// Setup and run just the initial segsearch on an established matrix, // without doing any additional chopping or joining. // (Internal factored version that can be used as part of the main SegSearch.) void Wordrec::InitialSegSearch(WERD_RES* word_res, LMPainPoints* pain_points, GenericVector<SegSearchPending>* pending, BestChoiceBundle* best_choice_bundle, BlamerBundle* blamer_bundle) { if (segsearch_debug_level > 0) { tprintf("Starting SegSearch on ratings matrix%s:\n", wordrec_enable_assoc ? " (with assoc)" : ""); word_res->ratings->print(getDict().getUnicharset()); } pain_points->GenerateInitial(word_res); // Compute scaling factor that will help us recover blob outline length // from classifier rating and certainty for the blob. float rating_cert_scale = -1.0 * getDict().certainty_scale / rating_scale; language_model_->InitForWord(prev_word_best_choice_, assume_fixed_pitch_char_segment, segsearch_max_char_wh_ratio, rating_cert_scale); // Initialize blamer-related information: map character boxes recorded in // blamer_bundle->norm_truth_word to the corresponding i,j indices in the // ratings matrix. We expect this step to succeed, since when running the // chopper we checked that the correct chops are present. if (blamer_bundle != NULL) { blamer_bundle->SetupCorrectSegmentation(word_res->chopped_word, wordrec_debug_blamer); } // pending[col] tells whether there is update work to do to combine // best_choice_bundle->beam[col - 1] with some BLOB_CHOICEs in matrix[col, *]. // As the language model state is updated, pending entries are modified to // minimize duplication of work. It is important that during the update the // children are considered in the non-decreasing order of their column, since // this guarantees that all the parents would be up to date before an update // of a child is done. pending->init_to_size(word_res->ratings->dimension(), SegSearchPending()); // Search the ratings matrix for the initial best path. (*pending)[0].SetColumnClassified(); UpdateSegSearchNodes(rating_cert_scale, 0, pending, word_res, pain_points, best_choice_bundle, blamer_bundle); }
/** * @name improve_by_chopping * * Repeatedly chops the worst blob, classifying the new blobs fixing up all * the data, and incrementally runs the segmentation search until a good word * is found, or no more chops can be found. */ void Wordrec::improve_by_chopping(float rating_cert_scale, WERD_RES* word, BestChoiceBundle* best_choice_bundle, BlamerBundle* blamer_bundle, LMPainPoints* pain_points, GenericVector<SegSearchPending>* pending) { int blob_number; do { // improvement loop. // Make a simple vector of BLOB_CHOICEs to make it easy to pick which // one to chop. GenericVector<BLOB_CHOICE*> blob_choices; int num_blobs = word->ratings->dimension(); for (int i = 0; i < num_blobs; ++i) { BLOB_CHOICE_LIST* choices = word->ratings->get(i, i); if (choices == NULL || choices->empty()) { blob_choices.push_back(NULL); } else { BLOB_CHOICE_IT bc_it(choices); blob_choices.push_back(bc_it.data()); } } SEAM* seam = improve_one_blob(blob_choices, &best_choice_bundle->fixpt, false, false, word, &blob_number); if (seam == NULL) break; // A chop has been made. We have to correct all the data structures to // take into account the extra bottom-level blob. // Put the seam into the seam_array and correct everything else on the // word: ratings matrix (including matrix location in the BLOB_CHOICES), // states in WERD_CHOICEs, and blob widths. word->InsertSeam(blob_number, seam); // Insert a new entry in the beam array. best_choice_bundle->beam.insert(new LanguageModelState, blob_number); // Fixpts are outdated, but will get recalculated. best_choice_bundle->fixpt.clear(); // Remap existing pain points. pain_points->RemapForSplit(blob_number); // Insert a new pending at the chop point. pending->insert(SegSearchPending(), blob_number); // Classify the two newly created blobs using ProcessSegSearchPainPoint, // as that updates the pending correctly and adds new pain points. MATRIX_COORD pain_point(blob_number, blob_number); ProcessSegSearchPainPoint(0.0f, pain_point, "Chop1", pending, word, pain_points, blamer_bundle); pain_point.col = blob_number + 1; pain_point.row = blob_number + 1; ProcessSegSearchPainPoint(0.0f, pain_point, "Chop2", pending, word, pain_points, blamer_bundle); if (language_model_->language_model_ngram_on) { // N-gram evaluation depends on the number of blobs in a chunk, so we // have to re-evaluate everything in the word. ResetNGramSearch(word, best_choice_bundle, pending); blob_number = 0; } // Run language model incrementally. (Except with the n-gram model on.) UpdateSegSearchNodes(rating_cert_scale, blob_number, pending, word, pain_points, best_choice_bundle, blamer_bundle); } while (!language_model_->AcceptableChoiceFound() && word->ratings->dimension() < kMaxNumChunks); // If after running only the chopper best_choice is incorrect and no blame // has been yet set, blame the classifier if best_choice is classifier's // top choice and is a dictionary word (i.e. language model could not have // helped). Otherwise blame the tradeoff between the classifier and // the old language model (permuters). if (word->blamer_bundle != NULL && word->blamer_bundle->incorrect_result_reason() == IRR_CORRECT && !word->blamer_bundle->ChoiceIsCorrect(word->best_choice)) { bool valid_permuter = word->best_choice != NULL && Dict::valid_word_permuter(word->best_choice->permuter(), false); word->blamer_bundle->BlameClassifierOrLangModel(word, getDict().getUnicharset(), valid_permuter, wordrec_debug_blamer); } }