A Memory-Based Model of Syntactic Analysis: Data Oriented Parsing Remko Scha, Rens Bod, Khalil Sima ’ an Institute for Logic, Language and Computation University of Amsterdam
Outline of the lecture Introduction Disambiguation Data Oriented Parsing DOP1 computational aspects and experiments Memory Based Learning framework Conclusions
Introduction Human language cognition: Analogy-based processes on a store of past experiences Modern linguistics Set of rules Language processing algorithms Performance model of human language processing Competence grammar as broad framework to performance models. Memory / Analogy - based language processing
The Problem of Ambiguity Resolution Every input string has unmanageable large number of analyses Uncertain input – generate guesses and choose one Syntactic disambiguation might be a side effect of semantic one
The Problem of Ambiguity Resolution Frequency of occurrence of lexical item and syntactic structures: People register frequencies People prefer analyses they already experienced than constructing a new ones More frequent analyses are preferred to less frequent ones
From Probabilistic Competence- Grammars to Data-Oriented Parsing Probabilistic information derived from past experience Characterization of the possible sentence-analyses of the language Stochastic Grammar Define : all sentences, all analyses. Assign : probability for each Achieve : preference that people display when they choose sentence or analyses.
Stochastic Grammar These predictions are limited Platitudes and conventional phrases Allow redundancy Use Tree Substitution Grammar
Stochastic Tree Substitution Grammar Set of elementary trees Tree rewrite process Redundant model Statistically relevant phrases Memory based processing model
Memory based processing model Data oriented parsing approach: Corpus of utterances – past experience STSG to analyze new input In order to describe a specific DOP model A formalism for representing utterance- analyses An extraction function Combination operations A probability model
A Simple Data Oriented Parsing Model: DOP1 Our corpus: DOP1 - Imaginary corpus of two treesDOP1 - Imaginary corpus of two trees Possible sub trees: t consists of more than one node t is connected except for the leaf nodes of t, each node in t has the same daughter-nodes as the corresponding node in T Stochastic Tree Substitution Grammar – set of sub trees Generation process – composition: A B – B is substituted on the leftmost non terminal leaf node of A
Example of sub trees
DOP1 - Imaginary corpus of two trees
Derivation and parse #1 She saw the dress with the telescope.
Derivation and parse #2 She saw the dress with the telescope.
Probability Computations: Probability of substituting a sub tree t on a specific node Probability of Derivation Probability of Parse Tree
Computational Aspects of DOP1 Parsing Disambiguation Most Probable Derivation Most Probable Parse Optimizations
Parsing Chart-like parse forest Derivation forest Elementary tree t as a context-free rule: root(t) — > yield(t) Label phrase with it ’ s syntactic category and its full elementary tree
Elementary trees of an example STSG
Derivation forest for the string abcd
Derivations and parse trees for the string abcd
Disambiguation Derivation forest define all derivation and parses Most likely parse must be chosen MPP in DOP1 MPP vs. MPD
Most Probable Derivation Viterbi algorithm: Eliminate low probability sub derivations using bottom-up fashion Select the most probable sub derivation at each chart entry, eliminate other sub derivation of that root node.
Viterbi algorithm Two derivations for abc d1 > d2 : eliminate the right derivation
Algorithm 1 – Computing the probability of most probable derivation Input : STSG, S, R, P Elementary trees in R are in CNF A — >t H : tree t, root A, sequence of labels H. - non terminal A in chart entry (i,j) after parsing the input W1,...,Wn. PPMPD – probability of MPD of input string W1,...,Wn.
Algorithm 1 – Computing the probability of most probable derivation
The Most Probable Parse Computing MPP in STSG is NP hard Monte Carlo method Sample derivations Observe frequent parse tree Estimate parse tree probability Random – first search The algorithm Law of Large Numbers
Algorithm 2: Sampling a random derivation for length := 1 to n do for start := 0 to n - length do for each root node X chart-entry (start, start + length) do: 1. select at random a tree from the distribution of elementary trees with root node X 2. eliminate the other elementary trees with root node X from this chart-entry
Results of Algorithm 2 Random derivation for the whole sentence First guess for MPP Compute the size of the sampling set Probability of error Upper bound 0 index of MPP,i index of parse i, N derivation No unique MPP – ambiguity
Reminder
Conclusions – lower bound for N Lower bound for N: Pi is probability of parse i B - Estimated probability by frequencies in N Var(B) = Pi*(1-Pi)/N 0 Var(B) <= 1/(4*N) s = sqrt(Var(B)) -> S <= 1/(2*sqrt(N)) 1/(4*s^2) <= N 100 s <= 0.05
Algorithm 3: Estimating the parse probabilities Given a derivation forest of a sentence and a threshold sm for the standard error: N := the smallest integer larger than 1/(4 sm 2) repeat N times: sample a random derivation from the derivation forest store the parse generated by this derivation for each parse i: estimate the conditional probability given the sentence by pi := #(i) / N
Complexity of Algorithm 3 Assumes value of max allowed standard error Samples number of derivations which is guaranteed to achieve the error Number of needed samples is quadratic in chosen error
Optimizations Sima ’ an : MPD in linear time in STSG size Bod : MPP on small random corpus of sub trees Sekine and Grishman : use only sub trees rooted with S or NP Goodman : different polynomial time
Experimental Properties of DOP1 Experiments on the ATIS corpus MPP vs. MPD Impact of fragment size Impact of fragment lexicalization Impact of fragment frequency Experiments on SRI-ATIS and OVIS Impact of sub tree depth
Experiments on ATIS corpus ATIS = Air Travel Information System 750 annotated sentence analyses Annotated by Penn Treebank Purpose: compare accuracy obtained in undiluted DOP1 with the one obtained in restricted STSG
Experiments on ATIS corpus Divide into training and test sets 90% = 675 in training set 10% = 75 in test set Convert training set into fragments and enrich with probabilities Test set sentences parsed with sub trees from the training set MPP was estimated from 100 sampled derivations Parse accuracy = % of MPP that are identical to test set parses
Results On 10 random training / test splits of ATIS: Average parse accuracy = 84.2% Standard deviation = 2.9 %
Impact of overlapping fragments MPP vs. MPD Can MPD achieve parse accuracies similar to MPP Can MPD do better than MPP Overlapping fragments Accuracies generated by MPD on test set The result is 69% Comparing to accuracy achieved with MPP on test set : 69% vs. 85% Conclusion: overlapping fragments play important role in predicting the appropriate analysis of a sentence
The impact of fragment size Large fragments capture more lexical/syntactic dependencies than small ones. The experiment: Use DOP1 with restricted maximum depth Max depth 1 -> DOP1 = SCFG Compute the accuracies both for MPD and MPP for each max depth
Impact of fragment size
Impact of fragment lexicalization Lexicalized fragment More words -> more lexical dependencies Experiment: Different version of DOP1 Restrict max number of words per fragment Check accuracy for MPP and MPD
Impact of fragment lexicalization
Impact of fragment frequency Frequent fragments contribute more large fragments are less frequent than small ones but might contribute more Experiment: Restrict frequency to min number of occurrences Not other restrictions Check accuracy for MPP
Impact of fragment frequency
Experiments on SRI-ATIS and OVIS Employ MPD because the corpus is bigger Tests performed on DOP1 and SDOP Use set of heuristic criteria for selecting the fragments: Constraints of the form of sub trees d - upper bound on depth n – number of substitution sites l – number of terminals L – number of consecutive terminals Apply constraints on all sub trees besides those with depth 1
Experiments on SRI-ATIS and OVIS d4 n2 l7 L3 DOP(i) Evaluation metrics: Recognized Tree Language Coverage – TLC Exact match Labeled bracketing recall and precision
Experiments on SRI-ATIS annotated syntactically utterances Annotation scheme originated from Core Language Engine system Fixed parameters except sub tree bound: n2 l4 L3 Training set – trees Test set – 1000 trees Experiment: Train and test on different depths upper bounds (takes more than 10 days for DOP(4) !!! )
Impact of sub tree depth SRI-ATIS
Experiments on OVIS corpus syntactically and semantically annotated trees Both annotations treated as one More non terminal symbols Utterances are answers to questions in dialog -> short utterances (avg. 3.43) Sima ’ an results – sentences with at least 2 words, avg n2 l7 L3
Experiments on OVIS corpus Experiment: Check different sub tree depth 1,3,4,5 Test set with 1000 trees Train set with 9000 trees
Impact of sub tree depth - OVIS
Summary of results ATIS: Accuracy of parsing is 85% Overlapping fragments have impact on accuracy Accuracy increases as fragment depth increases both for MPP and MPD Optimal lexical maximum for ATIS is 8 Accuracy decreases if lower bound of fragment frequency increases (for MPP)
Summary of results SRI-ATIS: Availability of more data is more crucial to accuracy of MPD. Depth has impact Accuracy is improved when using memory based parsing(DOP(2)) and not SCFG (DOP(1))
Summary of results OVIS: Recognition power isn ’ t affected by depth No big difference between exact match in DOP1(1) and DOP1(4) mean and standard deviations
DOP: probabilistic recursive MBL Relationship between present DOP framework and Memory Based Learning framework DOP extends MBL to deal with disambiguation MBL vs. DOP Flat or intermediate description vs. hierarchical
Case Based Reasoning - CBR Case Based learning Lazy learning, doesn ’ t generalize Lazy generalization Classify by means of similarity function Refer this paradigm as MBL CBR vs. other variants of MBL Task concept Similarity function Learning task
The DOP framework and CBR CBR method A formalism for representing utterance- analyses - case description language An extraction function – retrieve units Combination operations – reuse and revision Missing in DOP: Similarity function Extend CBR: A probability model DOP model defines CBR system for natural language analysis
DOP1 and CBR methods DOP1 as extension to CBR system = classified instance Retrieve sub trees and construct tree Sentence = instance Tree = class Set of sentences = instance space Set of trees – class space Frontier, SSF, Infinite runtime case-base containing instance-class-weight triples:
DOP1 and CBR methods Task and similarity function: Task = disambiguation Similarity function: Parsing -> recursive string matching procedure Ambiguity -> computing probability and selecting the highest. Conclusion: DOP1 is a lazy probabilistic recursive CBR classifier
DOP vs. other MBL approached in NLP K-NN vs. DOP Memory Based Sequence Learning DOP – stochastic model fro computing probabilities MBSL – ad hoc heuristics for computing scores DOP – globally based ranking strategy of alternative analyzes MBSL – locally based one Different generalization power
Conclusions Memory Based aspects of DOP model Disambiguation Probabilities to account frequencies DOP as probabilistic recursive Memory Based model DOP1 - properties, computational aspects and experiments. DOP and MBL - differences