Human Parsing Human Parsing Probabilistic Model Probabilistic Model Modeling Results Modeling Results Open Issues Open Issues 1 Human Parsing Garden Paths Probabilistic Models of Human Parsing Parser Architectures Informatics 2A: Lecture 23 2 Probabilistic Model Probabilistic Grammars Frame Probabilities Mirella Lapata (slides by Frank Keller) 3 Modeling Results School of Informatics Frame Preferences University of Edinburgh mlap@inf.ed.ac.uk Garden Paths Beam Width November 10, 2011 4 Open Issues 1 / 32 2 / 32 Human Parsing Human Parsing Probabilistic Model Garden Paths Probabilistic Model Garden Paths Modeling Results Parser Architectures Modeling Results Parser Architectures Open Issues Open Issues Overview Garden Paths Main Clause vs. Reduced Relative Ambiguity In this lecture, we will discuss a classic probabilistic model of (1) a. ?The horse raced past the barn fell. human parsing (Jurafsky, 1996): b. ?The teachers taught by the Berlitz method passed the the model integrates lexical and syntactic access and test. disambiguation; c. The children taught by the Berlitz method passed the it accounts for psycholinguistic data using concepts from NLP: test. probabilistic CFGs, Bayesian modeling, frame probabilities; Frame Ambiguity here, we focus on: syntactic disambiguation in human parsing. (2) a. ?The landlord painted all the walls with cracks. See previous lecture for background on human parsing (garden b. ?Ross baked the cake in the freezer. paths, parser architectures). Note: ? means garden path. 3 / 32 4 / 32
Human Parsing Human Parsing Probabilistic Model Garden Paths Probabilistic Model Garden Paths Modeling Results Parser Architectures Modeling Results Parser Architectures Open Issues Open Issues Garden Paths Frame Preferences A verb can have several subcategorization frames (phrases it selects for). Some frames are preferred over others: Lexical Category Ambiguity (4) The women discussed the dogs on the beach. (3) a. ?The complex houses married and single students and a. The women discussed the dogs which were on the beach. their families. (90%) b. ?The warehouse fires destroyed all the buildings. b. The women discussed them (the dogs) while on the beach. c. ?The warehouse fires a dozen employees each year. (10%) d. ?The prime number few. e. ?The old man the boats. (5) The women kept the dogs on the beach. f. ?The grappling hooks on to the enemy ship. a. The women kept the dogs which were on the beach. (5%) b. The women kept them (the dogs) while on the beach. (95%) Results from rating study by Ford et al. (1982). 5 / 32 6 / 32 Human Parsing Human Parsing Probabilistic Model Garden Paths Probabilistic Model Garden Paths Modeling Results Parser Architectures Modeling Results Parser Architectures Open Issues Open Issues Clicker Question (1) Parser Architectures Serial Parser Which one of the following is not a plausible architecture for a human parser? build parse trees through successive rule selection; if more than one rule applies (choice point), chose one 1 A serial parser maintains only one analysis at a time possible tree based on a selection rule; 2 A parallel parser maintains several analyses if the tree turns out to be impossible, return to the choice 3 A parser that computes analyses sentence-by-sentence point (backtracking) and reparse from there; 4 A parser that combines serial processing with limited example for selection rule: minimal attachment (choose the parallelism tree with the least nodes). 7 / 32 8 / 32
Human Parsing Human Parsing Probabilistic Model Garden Paths Probabilistic Model Garden Paths Modeling Results Parser Architectures Modeling Results Parser Architectures Open Issues Open Issues Parser Architectures Modeling Human Parsing Serial Parser Parallel Parser garden path means: wrong tree selected at a choice point; build parse trees through successive rule selection; backtracking occurs, causes increased processing times. if more than one rule applies, create a new tree for each rule; Parallel Parser pursue all possibilities in parallel; garden path means: correct tree was pruned; if one turns out to be impossible, drop it; backtracking occurs, causes increased processing times. problem: number of parse trees can grow exponentially. Jurafsky (1996) assumes bounded parallelism in a parsing model solution: bounded parallelism, only pursue a limited number of based on probabilistic CFGs. possibilities (prune trees). Pruning occurs if a parse tree is sufficiently improbable (beam search algorithm). 9 / 32 10 / 32 Human Parsing Human Parsing Probabilistic Model Probabilistic Grammars Probabilistic Model Probabilistic Grammars Modeling Results Frame Probabilities Modeling Results Frame Probabilities Open Issues Open Issues Probabilistic Context-free Grammars Probabilistic Context-free Grammars Context-free rules annotated with probabilities; Example probabilities of all rules with the same lefthand side sum to S 1 . 0 one; ✟ ❍ ✟✟✟ ❍ ❍ probability of a parse is the product of the probabilities of all ❍ NP 0 . 1 VP 0 . 7 rules applied in the parse. ✟ ❍ ✟✟✟ ❍ ❍ astronomers ❍ V 1 . 0 NP 0 . 4 Example ✟ ❍ ✟✟ ❍ ❍ saw S → NP VP 1.0 NP → NP PP 0.4 NP 0 . 18 PP 1 . 0 PP → P NP 1.0 NP → astronomers 0.1 ✟ ❍ ❍ ✟ P 1 . 0 NP 0 . 18 stars VP → V NP 0.7 NP → ears 0.18 VP → VP PP 0.3 NP → saw 0.04 with ears P → with 1.0 NP → stars 0.18 V → saw 1.0 NP → telescopes 0.1 P ( t 1 ) = 1 . 0 · 0 . 1 · 0 . 7 · 1 . 0 · 0 . 4 · 0 . 18 · 1 . 0 · 1 . 0 · 0 . 18 = 0 . 0009072 11 / 32 12 / 32
Human Parsing Human Parsing Probabilistic Model Probabilistic Grammars Probabilistic Model Probabilistic Grammars Modeling Results Frame Probabilities Modeling Results Frame Probabilities Open Issues Open Issues Probabilistic Context-free Grammars Frame Probabilities Example Subcategorization frames of the verb keep: S 1 . 0 ✟✟✟✟✟ ❍ ❍ NP AP keep the prices reasonable ❍ ❍ ❍ NP VP keep his foes guessing NP 0 . 1 VP 0 . 3 NP VP keep their eyes peeled ✟ ❍ ✟✟✟ ❍ ❍ astronomers NP PRT keep the people in ❍ VP 0 . 7 PP 1 . 0 NP PP keep his nerves from jangling ✟ ❍ ❍ ✟ ❍ ❍ ✟ ✟ V 1 . 0 NP 0 . 18 P 1 . 0 NP 0 . 18 Frame probabilities tell us how likely each of these frames is. This saw stars with ears information can be combined with construction probabilities generated by a probabilistic CFG. P ( t 2 ) = 1 . 0 · 0 . 1 · 0 . 3 · 0 . 7 · 1 . 0 · 0 . 18 · 1 . 0 · 1 . 0 · 0 . 18 = 0 . 0006804 t 1 more probable than t 2 : improbable analyses can be pruned. 13 / 32 14 / 32 Human Parsing Human Parsing Frame Preferences Probabilistic Model Probabilistic Grammars Probabilistic Model Garden Paths Modeling Results Frame Probabilities Modeling Results Beam Width Open Issues Open Issues Frame Probabilities Modeling Frame Preferences p (keep , � NP XP[pred +] � ) = 0 . 81 Problem: how can frame probabilities be computed? VP → V NP XP 0.15 Solution: use a corpus that’s annotated with tree structures (Penn Treebank); estimate frame probabilities from the corpus. t 1 : VP ✟ ❍ ✟✟✟✟✟✟ ❍ Example ❍ ❍ ❍ discuss � NP PP � .24 ❍ ❍ V NP PP � NP � .76 keep � NP XP[pred +] � .81 keep the dogs on the beach � NP � .19 p ( t 1 ) = 0 . 15 · 0 . 81 = 0 . 12 (preferred) 15 / 32 16 / 32
Human Parsing Human Parsing Frame Preferences Frame Preferences Probabilistic Model Probabilistic Model Garden Paths Garden Paths Modeling Results Modeling Results Beam Width Beam Width Open Issues Open Issues Modeling Frame Preferences Modeling Frame Preferences p (keep , � NP � ) = 0 . 19 VP → V NP 0.39 p (discuss , � NP PP � ) = 0 . 24 NP → NP XP 0.14 VP → V NP XP 0.15 t 2 : t 1 : VP VP ✟ ❍ ✟✟✟ ❍ ✟ ❍ ❍ ✟✟✟✟✟✟ ❍ ❍ ❍ ❍ V NP ❍ ❍ ✟ ❍ ❍ ✟✟✟ ❍ ❍ V NP PP ❍ keep NP PP discuss the dogs on the beach the dogs on the beach p ( t 1 ) = 0 . 15 · 0 . 24 = 0 . 036 (dispreferred) p ( t 2 ) = 0 . 19 · 0 . 39 · 0 . 14 = 0 . 01 (dispreferred) 17 / 32 18 / 32 Human Parsing Human Parsing Frame Preferences Frame Preferences Probabilistic Model Probabilistic Model Garden Paths Garden Paths Modeling Results Modeling Results Beam Width Beam Width Open Issues Open Issues Modeling Frame Preferences Modeling Garden Path Effects Garden path caused by construction probabilities: p (discuss , � NP � ) = 0 . 76 VP → V NP 0.39 S → NP . . . 0.92 N → house 0.0024 NP → NP XP 0.14 NP → Det Adj N 0.28 Adj → complex 0.00086 N → ROOT s 0.23 t 2 : VP t 1 : ✟✟✟✟ ❍ S ❍ ❍ ✟ ❍ ❍ ✟✟✟ ❍ V NP ❍ ❍ NP ✟✟✟ ✟ ❍ ❍ ❍ . . . ❍ discuss ✟ ❍ ✟✟✟✟ ❍ NP PP ❍ ❍ ❍ Det Adj N the dogs on the beach the complex houses p ( t 2 ) = 0 . 76 · 0 . 39 · 0 . 14 = 0 . 041 (preferred) p ( t 1 ) = 1 . 2 · 10 − 7 (preferred) 19 / 32 20 / 32
Recommend
More recommend
