///functor: Run composition itself, separate from load times (e.g. which may include an fst expansion).
VectorFst<Arc> * operator() () {
if (!fst_.NumStates() ) {
LWARN ("Empty lattice. ... Skipping LM application!");
return NULL;
}
while ( qc_.size() ) {
StateId s = qc_.front();
qc_.pop();
pair<StateId, const typename KenLMModelT::State> p = get ( s );
StateId& s1 = p.first;
const typename KenLMModelT::State s2 = p.second;
for ( ArcIterator< VectorFst<Arc> > arc1 ( fst_, s1 ); !arc1.Done();
arc1.Next() ) {
const Arc& a1 = arc1.Value();
float w = 0;
float wp = wp_;
typename KenLMModelT::State nextlmstate;
if ( epsilons_.find ( a1.olabel ) == epsilons_.end() ) {
w = lmmodel_.Score ( s2, idbridge_.map (a1.olabel), nextlmstate ) * natlog10_;
//silly hack
if ( a1.olabel <= 2 ) {
wp = 0;
if (a1.olabel == 1 ) w = 0; //get same result as srilm
}
} else {
nextlmstate = s2;
wp = 0; //We don't count epsilon labels
}
pair<StateId, bool> nextp = add ( nextlmstate
, a1.nextstate
, fst_.Final ( a1.nextstate ) );
StateId& newstate = nextp.first;
bool visited = nextp.second;
composed_->AddArc ( s
, Arc ( a1.ilabel, a1.olabel
, Times ( a1.weight, Times (mw_ ( w ) , mw_ (wp) ) )
, newstate ) );
//Finally, only add newstate to the queue if it hasn't been visited previously
if ( !visited ) {
qc_.push ( newstate );
}
}
}
LINFO ( "Done! Number of states=" << composed_->NumStates() );
return composed_;
};
/**
* \brief Adds a state.
* \return true if the state requested has already been visited, false otherwise.
*/
inline pair <StateId, bool> add ( typename KenLMModelT::State& m2nextstate,
StateId m1nextstate, Weight m1stateweight ) {
static StateId lm = 0;
getIdx ( m2nextstate );
///New history:
if ( seenlmstates_.find ( history ) == seenlmstates_.end() ) {
seenlmstates_[history] = ++lm;
}
uint64_t compound = m1nextstate * sid + seenlmstates_[history];
LDEBUG ( "compound id=" << compound );
if ( stateexistence_.find ( compound ) == stateexistence_.end() ) {
LDEBUG ( "New State!" );
statemap_[composed_->NumStates()] =
pair<StateId, const typename KenLMModelT::State > ( m1nextstate, m2nextstate );
composed_->AddState();
if ( m1stateweight != mw_ ( ZPosInfinity() ) ) composed_->SetFinal (
composed_->NumStates() - 1, m1stateweight );
stateexistence_[compound] = composed_->NumStates() - 1;
return pair<StateId, bool> ( composed_->NumStates() - 1, false );
}
return pair<StateId, bool> ( stateexistence_[compound], true );
};
/**
* Constructor. Initializes on-the-fly composition with a language model.
* \param fst Machine you want to apply the language to. Pass a delayed machine if you can, as it will expand it in constructor.
* \param model A KenLM language model
* \param epsilons List of words to work as epsilons
* \param natlog Use or not natural logs
* \param lmscale Language model scale
*/
ApplyLanguageModelOnTheFly ( const Fst<Arc>& fst,
KenLMModelT& model,
#ifndef USE_GOOGLE_SPARSE_HASH
unordered_set<Label>& epsilons,
#else
google::dense_hash_set<Label>& epsilons,
#endif
bool natlog,
float lmscale ,
float lmwp,
const IdBridgeT& idbridge
) :
composed_ ( NULL ) ,
natlog10_ ( natlog ? -lmscale* ::log ( 10.0 ) : -lmscale ),
fst_ ( fst ),
lmmodel_ ( model ),
vocab_ ( model.GetVocabulary() ),
wp_ ( lmwp ) ,
epsilons_ ( epsilons ) ,
history ( model.Order(), 0),
idbridge_ (idbridge) {
#ifdef USE_GOOGLE_SPARSE_HASH
stateexistence_.set_empty_key ( numeric_limits<ull>::max() );
statemap_.set_empty_key ( numeric_limits<uint64_t>::max() );
basic_string<unsigned> aux (KENLM_MAX_ORDER, numeric_limits<unsigned>::max() );
seenlmstates_.set_empty_key ( aux );
#endif
buffersize = ( model.Order() - 1 ) * sizeof ( unsigned int );
buffer = const_cast<unsigned *> ( history.c_str() );
if (!fst_.NumStates() ) {
LWARN ("Empty lattice");
return;
}
composed_ = new VectorFst<Arc>;
typename KenLMModelT::State bs = model.NullContextState();
///Initialize with first state
pair<StateId, bool> nextp = add ( bs, fst_.Start(),
fst_.Final ( fst_.Start() ) );
qc_.push ( nextp.first );
composed_->SetStart ( nextp.first );
};
vector<PathData> M2MFstAligner::write_alignment(const VectorFst<LogArc> &ifst,
int nbest)
{
//Generic alignment generator
VectorFst<StdArc> fst;
Map(ifst, &fst, LogToStdMapper());
for (StateIterator<VectorFst<StdArc> > siter(fst); !siter.Done();
siter.Next()) {
StdArc::StateId q = siter.Value();
for (MutableArcIterator<VectorFst<StdArc> > aiter(&fst, q);
!aiter.Done(); aiter.Next()) {
//Prior to decoding we make several 'heuristic' modifications to the weights:
// 1. A multiplier is applied to any multi-token substrings
// 2. Any LogWeight::Zero() arc weights are reset to '99'.
// We are basically resetting 'Infinity' values to a 'smallest non-Infinity'
// so that the ShortestPath algorithm actually produces something no matter what.
// 3. Any arcs that consist of subseq1:subseq2 being the same length and subseq1>1
// are set to '99' this forces shortestpath to choose arcs where one of the
// following conditions holds true
// * len(subseq1)>1 && len(subseq2)!=len(subseq1)
// * len(subseq2)>1 && len(subseq1)!=len(subseq2)
// * len(subseq1)==len(subseq2)==1
//I suspect these heuristics can be eliminated with a better choice of the initialization
// function and maximization function, but this is the way that m2m-aligner works, so
// it makes sense for our first cut implementation.
//In any case, this guarantees that M2MFstAligner produces results identical to those
// produced by m2m-aligner - but with a bit more reliability.
//UPDATE: this now produces a better alignment than m2m-aligner.
// The maxl heuristic is still in place. The aligner will produce *better* 1-best alignments
// *without* the maxl heuristic below, BUT this comes at the cost of producing a less
// flexible corpus. That is, for a small training corpus like nettalk, if we use the
// best alignment we wind up with more 'chunks' and thus get a worse coverage for unseen
// data. Using the aignment lattices to train the joint ngram model solves this problem.
// Oh baby. Can't wait to for everyone to see the paper!
//NOTE: this is going to fail if we encounter any alignments in a new test item that never
// occurred in the original model.
StdArc
arc = aiter.Value();
int
maxl = get_max_length(isyms->Find(arc.ilabel));
if (maxl == -1) {
arc.weight = 999;
}
else {
//Optionally penalize m-to-1 / 1-to-m links. This produces
// WORSE 1-best alignments, but results in better joint n-gram
// models for small training corpora when using only the 1-best
// alignment. By further favoring 1-to-1 alignments the 1-best
// alignment corpus results in a more flexible joint n-gram model
// with regard to previously unseen data.
//if( penalize==true ){
arc.weight = alignment_model[arc.ilabel].Value() * maxl;
//}else{
//For larger corpora this is probably unnecessary.
//arc.weight = alignment_model[arc.ilabel].Value();
//}
}
if (arc.weight == LogWeight::Zero())
arc.weight = 999;
if (arc.weight != arc.weight)
arc.weight = 999;
aiter.SetValue(arc);
}
}
VectorFst<StdArc> shortest;
ShortestPath(fst, &shortest, nbest);
RmEpsilon(&shortest);
//Skip empty results. This should only happen
// in the following situations:
// 1. seq1_del=false && len(seq1)<len(seq2)
// 2. seq2_del=false && len(seq1)>len(seq2)
//In both 1.and 2. the issue is that we need to
// insert a 'skip' in order to guarantee at least
// one valid alignment path through seq1*seq2, but
// user params didn't allow us to.
//Probably better to insert these where necessary
// during initialization, regardless of user prefs.
if (shortest.NumStates() == 0) {
vector<PathData> dummy;
return dummy;
}
FstPathFinder
pathfinder(skipSeqs);
pathfinder.isyms = isyms;
pathfinder.findAllStrings(shortest);
return pathfinder.paths;
}
