Download presentation
Presentation is loading. Please wait.
Published byAbel Hamilton Modified over 9 years ago
1
Dependency Parsing Joakim Nivre
2
Dependency Grammar Old tradition in descriptive grammar Modern theroretical developments: –Structural syntax (Tesnière) –Meaning-Text Theory (Mel’čuk) –Word Grammar (Hudson) –Functional Generative Description (Prague) Basic idea: –Syntactic structure consists of binary, asymmetrical relations between the words of a sentence
3
Dependency Representations Directed graphs: –V is a set of nodes (tokens) –E is a set of arcs (dependency relations) –L is a labeling function on E (dependency types) Example: PP På (In) NN 60-talet (the-60’s) VB målade (painted) PN han (he) JJ djärva (bold) NN tavlor (pictures) ADV PR OBJ SUBATT
4
Graph Constraints Commonly imposed constraints: –Single-head (at most one head per node) –Connectedness (no dangling nodes) –Acyclicity (no cycles in the graph) –Projectivity: An arc i j is projective iff, for every k occurring between i and j in the input string, i j. A graph is projective iff every arc in A is projective.
5
Dependency Parsing Dependency-based grammar parsing: –Given a dependency grammar G and an input string x *, derive some or all of the dependency graphs y assigned to x by G. Dependency-based text parsing: –Given a text T = (x 1, …, x n ), derive the correct dependency graph y i for every sentence x i T. Text parsing may be grammar-driven or not.
6
Parsing Methods Three main approaches: –Dynamic programming algorithms applied to context-free (projective) dependency grammars –Eliminative parsing techniques applied to constraint-based formulations of (non- projective) dependency grammar –Deterministic parsing algorithms combined with weak grammars or data-driven methods
7
Dynamic Programming Early formalizations: –Hays/Gaifman: Equivalent to a subset of context-free grammars (roughly lexicalized) –Tabular parsing techniques (cf. CKY parsing) Modern developments: –Link grammar (Sleator and Temperley) –Bilexical grammar (Eisner): Lexicalized parsing in O(n 3 ) time by combining spans instead of constituents.
8
Constraint Satisfaction Constraints on dependency graphs (Maruyama): pos(i) = D [dep(i) = DET pos(head(i)) = N i < head(i)] pos(i) = N [dep(i) = SUBJ pos(head(i)) = V i < head(i)] pos(i) = V [dep(i) = ROOT head(i) = nil] [head(i) = head(j) dep(i) = dep(j)] i = j Graph satisfying the above constraints: DaDa N dog V runs SUBJDET
9
Parsing with Constraints Eliminative parsing: –Start with all formally possible analyses (in a compact representation) –Eliminate representations that violate constraints until only valid analyses remain Variations: –Weighted constraints –Transformational parsing
10
Deterministic Parsing Covington’s fundamental algorithm: –Accept words one by one starting at the beginning of the sentence, and try linking each word as head or dependent of every previous word. Variations on shift-reduce parsing: –Standard (Kudo, Matsumoto, Yamada) –Arc-eager (Nivre)
11
Arc-Eager Parsing Configuration: C = S, I, A S = Stack I = Input (remaining) A = Arc relation (current) Initialization: nil, W, Termination: S, nil, A for any S, A Acceptance: S, nil, A if (W, A) is connected
12
Transitions Left-Arc (LA): w i |S, w j |I, A S, w j |I, A {(w j, w i )} if a : a A dep(a) = w i Right-Arc (RA): w i |S, w j |I, A w j |w i |S, I, A {(w i, w j )} if a : a A dep(a) = w j Reduce (RE): w i |S, I, A S, I, A if a : a A dep(a) = w i Shift (SH): S, w i |I, A w i |S, I, A
13
MaltParser Robust, data-driven dependency parsing using a combination of: –Deterministic parsing (e.g. arc-eager) –Discriminative machine learning (e.g. MBL) –User-defined feature models: Lexical features Part-of-speech features Dependency type features
14
Why (Not) Dependency Parsing? Potential advantages of dependency parsing: –Dependency relations are close to semantic relations, which facilitates semantic interpretation –Dependency representations are more constrained (less complex), which facilitates parsing –Dependency representations are more suitable for languages with free or flexible word order Potential disadvantages: –Dependency representations are less expressive –Dependency representations are less well understood formally and computationally
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.