| Class: | Parsing grammars that are context-free |
| Data: | String |
| Time: | O(n3) |
| Best-Time: |
|
| Average-Time: | \Theta(n3) |
In computer science, the Earley parser is an algorithm for parsing strings that belong to a given context-free language, though (depending on the variant) it may suffer problems with certain nullable grammars.[1] The algorithm, named after its inventor, Jay Earley, is a chart parser that uses dynamic programming; it is mainly used for parsing in computational linguistics. It was first introduced in his dissertation[2] in 1968 (and later appeared in an abbreviated, more legible, form in a journal).
Earley parsers are appealing because they can parse all context-free languages, unlike LR parsers and LL parsers, which are more typically used in compilers but which can only handle restricted classes of languages. The Earley parser executes in cubic time in the general case
{O}(n3)
{O}(n2)
The following algorithm describes the Earley recogniser. The recogniser can be modified to create a parse tree as it recognises, and in that way can be turned into a parser.
In the following descriptions, α, β, and γ represent any string of terminals/nonterminals (including the empty string), X and Y represent single nonterminals, and a represents a terminal symbol.
Earley's algorithm is a top-down dynamic programming algorithm. In the following, we use Earley's dot notation: given a production X → αβ, the notation X → α • β represents a condition in which α has already been parsed and β is expected.
Input position 0 is the position prior to input. Input position n is the position after accepting the nth token. (Informally, input positions can be thought of as locations at token boundaries.) For every input position, the parser generates a state set. Each state is a tuple (X → α • β, i), consisting of
(Earley's original algorithm included a look-ahead in the state; later research showed this to have little practical effect on the parsing efficiency, and it has subsequently been dropped from most implementations.)
A state is finished when its current position is the last position of the right side of the production, that is, when there is no symbol to the right of the dot • in the visual representation of the state.
The state set at input position k is called S(k). The parser is seeded with S(0) consisting of only the top-level rule. The parser then repeatedly executes three operations: prediction, scanning, and completion.
Duplicate states are not added to the state set, only new ones. These three operations are repeated until no new states can be added to the set. The set is generally implemented as a queue of states to process, with the operation to be performed depending on what kind of state it is.
The algorithm accepts if (X → γ •, 0) ends up in S(n), where (X → γ) is the top level-rule and n the input length, otherwise it rejects.
Adapted from Speech and Language Processing[4] by Daniel Jurafsky and James H. Martin,
function INIT(words) S ← CREATE_ARRAY(LENGTH(words) + 1) for k ← from 0 to LENGTH(words) do S[k] ← EMPTY_ORDERED_SET
function EARLEY_PARSE(words, grammar) INIT(words) ADD_TO_SET((γ → •S, 0), S[0]) for k ← from 0 to LENGTH(words) do for each state in S[k] do // S[k] can expand during this loop if not FINISHED(state) then if NEXT_ELEMENT_OF(state) is a nonterminal then PREDICTOR(state, k, grammar) // non_terminal else do SCANNER(state, k, words) // terminal else do COMPLETER(state, k) end end return chart
procedure PREDICTOR((A → α•Bβ, j), k, grammar) for each (B → γ) in GRAMMAR_RULES_FOR(B, grammar) do ADD_TO_SET((B → •γ, k), S[k]) end
procedure SCANNER((A → α•aβ, j), k, words) if j < LENGTH(words) and a ⊂ PARTS_OF_SPEECH(words[k]) then ADD_TO_SET((A → αa•β, j), S[k+1]) end
procedure COMPLETER((B → γ•, x), k) for each (A → α•Bβ, j) in S[x] do ADD_TO_SET((A → αB•β, j), S[k]) end
Consider the following simple grammar for arithmetic expressions: ::= # the start rule ::= "+"
This is the sequence of state sets:
| (state no.) | Production | (Origin) | Comment |
|---|---|---|---|
| S(0): • 2 + 3 * 4 | |||
| 1 | P → • S | 0 | start rule |
| 2 | S → • S + M | 0 | predict from (1) |
| 3 | S → • M | 0 | predict from (1) |
| 4 | M → • M * T | 0 | predict from (3) |
| 5 | M → • T | 0 | predict from (3) |
| 6 | T → • number | 0 | predict from (5) |
| S(1): 2 • + 3 * 4 | |||
| 1 | T → number • | 0 | scan from S(0)(6) |
| 2 | M → T • | 0 | complete from (1) and S(0)(5) |
| 3 | M → M • * T | 0 | complete from (2) and S(0)(4) |
| 4 | S → M • | 0 | complete from (2) and S(0)(3) |
| 5 | S → S • + M | 0 | complete from (4) and S(0)(2) |
| 6 | P → S • | 0 | complete from (4) and S(0)(1) |
| S(2): 2 + • 3 * 4 | |||
| 1 | S → S + • M | 0 | scan from S(1)(5) |
| 2 | M → • M * T | 2 | predict from (1) |
| 3 | M → • T | 2 | predict from (1) |
| 4 | T → • number | 2 | predict from (3) |
| S(3): 2 + 3 • * 4 | |||
| 1 | T → number • | 2 | scan from S(2)(4) |
| 2 | M → T • | 2 | complete from (1) and S(2)(3) |
| 3 | M → M • * T | 2 | complete from (2) and S(2)(2) |
| 4 | S → S + M • | 0 | complete from (2) and S(2)(1) |
| 5 | S → S • + M | 0 | complete from (4) and S(0)(2) |
| 6 | P → S • | 0 | complete from (4) and S(0)(1) |
| S(4): 2 + 3 * • 4 | |||
| 1 | M → M * • T | 2 | scan from S(3)(3) |
| 2 | T → • number | 4 | predict from (1) |
| S(5): 2 + 3 * 4 • | |||
| 1 | T → number • | 4 | scan from S(4)(2) |
| 2 | M → M * T • | 2 | complete from (1) and S(4)(1) |
| 3 | M → M • * T | 2 | complete from (2) and S(2)(2) |
| 4 | S → S + M • | 0 | complete from (2) and S(2)(1) |
| 5 | S → S • + M | 0 | complete from (4) and S(0)(2) |
| 6 | P → S • | 0 | complete from (4) and S(0)(1) |
Earley's dissertation[5] briefly describes an algorithm for constructing parse trees by adding a set of pointers from each non-terminal in an Earley item back to the items that caused it to be recognized. But Tomita noticed[6] that this does not take into account the relations between symbols, so if we consider the grammar S → SS | b and the string bbb, it only notes that each S can match one or two b's, and thus produces spurious derivations for bb and bbbb as well as the two correct derivations for bbb.
Another method[7] is to build the parse forest as you go, augmenting each Earley item with a pointer to a shared packed parse forest (SPPF) node labelled with a triple (s, i, j) where s is a symbol or an LR(0) item (production rule with dot), and i and j give the section of the input string derived by this node. A node's contents are either a pair of child pointers giving a single derivation, or a list of "packed" nodes each containing a pair of pointers and representing one derivation. SPPF nodes are unique (there is only one with a given label), but may contain more than one derivation for ambiguous parses. So even if an operation does not add an Earley item (because it already exists), it may still add a derivation to the item's parse forest.
SPPF nodes are never labeled with a completed LR(0) item: instead they are labelled with the symbol that is produced so that all derivations are combined under one node regardless of which alternative production they come from.
Philippe McLean and R. Nigel Horspool in their paper "A Faster Earley Parser" combine Earley parsing with LR parsing and achieve an improvement in an order of magnitude.