-spec travel-verb from City on Date for City The semantic representation is a slot-filler frame-like representation – crafted for that domain. Portability: Lack of generality – A new one for each application – Large cost in development time Robustness: If users go outside the grammar, things may break disastrously I want to go from ah to Boston from Newark Expressivity: – I want to go to Boston from Newark or New York

44. Information Extraction Another ‘robust’ alternative Idea: ‘extract’ particular types of information from arbitrary text or transcribed speech Examples: Named entities: people, places, organizations, times, dates – MIPS Vice President John Hime MUC evaluations Domains: Medical texts, broadcast news (terrorist reports), company mergers, customer care voicemail,...

45. Appropriate where Semantic Grammars and Syntactic Parsers are Not Appropriate where information needs very specific and specifiable in advance Question answering systems, gisting of news or mail… Job ads, financial information, terrorist attacks Input too complex and far-ranging to build semantic grammars But full-blown syntactic parsers are impractical Too much ambiguity for arbitrary text 50 parses or none at all Too slow for real-time applications

46. Information Extraction Techniques Often use a set of simple templates or frames with slots to be filled in from input text Ignore everything else My number is 212-555-1212. The inventor of the wiggleswort was Capt. John T. Hart. The king died in March of 1932. Generative Model: POS-style HMM model (with novel encoding) The/O king/O died/O in/O March/I of/I 1932/I in/O France/O T* = argmax T P(W|T) * P(T) Context neighboring words, capitalization, punctuation can be used as well.

47. Discriminative Disambiguation Techniques Large set of features makes MLE estimation of the parameters unreliable. P(T|W) = π P(t i | W, POS, Ortho) = P(t i | w i-k …w i+k, pos i-k …pos i+k, ortho i ) Direct approach: – F (t i,w i-k …w i+k, pos i-k …pos i+k, ortho i ) = F(y,X) – F(y,X) = Maximum Entropy Markov Models, Conditional Random Fields

48. ScanMail Transcription gender F age A caller_name NA native_speaker N speech_pathology N sample_rate 8000 label 0 804672 " [ Greeting: hi R__ ] [ CallerID: it's me ] give me a call [ um ] right away cos there's [.hn ] I guess there's some [.hn ] change [ Date: tomorrow ] with the nursery school and they [ um ] [.hn ] anyway they had this idea [ cos ] since I think J__'s the only one staying [ Date: tomorrow ] for play club so they wanted to they suggested that [.hn ] well J2__actually offered to take J__home with her and then would she would meet you back at the synagogue at [ Time: five thirty ] to pick her up [.hn ] [ uh ] so I don't know how you feel about that otherwise Miriam and one other teacher would stay and take care of her till [ Date: five thirty tomorrow ] but if you [.hn ] I wanted to know how you feel before I tell her one way or the other so call me [.hn ] right away cos I have to get back to her in about an hour so [.hn ] okay [ Closing: bye [.nhn ] [.onhk ] ]" duration "50.3 seconds"

49. SCANMail Access Devices PC Pocket PC Dataphone Voice Phone Flash E-mail

50. Word Sense Disambiguation

51. Disambiguation via Selectional Restrictions A step toward semantic parsing Different verbs select for different thematic roles wash the dishes (takes washable-thing as patient) serve delicious dishes (takes food-type as patient) Method: rule-to-rule syntactico-semantic analysis Semantic attachment rules are applied as sentences are syntactically parsed VP --> V NP V  serve {theme:food-type} Selectional restriction violation: no parse

52. Requires: Write selectional restrictions for each sense of each predicate – or use FrameNetFrameNet – serve alone has 15 verb senses Hierarchical type information about each argument (a la WordNet) WordNet – How many hypernyms does dish have? – How many lexemes are hyponyms of dish?hyponyms But also: Sometimes selectional restrictions don’t restrict enough (Which dishes do you like?) Sometimes they restrict too much (Eat dirt, worm! I’ll eat my hat!)

53. Can we take a more statistical approach? How likely is dish/crockery to be the object of serve? dish/food? A simple approach (baseline): predict the most likely sense Why might this work? When will it fail? A better approach: learn from a tagged corpus What needs to be tagged? An even better approach: Resnik’s selectional association (1997, 1998) Estimate conditional probabilities of word senses from a corpus tagged only with verbs and their arguments (e.g. ragout is an object of served -- Jane served/V ragout/Obj

54. How do we get the word sense probabilities? For each verb object (e.g. ragout) – Look up hypernym classes in WordNet – Distribute “credit” for this object sense occurring with this verb among all the classes to which the object belongs Brian served/V the dish/Obj Jane served/V food/Obj – If ragout has N hypernym classes in WordNet, add 1/N to each class count (including food) as object of serve – If tureen has M hypernym classes in WordNet, add 1/M to each class count (including dish) as object of serve Pr(Class|v) is the count(c,v)/count(v) How can this work? – Ambiguous words have many superordinate classes John served food/the dish/tuna/curry – There is a common sense among these which gets “credit” in each instance, eventually dominating the likelihood score

55. To determine most likely sense of ‘bass’ in Bill served bass Having previously assigned ‘credit’ for the occurrence of all hypernyms of things like fish and things like musical instruments to all their hypernym classes (e.g. ‘fish’ and ‘musical instruments’) Find the hypernym classes of bass (including fish and musical instruments) Choose the class C with the highest probability, given that the verb is serve Results: Baselines: – random choice of word sense is 26.8% – choose most frequent sense (NB: requires sense-labeled training corpus) is 58.2% Resnik’s: 44% correct with only pred/arg relations labeled

56. Machine Learning Approaches Learn a classifier to assign one of possible word senses for each word Acquire knowledge from labeled or unlabeled corpus Human intervention only in labeling corpus and selecting set of features to use in training Input: feature vectors Target (dependent variable) Context (set of independent variables) Output: classification rules for unseen text

57. Supervised Learning Training and test sets with words labeled as to correct sense (It was the biggest [fish: bass] I’ve seen.) Obtain values of independent variables automatically (POS, co-occurrence information, …) Run classifier on training data Test on test data Result: Classifier for use on unlabeled data

58. Input Features for WSD POS tags of target and neighbors Surrounding context words (stemmed or not) Punctuation, capitalization and formatting Partial parsing to identify thematic/grammatical roles and relations Collocational information: How likely are target and left/right neighbor to co-occur Co-occurrence of neighboring words Intuition: How often does sea or words with bass

59. How do we proceed? – Look at a window around the word to be disambiguated, in training data – Which features accurately predict the correct tag? – Can you think of other features might be useful in general for WSD? Input to learner, e.g. Is the bass fresh today? [w-2, w-2/pos, w-1,w-/pos,w+1,w+1/pos,w+2,w+2/pos… [is,V,the,DET,fresh,RB,today,N...

60. Types of Classifiers Naïve Bayes ŝ = p(s|V), or Where s is one of the senses possible and V the input vector of features Assume features independent, so probability of V is the product of probabilities of each feature, given s, so and p(V) same for any ŝ Then

61. Rule Induction Learners (e.g. Ripper) Given a feature vector of values for independent variables associated with observations of values for the training set (e.g. [fishing,NP,3,…] + bass 2 ) Produce a set of rules that perform best on the training data, e.g. bass 2 if w-1== ‘ fishing ’ & pos==NP …

62. Like case statements applying tests to input in turn fish within window--> bass 1 striped bass--> bass 1 guitar within window--> bass 2 bass player--> bass 1 … Ordering based on individual accuracy on entire training set based on log-likelihood ratio Decision Lists

63. Bootstrapping I Start with a few labeled instances of target item as seeds to train initial classifier, C Use high confidence classifications of C on unlabeled data as training data Iterate Bootstrapping II Start with sentences containing words strongly associated with each sense (e.g. sea and music for bass), either intuitively or from corpus or from dictionary entries One Sense per Discourse hypothesis

64. Unsupervised Learning Cluster feature vectors to ‘discover’ word senses using some similarity metric (e.g. cosine distance) Represent each cluster as average of feature vectors it contains Label clusters by hand with known senses Classify unseen instances by proximity to these known and labeled clusters Evaluation problem What are the ‘right’ senses? Cluster impurity How do you know how many clusters to create? Some clusters may not map to ‘known’ senses

65. Dictionary Approaches Problem of scale for all ML approaches Build a classifier for each sense ambiguity Machine readable dictionaries (Lesk ‘86) Retrieve all definitions of content words occurring in context of target (e.g. the happy seafarer ate the bass) Compare for overlap with sense definitions of target entry (bass 2 : a type of fish that lives in the sea) Choose sense with most overlap Limits: Entries are short --> expand entries to ‘related’ words