ChartRuleLookupManagerMemory::ChartRuleLookupManagerMemory(
const InputType &src,
const ChartCellCollectionBase &cellColl,
const PhraseDictionarySCFG &ruleTable)
: ChartRuleLookupManagerCYKPlus(src, cellColl)
, m_ruleTable(ruleTable)
{
CHECK(m_dottedRuleColls.size() == 0);
size_t sourceSize = src.GetSize();
m_dottedRuleColls.resize(sourceSize);
const PhraseDictionaryNodeSCFG &rootNode = m_ruleTable.GetRootNode();
for (size_t ind = 0; ind < m_dottedRuleColls.size(); ++ind) {
#ifdef USE_BOOST_POOL
DottedRuleInMemory *initDottedRule = m_dottedRulePool.malloc();
new (initDottedRule) DottedRuleInMemory(rootNode);
#else
DottedRuleInMemory *initDottedRule = new DottedRuleInMemory(rootNode);
#endif
DottedRuleColl *dottedRuleColl = new DottedRuleColl(sourceSize - ind + 1);
dottedRuleColl->Add(0, initDottedRule); // init rule. stores the top node in tree
m_dottedRuleColls[ind] = dottedRuleColl;
}
}
ChartRuleLookupManagerMemory::ChartRuleLookupManagerMemory(
const ChartParser &parser,
const ChartCellCollectionBase &cellColl,
const PhraseDictionaryMemory &ruleTable)
: ChartRuleLookupManagerCYKPlus(parser, cellColl)
, m_ruleTable(ruleTable)
{
UTIL_THROW_IF2(m_dottedRuleColls.size() != 0,
"Dotted rule collection not correctly initialized");
size_t sourceSize = parser.GetSize();
m_dottedRuleColls.resize(sourceSize);
const PhraseDictionaryNodeMemory &rootNode = m_ruleTable.GetRootNode();
for (size_t ind = 0; ind < m_dottedRuleColls.size(); ++ind) {
#ifdef USE_BOOST_POOL
DottedRuleInMemory *initDottedRule = m_dottedRulePool.malloc();
new (initDottedRule) DottedRuleInMemory(rootNode);
#else
DottedRuleInMemory *initDottedRule = new DottedRuleInMemory(rootNode);
#endif
DottedRuleColl *dottedRuleColl = new DottedRuleColl(sourceSize - ind + 1);
dottedRuleColl->Add(0, initDottedRule); // init rule. stores the top node in tree
m_dottedRuleColls[ind] = dottedRuleColl;
}
}
// Given a partial rule application ending at startPos-1 and given the sets of
// source and target non-terminals covering the span [startPos, endPos],
// determines the full or partial rule applications that can be produced through
// extending the current rule application by a single non-terminal.
void ChartRuleLookupManagerMemory::ExtendPartialRuleApplication(
const DottedRuleInMemory &prevDottedRule,
size_t startPos,
size_t endPos,
size_t stackInd,
DottedRuleColl & dottedRuleColl)
{
// source non-terminal labels for the remainder
const NonTerminalSet &sourceNonTerms =
GetSentence().GetLabelSet(startPos, endPos);
// target non-terminal labels for the remainder
const ChartCellLabelSet &targetNonTerms = GetTargetLabelSet(startPos, endPos);
// note where it was found in the prefix tree of the rule dictionary
const PhraseDictionaryNodeSCFG &node = prevDottedRule.GetLastNode();
const PhraseDictionaryNodeSCFG::NonTerminalMap & nonTermMap =
node.GetNonTerminalMap();
const size_t numChildren = nonTermMap.size();
if (numChildren == 0) {
return;
}
const size_t numSourceNonTerms = sourceNonTerms.size();
const size_t numTargetNonTerms = targetNonTerms.GetSize();
const size_t numCombinations = numSourceNonTerms * numTargetNonTerms;
// We can search by either:
// 1. Enumerating all possible source-target NT pairs that are valid for
// the span and then searching for matching children in the node,
// or
// 2. Iterating over all the NT children in the node, searching
// for each source and target NT in the span's sets.
// We'll do whichever minimises the number of lookups:
if (numCombinations <= numChildren*2) {
// loop over possible source non-terminal labels (as found in input tree)
NonTerminalSet::const_iterator p = sourceNonTerms.begin();
NonTerminalSet::const_iterator sEnd = sourceNonTerms.end();
for (; p != sEnd; ++p) {
const Word & sourceNonTerm = *p;
// loop over possible target non-terminal labels (as found in chart)
ChartCellLabelSet::const_iterator q = targetNonTerms.begin();
ChartCellLabelSet::const_iterator tEnd = targetNonTerms.end();
for (; q != tEnd; ++q) {
const ChartCellLabel &cellLabel = q->second;
// try to match both source and target non-terminal
const PhraseDictionaryNodeSCFG * child =
node.GetChild(sourceNonTerm, cellLabel.GetLabel());
// nothing found? then we are done
if (child == NULL) {
continue;
}
// create new rule
#ifdef USE_BOOST_POOL
DottedRuleInMemory *rule = m_dottedRulePool.malloc();
new (rule) DottedRuleInMemory(*child, cellLabel, prevDottedRule);
#else
DottedRuleInMemory *rule = new DottedRuleInMemory(*child, cellLabel,
prevDottedRule);
#endif
dottedRuleColl.Add(stackInd, rule);
}
}
}
else
{
// loop over possible expansions of the rule
PhraseDictionaryNodeSCFG::NonTerminalMap::const_iterator p;
PhraseDictionaryNodeSCFG::NonTerminalMap::const_iterator end =
nonTermMap.end();
for (p = nonTermMap.begin(); p != end; ++p) {
// does it match possible source and target non-terminals?
const PhraseDictionaryNodeSCFG::NonTerminalMapKey &key = p->first;
const Word &sourceNonTerm = key.first;
if (sourceNonTerms.find(sourceNonTerm) == sourceNonTerms.end()) {
continue;
}
const Word &targetNonTerm = key.second;
const ChartCellLabel *cellLabel = targetNonTerms.Find(targetNonTerm);
if (!cellLabel) {
continue;
}
// create new rule
const PhraseDictionaryNodeSCFG &child = p->second;
#ifdef USE_BOOST_POOL
DottedRuleInMemory *rule = m_dottedRulePool.malloc();
new (rule) DottedRuleInMemory(child, *cellLabel, prevDottedRule);
#else
DottedRuleInMemory *rule = new DottedRuleInMemory(child, *cellLabel,
prevDottedRule);
#endif
dottedRuleColl.Add(stackInd, rule);
}
}
}
