- LICENSE AND COPYRIGHT
Marpa::Doc::Algorithm - The Marpa Algorithm
Marpa combines many parsing techniques, and a full description would be a textbook. I have to content myself with briefly describing the more significant ideas new with Marpa. In this document I assume that the reader is familiar with parsing in general, and that he understands terms like Earley set and Earley item. It will be helpful to have read the other Marpa documents, up to and including Marpa::Doc::Internals.
All claims of originality are limited by my ignorance. The parsing literature is large, and I may have been preceded by someone whose work I'm not aware of. Readers familiar with scholarly literature will know that a claim of originality in combining A and B, is not to be read as a claim of originality in either A or B.
Marpa takes the recognition engine as described by Horspool and Aycock and turns it inside out and upside down. In the Horspool and Aycock version of Earley's, the main loop iterated over each item of an Earley set, first scanning for tokens, then completing items in the set. Marpa turns this into two separate loops over the Earley items, the first of which completes the current earley set, and the second of which scans for tokens.
The advantage of this double inversion is that the lexical analysis can be put between the two loops -- after completion, but before scanning. This makes predictive lexing possible. Being able to predict which lexables are legal at a given point in the parse can save a lot of processing, especially if the lexing is complex. Any lexemes which are not predicted by items in the current Earley set do not need to be scanned for. Now that Hayspool and Aycock have sped up Earley's algorithm, the time spent lexing is a significant factor in overall speed. Predictive lexing can reduce lexing time to a fraction of the original.
Marpa allows ambiguous lexing, including recognition of lexemes of different lengths starting at the same lexical position, and recognition of overlapping lexemes. To facilitate this, Marpa introduces the earleme (named after Jay Earley). Previous Earley implementations required the input to be broken up into tokens, usually by lexical analysis of the input using DFA's. (DFA's -- deterministic finite automata -- are the equivalent of regular expressions when that term is used in the strictest sense). Requiring that the first level of analysis be performed by a DFA hobbles a general parser like Earley's.
Marpa allows the scanning phase of Earley's algorithm to add items not just to the next Earley set, but to any later one. Alternative scans of the input can be put into the Earley sets, and the power of Earley's algorithm harnessed to deal with the indeterminism.
Marpa does this by allowing each scanned token to have a length in earlemes. The earleme is a unit of distance measured in Earley sets. The first Earley set at earleme 0. Subsequent Earley sets are at earleme 1, 2, etc. If the length of a scanned token is L, and the current Earley set is C, a newly scanned Earley item is added to Earley set C+L.
An implementation can sync earlemes up with any measure that's convenient. For example, the distance in earlemes may be the length of a string, measured either in bytes or Unicode graphemes. An implementation may also mimic traditional lexing by defining the earleme to be the same as the distance in a token stream, measured in tokens.
Another significant change to the Aycock and Horspool algorithm in Marpa is its internal rewriting of the grammar. Aycock and Horspool rewrote their grammars internally into what they called NNF (Nihilist Normal Form). Earley's original algorithm had serious issues with nullable symbols and productions, and the NNF rewrite fixes almost all of them. (A nullable symbol or production is one which derives the empty sentence.) Importantly, NNF also allows complete and easy mapping of the semantics of the original grammar to its NNF rewrite, so that NNF and the whole rewrite process can be made invisible to the user.
A problem with NNF is that the rewritten grammar is exponentially larger than the original in the theoretical worst case. I think cases could arise in practice where the NNF size explosion is a real problem. One such case might be Perl 6 rules in which whitespace is significant but optional.
Marpa's solution is Chomsky-Horspool-Aycock Form (CHAF). This is Horspool and Aycock's NNF, but with the further restriction that no more than two nullable symbols may appear in any production. (The idea that any context-free grammar can be rewritten into productions of at most a small fixed size appears to originate, like so much else, with Noam Chomsky.) The shortened CHAF production maps back to the original grammar, so that like NNF, the CHAF rewrite can be made invisible to the user. With CHAF, the theoretical worst behavior is linear, and in those difficult cases likely to arise in practice the multiplier is much smaller.
Aycock and Horspool give an algorithm for constructing a rightmost derivation from their version of the Earley sets. They suggest that in the case of multiple parses, their Earley sets could be iterated through, and they point out where the decision points occur in their algorithm. But they give no algorithm to do that iteration.
Marpa's method of evaluation is to first create a parse bocage from the Earley sets. Marpa generates the parse trees from the parse bocage. Parse trees are kept as an array, in pre-order form. Parse trees are evaluated by scanning the array in reverse, so that they are seen in post-order and a conventional evaluation stack can be used.
A ambiguous parse is a set of parse trees, and in the parsing literature there is an efficient and compact means of storing a set of closely related parse trees. It is called, aptly, a parse forest. Nodes in a parse forest are divided into and-nodes and or-nodes. And-nodes are individual pieces of parse trees. In conventional parse forests, each and-node represents a production, as in a parse tree, with the lhs as the parent node and the rhs symbols as child nodes. Or-nodes represent choices between and-nodes, with each child of an or-node being a choice. A parse tree is generated from a parse forest by traversing the forest and selecting one child node at every or-node.
Marpa uses a modified form of parse forest, which I will call a parse bocage. Marpa could not use standard parse forests for two reasons. First, parse forests not only contain trees, but they themselves form a tree. However, the data produced by Marpa's recognizer may contain cycles. Therefore, the data structure Marpa needs is not a tree, strictly speaking, but a graph.
Second, the productions of the grammar are not repesented intact when the Marpa evaluator finds them in the Earley items. Instead, each production is broken up, and represented as links between Earley items. This system of links comes from Aycock and Horspool.
In Marpa's elaboration of the Aycock-Horspool scheme, each Earley item has a list of sources. Each source can have two elements: a predecessor and a cause.
The predecessor is optional. If present, it is a link to an Earley item. It's a predecessor link in the sense that both the linking and the linked Earley item are pieces of the same production, and the linked Earley item is an earlier piece.
The cause element of the source is always present. It can be a token or a link to a child Earley item. The Earley item linked by the cause element is a child in the sense that if the productions from the two Earley items were represented in a parse tree, the production from the linking Earley item would be the parent, and the production from the linked Earley item would be the child.
In effect, the sources in the Earley items contain the original grammar rewritten into productions with at most two symbols on the right hand side. This is basically Chomsky Normal Form (CNF).
A CNF grammar can be represented conveniently as a binary tree. Marpa could have restored the original, non-binary structure of the grammar when it created the parse bocage from the Earley items. But, for two reasons, that doesn't make sense.
First, combining CNF with QDFA states produces a representation even more compact than a conventional parse forest. Secondly, the code to traverse, iterate and evaluate binary trees can be considerably simpler than code which needs to deal with nodes which can have an arbitrary number of child nodes. Rather than rewrite the grammar into a form that's harder to process, it makes sense to leave the grammar the way it is found in the Earley items.
For this reason, the and-nodes in parse bocages do not directly represent productions of the grammar. Instead productions are broken up, so that every and-node has at most two children.
The bocage forests of Normandy and Brittany are networks of dense hedgerows cultivated over centuries as obstacles to livestock and armies. These hedges are impressive obstacles even to modern infantry, and the Norman bocage played a major role in World War II. The term "parse bocage" seemed appropriate for the strategic thicket that Marpa uses to store parses.
To find the next parse tree in an ambiguous parse, Marpa scans the current parse tree, comparing it to the parse bocage. The current parse tree is kept in pre-order in an array. To iterate the parse tree, Marpa treats the array like a stack. It starts at the end of the array/stack, and pops node after node, looking for a node to iterate.
Because the pre-order array is scanned from the end to the beginning, the nodes of the parse tree are popped in post-order. As each node is popped, it is compared with the parse bocage. Every node in the parse tree corresponds both to an or-node, and to a current choice of and-node from that or-node. If all the choices of and-node for a parse tree node have already been tried, the popped node is temporarily set aside and the next node on the parse tree array/stack is popped.
If the node just popped from the parse tree has another choice of and-node available, that just-popped node is selected as the interation node. The iteration node is modified to reflect the new choice of and-node. The modified iteration node is then pushed back onto the parse tree array/stack.
Next, the child nodes of the new iteration node are generated and pushed on the parse tree array/stack. Finally, there may have been some parse tree nodes, which were popped from the array/stack, but which don't change and which shouldn't be removed from the parse tree. The parse tree nodes which were popped, but turned out not to be descendents of the iterated node, are pushed back onto the parse tree array/stack.
In order to iterate through the parses, Marpa stores, with most of the Earley items, a list of sources -- information about why the Earley items were added. The order of the sources in these lists controls the parse order. I realized that the sources could be sorted based on priorities assigned to the rules and tokens of the grammar. These priorities could be user-assigned and used to put the most desirable parse first.
In effect, Marpa's rule and terminal priorities turn ambiguous parsing into a convenient form of precedence parsing. I don't believe Earley parsing and precedence parsing have been combined before.
Aycock and Horspool's precomputation uses what they call a split LR(0) ε-DFA. Not exactly the most transparent name, but a brilliant concept: if you leave in some of the non-determinism when you transform an NFA to an LR(0) DFA, you can use it to guide the Earley recognizer.
While the concept gave me an essential insight that I could never have come up on my own, I found the split LR(0) ε-DFA awkward to program. A point came when I wanted to add some more non-determinism, and then I realized that not only would an ordinary NFA allow my change, but that it would make the code shorter and better.
What Marpa does, therefore, is to transform the original NFA into a second NFA. The second NFA keeps the form of an NFA, but has most of the non-determinism taken out of it. It's an NFA with a lot less "N". It is almost an LR(0) DFA, so I call it a quasi-deterministic finite automaton, or QDFA.
See the support section in the main module.
Copyright 2007 - 2009 Jeffrey Kegler
This program is free software; you can redistribute it and/or modify it under the same terms as Perl 5.10.0.