/ 
Marpa-PP-0.005_008
/ Marpa::PP::Semantics::Order

NAME

Marpa::PP::Semantics::Order - How Marpa Ranks Ambiguous Parses

DESCRIPTION

This document details the order in which the recognizer's value method returns parse results. The same mechanism allows the selection of parse results. It can also be exploited to do the actual processing of parses, using side effects.

Duplicate Parses are Eliminated

In a single parse series, Marpa will never return the same parse result twice. Marpa regards two parses as being the same if they are semantic duplicates.

Two parses are semantic duplicates if a recursive, top-down evaluation of each applies the same rules in the same order at the same earleme locations. When this is the case, a deterministic semantics will always produce the same value for both parses -- hence the term "semantic duplicate". When the Marpa documentation refers to duplicate parses, it will mean semantic duplicates unless otherwise stated.

Default Parse Order

By calling the recognizer's value method repeatedly, Marpa can produce all the parse results for a given parse. The default is for the parse results to be returned in an arbitrary order. This corresponds to the "none" value of the recognizer's ranking_method named argument.

A General Approach to Sorting Parses

The most general way to sort Marpa parses is for the application to take control. The application can set up the Marpa semantic actions so that the value of every parse result is a <rank, true_value> duple. The duples can then be sorted by rank. Once the resuls are sorted, the rank element of the duple can be discarded. (Those familiar with the Schwartzian transform may note a resemblance. In Perl, duples can be implemented as references to arrays of 2 elements.)

The user needs to be careful. In theory, ambiguity can cause an exponential explosion in the number of results. In practice, ambiguity tends to get out of hand very easily. Producing and sorting all the parses can take a very long time.

The Constant Ranking Method

Marpa supports a simplified approach to sorting parses. The Constant Ranking Method is general enough to handle many, perhaps even most, practical applications, and simple enough that Marpa is able to optimize it. The Constant Ranking Method is specified by giving the recognizer's ranking_method a value of "constant".

The basic idea is to allow the user to specify constant values for rules, and to rank all other nodes according to the sum of the values of their children. Leaf nodes default to a value of 0.

When specified, the value of a rule must be "constant" in the sense that it cannot depend on the values of its children. This is a major limitation, but it greatly simplifies the logic for re-ranking parses as they are iterated. And it is less of a limitation than it may appear, because rules, if they do not have ranking actions, will take into account the values of their children. By strategically mixing rules that take into account child values, with rules that can have a user-specified rank, applications will often be able to accomplish what they need to.

The ranking action of a token leaf node is specified using the token symbol's ranking_action property. The ranking action of a nulled leaf node is specified using the null node symbol's ranking_action property. The ranking action of a rule is specified using the rule's ranking_action property.

Ranking actions must return a reference to the rank. Ranks must be Perl numbers. Negative values and non-integer values are allowed. The highest numeric value is considered to be the highest rank, and the lowest numeric value is considered to be the lowest rank.

As a special case, if a ranking action returns a Perl undef, Marpa will skip that possibility, rather than ranking it. Note that any "skipped" node in a parse result causes that whole parse result to be skipped. A consequence of this is that any skipped node in an unambiguous parse will result in no parse results being found.

This behavior may seem to be draconian, but in fact skipping the entire tree is the most natural way to deal with skipped nodes. Anything resembling a traditional semantics requires parse trees to have a full set of nodes. And it is unclear what purpose an alternative semantics might be expected to serve.

An instance of a rule is a rule, a start location, and an end location. Ranking actions are called once for each rule instance. While ranking actions return constants in the sense that they cannot be aware of the ranks of their child nodes, the rank returned can vary based on the rule's start and end location. Ranking actions can determine their location using the context-aware static methods.

For the rank of a node to be calculated, the ranking action must first be resolved to a ranking Perl closure. Ranking action names are resolved to ranking Perl closures in the Ranking Phase, using the same logic that resolves semantic actions to semantic Perl closures. The logic that resolves action names to Perl closures is described elsewhere ("RESOLVING ACTION NAMES" in Marpa::PP::Semantics). The ranking Perl closures are both resolved and called in a single Ranking Phase.

Exploiting Side Effects

In every parse series, ranking actions are guaranteed to be called once and only once for each rule instance. As a reminder, a rule instance is a rule, together with a start and end location. This guarantee makes ranking actions useful for their side effects, even when there is no interest in changing the order of the parse results. In fact, ranking actions can be used in cases when there is no interest in evaluating the actual parse results.

For example, an application may be interested in detecting a particular kind of ambiguity: it may wants to know, for two specific rules, when they derive the same input string. To do this, the application can create ranking actions which have the side effect of tracking all instances of these two rules, by location. If there is no interest in an actual parse, the ranking actions can return undef, which will cause all parses to be discarded.

For applications like ambiguity detection, an implementation that uses the ranking actions can be much faster than one which produces all the parse results. In practical cases, the number of parse results grows much more rapidly than the number of ambiguities. In the worst cases, the number of calls to ranking actions can be O(n**3) in the length of the input. While far from fast, this is much better than the worst case for evaluating all the parses, which is O(e**n).

COPYRIGHT AND LICENSE

Copyright 2011 Jeffrey Kegler
This file is part of Marpa::PP.  Marpa::PP is free software: you can
redistribute it and/or modify it under the terms of the GNU Lesser
General Public License as published by the Free Software Foundation,
either version 3 of the License, or (at your option) any later version.

Marpa::PP is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
Lesser General Public License for more details.

You should have received a copy of the GNU Lesser
General Public License along with Marpa::PP.  If not, see
http://www.gnu.org/licenses/.