vpc.c

/* Produce a vector-per-class description of the model data in a barrel */

/* Copyright (C) 1997, 1998, 1999 Andrew McCallum

   Written by:  Andrew Kachites McCallum <mccallum@cs.cmu.edu>

   This file is part of the Bag-Of-Words Library, `libbow'.

   This library is free software; you can redistribute it and/or
   modify it under the terms of the GNU Library General Public License
   as published by the Free Software Foundation, version 2.
   
   This library 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
   Library General Public License for more details.

   You should have received a copy of the GNU Library General Public
   License along with this library; if not, write to the Free Software
   Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111, USA */

#include <bow/libbow.h>

double bow_wi2dvf_sum (bow_wi2dvf *wi2dvf)
{
  int wi, max_wi, dvi;
  double sum = 0;
  bow_dv *dv;
  max_wi = MIN (wi2dvf->size, bow_num_words ());

  for (wi = 0; wi < max_wi; wi++)
    {
      dv = bow_wi2dvf_dv (wi2dvf, wi);
      if (!dv)
	continue;
      for (dvi = 0; dvi < dv->length; dvi++)
	sum += dv->entry[dvi].weight;
    }
  return sum;
}

/* Given a barrel of documents, create and return another barrel with
   only one vector per class. The classes will be represented as
   "documents" in this new barrel. */
bow_barrel *
bow_barrel_new_vpc (bow_barrel *doc_barrel)
{
  bow_barrel* vpc_barrel;	/* The vector per class barrel */
  int max_ci = -1;		/* The highest index of encountered classes */
  int num_classes = bow_barrel_num_classes (doc_barrel);
  int wi;
  int max_wi;
  int dvi;
  int ci;
  bow_dv *dv;
  bow_dv *vpc_dv;
  int di;
  int num_docs_per_ci[num_classes];
  bow_cdoc *cdoc;
  double sum = 0;

  assert (doc_barrel->classnames);

  max_wi = MIN (doc_barrel->wi2dvf->size, bow_num_words ());

  /* Create an empty barrel; we fill fill it with vector-per-class
     data and return it. */
  /* This assertion can fail when DOC_BARREL was read from a disk
     archive that was created before CLASS_PROBS was added to BOW_CDOC */
  assert (doc_barrel->cdocs->entry_size >= sizeof (bow_cdoc));
  vpc_barrel = bow_barrel_new (doc_barrel->wi2dvf->size,
			       num_classes,
			       doc_barrel->cdocs->entry_size,
			       doc_barrel->cdocs->free_func);
  vpc_barrel->method = doc_barrel->method;
  vpc_barrel->classnames = bow_int4str_new (0);
  /* Make sure to set the VPC indicator */
  vpc_barrel->is_vpc = 1;

  bow_verbosify (bow_verbose, "Making vector-per-class... words ::       ");

  /* Count the number of documents in each class */
  for (ci = 0; ci < num_classes; ci++)
    num_docs_per_ci[ci] = 0;
  for (di = 0; di < doc_barrel->cdocs->length; di++)
    {
      cdoc = bow_array_entry_at_index (doc_barrel->cdocs, di);
      if (cdoc->type == bow_doc_train)
	num_docs_per_ci[cdoc->class]++;
    }

  /* Update the CDOC->WORD_COUNT in the DOC_BARREL in order to match
     the (potentially) pruned vocabulary. */
  {
    bow_wv *wv = NULL;
    int wvi;
    bow_dv_heap *heap = bow_test_new_heap (doc_barrel);
    while ((di = bow_heap_next_wv (heap, doc_barrel, &wv,
				   bow_cdoc_yes)) != -1)
      {
	cdoc = bow_array_entry_at_index (doc_barrel->cdocs, di);
	cdoc->word_count = 0;
	for (wvi = 0; wvi < wv->num_entries; wvi++)
	  {
	    if (bow_wi2dvf_dv (doc_barrel->wi2dvf, wv->entry[wvi].wi))
	      cdoc->word_count += wv->entry[wvi].count;
	  }
      }
  }

  /* Initialize the WI2DVF part of the VPC_BARREL.  Sum together the
     counts and weights for individual documents, grabbing only the
     training documents. */
  for (wi = 0; wi < max_wi; wi++)
    {
      dv = bow_wi2dvf_dv (doc_barrel->wi2dvf, wi);
      if (!dv)
	continue;
      for (dvi = 0; dvi < dv->length; dvi++)
	{
	  di = dv->entry[dvi].di;
	  cdoc = bow_array_entry_at_index (doc_barrel->cdocs, di);
	  ci = cdoc->class;
	  assert (ci >= 0);
	  assert (ci < num_classes);
	  if (ci > max_ci)
	    max_ci = ci;
	  if (cdoc->type == bow_doc_train)
	    {
	      float weight;

	      /* The old version of bow_wi2dvf_add_di_text_fp() initialized
		 the dv WEIGHT to 0 instead of the word count.  If the weight 
		 is zero, then use the count instead.  Note, however, that
		 the TFIDF method might have set the weight, so we don't
		 want to use the count all the time. */
	      if (dv->entry[dvi].weight)
		weight = dv->entry[dvi].weight;
	      else
		weight = dv->entry[dvi].count;

	      if (bow_event_model == bow_event_document)
		{
		  assert (dv->entry[dvi].count);
		  bow_wi2dvf_add_wi_di_count_weight (&(vpc_barrel->wi2dvf), 
						     wi, ci, 1, 1);
		}
	      else if (bow_event_model == bow_event_document_then_word)
		{
		  bow_wi2dvf_add_wi_di_count_weight
		    (&(vpc_barrel->wi2dvf), wi, ci, dv->entry[dvi].count,
		     (bow_event_document_then_word_document_length
		      * weight / cdoc->word_count));
		  sum += (bow_event_document_then_word_document_length
			  * weight / cdoc->word_count);
		}
	      else
		{
		  bow_wi2dvf_add_wi_di_count_weight (&(vpc_barrel->wi2dvf), 
						     wi, ci, 
						     dv->entry[dvi].count,
						     weight);
		}
	    }
	}
      /* Set the IDF of the class's wi2dvf directly from the doc's wi2dvf */
      vpc_dv = bow_wi2dvf_dv (vpc_barrel->wi2dvf, wi);
      if (vpc_dv)		/* this could be null if all of this word's
				   occurrences are in non training docs */
	vpc_dv->idf = dv->idf;
      if (max_wi - wi % 100 == 0)
	bow_verbosify (bow_verbose, "\b\b\b\b\b\b%6d", max_wi - wi);
    }
#if 0
  bow_verbosify (bow_progress, "vpc_sum=%f\n", sum);
  bow_verbosify (bow_progress, "wi2dvf_sum=%f\n", 
		 bow_wi2dvf_sum (vpc_barrel->wi2dvf));
#endif
  bow_verbosify (bow_verbose, "\b\b\b\b\b\b");
  /* xxx OK to have some classes with no words
     assert (num_classes-1 == max_ci); */
  if (max_ci < 0)
    {
      int i;
      bow_verbosify (bow_progress, "%s: No data found for ",
		     __PRETTY_FUNCTION__);
      for (i = 0; i < num_classes; i++)
	bow_verbosify (bow_progress, "%s ", 
		       bow_barrel_classname_at_index (doc_barrel, i));
      bow_verbosify (bow_progress, "\n");
    }
  bow_verbosify (bow_verbose, "\n");

  /* Initialize the CDOCS and CLASSNAMES parts of the VPC_BARREL.
     Create BOW_CDOC structures for each class, and append them to the
     VPC->CDOCS array. */
  for (ci = 0; ci < num_classes; ci++)
    {
      bow_cdoc cdoc;
      const char *classname = NULL;

      cdoc.type = bow_doc_train;
      cdoc.normalizer = -1.0f;
      /* Make WORD_COUNT be the number of documents in the class.
         This is for the document event model.*/
      cdoc.word_count = num_docs_per_ci[ci];
      if (doc_barrel->classnames)
	{
	  classname = bow_barrel_classname_at_index (doc_barrel, ci);
	  cdoc.filename = strdup (classname);
	  if (!cdoc.filename)
	    bow_error ("Memory exhausted.");
	}
      else
	{
	  cdoc.filename = NULL;
	}
      cdoc.class_probs = NULL;
      cdoc.class = ci;
      bow_verbosify (bow_verbose, "%20d model documents in class `%s'\n",
		     num_docs_per_ci[ci], cdoc.filename);
      /* Add a CDOC for this class to the VPC_BARREL */
      bow_array_append (vpc_barrel->cdocs, &cdoc);
      /* Add an entry for this class into the VPC_BARREL->CLASSNAMES map. */
      bow_str2int (vpc_barrel->classnames, classname);
    }

  if (doc_barrel->method->vpc_set_priors)
    {
      /* Set the prior probabilities on classes, if we're doing
	 NaiveBayes or something else that needs them.  */
      (*doc_barrel->method->vpc_set_priors) (vpc_barrel, doc_barrel);
    }
  else
    {
      /* We don't need priors, so set them to obviously bogus values,
	 so we'll notice if they accidently get used. */
      for (ci = 0; ci < num_classes; ci++)
	{
	  bow_cdoc *cdoc;
	  cdoc = bow_array_entry_at_index (vpc_barrel->cdocs, ci);
	  cdoc->prior = -1;
	}
    }

  return vpc_barrel;
}

/* Like bow_barrel_new_vpc(), but it also sets and normalizes the
   weights appropriately by calling SET_WEIGHTS from the METHOD of
   DOC_BARREL on the `vector-per-class' barrel that will be returned. */
bow_barrel *
bow_barrel_new_vpc_merge_then_weight (bow_barrel *doc_barrel)
{
  bow_barrel *vpc_barrel;

  assert (doc_barrel->method->name);
  /* Merge documents into classes, then set weights. */
  vpc_barrel = bow_barrel_new_vpc (doc_barrel);
  bow_barrel_set_weights (vpc_barrel);
  /* Scale the weights */
  bow_barrel_scale_weights (vpc_barrel, doc_barrel);
  /* Normalize the weights. */
  bow_barrel_normalize_weights (vpc_barrel);
  return vpc_barrel;
}

/* Same as above, but set the weights in the DOC_BARREL, create the
   `Vector-Per-Class' barrel, and set the weights in the VPC barrel by
   summing weights from the DOC_BARREL. */
bow_barrel *
bow_barrel_new_vpc_weight_then_merge (bow_barrel *doc_barrel)
{
  bow_barrel *vpc_barrel;

  /* Set weights, then merge documents into classes. */
  bow_barrel_set_weights (doc_barrel);
  vpc_barrel = bow_barrel_new_vpc (doc_barrel);
  bow_barrel_scale_weights (vpc_barrel, doc_barrel);
  bow_barrel_normalize_weights (vpc_barrel);
  return vpc_barrel;
}

/* Set the class prior probabilities by counting the number of
   documents of each class. */
void
bow_barrel_set_vpc_priors_by_counting (bow_barrel *vpc_barrel,
				       bow_barrel *doc_barrel)
					
{
  double prior_sum = 0;
  int ci;
  int max_ci = vpc_barrel->cdocs->length - 1;
  int di;

  /* Zero them. */
  for (ci = 0; ci <= max_ci; ci++)
    {
      bow_cdoc *cdoc;
      cdoc = bow_array_entry_at_index (vpc_barrel->cdocs, ci);
      cdoc->prior = 0;
    }
  /* Add in document counts. */
  for (di = 0; di < doc_barrel->cdocs->length; di++)
    {
      bow_cdoc *doc_cdoc;
      bow_cdoc *vpc_cdoc;
      doc_cdoc = bow_array_entry_at_index (doc_barrel->cdocs, di);
      if (doc_cdoc->type != bow_doc_train)
	continue;
      if (doc_cdoc->class >= vpc_barrel->cdocs->length)
	{
	  /* This can happen if all of the documents in a certain class
	     contain only words that are not in the vocabulary used
	     when running bow_barrel_new_vpc() above. */
	  bow_error ("Number of classes in class barrel do not match\n"
		     "number of classes in document barrel!");
	}
      vpc_cdoc = bow_array_entry_at_index (vpc_barrel->cdocs, 
					   doc_cdoc->class);
      vpc_cdoc->prior += doc_cdoc->prior;
    }
  /* Sum them all. */
  for (ci = 0; ci <= max_ci; ci++)
    {
      bow_cdoc *cdoc;
      cdoc = bow_array_entry_at_index (vpc_barrel->cdocs, ci);
      prior_sum += cdoc->prior;
    }
  if (prior_sum)
    {
      /* Normalize to set the prior. */
      for (ci = 0; ci <= max_ci; ci++)
	{
	  bow_cdoc *cdoc;
	  cdoc = bow_array_entry_at_index (vpc_barrel->cdocs, ci);
	  cdoc->prior /= prior_sum;
	  if (cdoc->prior == 0)
	    bow_verbosify (bow_progress, 
			   "WARNING: class `%s' has zero prior\n",
			   cdoc->filename);
	  /* printf ("ci=%d  prior_sum=%f  prior=%f\n", ci,prior_sum,
	     cdoc->prior);*/
	  /* xxx We allow "cdoc->prior >= 0.0" because there may be no
	     training data for some class.  Is this good? */
	  assert (cdoc->prior >= 0.0 && cdoc->prior <= 1.0);
	}
    }
  else
    {
      bow_verbosify (bow_progress, "WARNING: All classes have zero prior\n");
    }
}

/* Like bow_barrel_new_vpc, but uses both labeled and unlabeled data.
   It uses the class_probs of each doc to determine its class
   membership. The counts in the wi2dvf are set to bogus numbers.  The
   weights of the wi2dvf contain the real information. The normalizer
   of each vpc cdoc is set to the fractional number of documents per
   class.  The word_count of each vpc cdoc is rounded integer for the
   number of documents per class.  The word_count of each document
   cdoc is set to the sum of the counts of its corresponding word
   vector.  This is to get correct numbers for the doc-then-word event
   model.  */
bow_barrel *
bow_barrel_new_vpc_using_class_probs (bow_barrel *doc_barrel)
{
  bow_barrel* vpc_barrel;	/* The vector per class barrel */
  int num_classes = bow_barrel_num_classes (doc_barrel);
  int wi;
  int max_wi;
  int dvi;
  int ci;
  bow_dv *dv;
  bow_dv *vpc_dv;
  int di;
  float num_docs_per_ci[num_classes];
  bow_cdoc *cdoc;

  assert (doc_barrel->classnames);

  max_wi = MIN (doc_barrel->wi2dvf->size, bow_num_words ());

  /* Create an empty barrel; we fill it with vector-per-class
     data and return it. */
  /* This assertion can fail when DOC_BARREL was read from a disk
     archive that was created before CLASS_PROBS was added to BOW_CDOC */
  assert (doc_barrel->cdocs->entry_size >= sizeof (bow_cdoc));
  vpc_barrel = bow_barrel_new (doc_barrel->wi2dvf->size,
			       num_classes,
			       doc_barrel->cdocs->entry_size,
			       doc_barrel->cdocs->free_func);
  vpc_barrel->method = doc_barrel->method;
  vpc_barrel->classnames = bow_int4str_new (0);
  /* Make sure to set the VPC indicator */
  vpc_barrel->is_vpc = 1;

  bow_verbosify (bow_verbose, "Making vector-per-class... words ::       ");

  /* Count the number of documents in each class using the class probs */
  for (ci = 0; ci < num_classes; ci++)
    num_docs_per_ci[ci] = 0.0;
  for (di = 0; di < doc_barrel->cdocs->length; di++)
    {
      cdoc = bow_array_entry_at_index (doc_barrel->cdocs, di);
      if (cdoc->type == bow_doc_train ||
	  cdoc->type == bow_doc_unlabeled) {
	for (ci = 0; ci < num_classes; ci++) 
	  num_docs_per_ci[ci] += cdoc->class_probs[ci];
      }
    }

  /* Update the CDOC->WORD_COUNT in the DOC_BARREL in order to match
     the (potentially) pruned vocabulary. */
  {
    bow_wv *wv = NULL;
    int wvi;
    bow_dv_heap *heap = bow_test_new_heap (doc_barrel);
    while ((di = bow_heap_next_wv (heap, doc_barrel, &wv,
				   bow_cdoc_yes)) != -1)
      {
	cdoc = bow_array_entry_at_index (doc_barrel->cdocs, di);
	cdoc->word_count = 0;
	for (wvi = 0; wvi < wv->num_entries; wvi++)
	  {
	    if (bow_wi2dvf_dv (doc_barrel->wi2dvf, wv->entry[wvi].wi))
	      cdoc->word_count += wv->entry[wvi].count;
	  }
      }
  }

  /* Initialize the WI2DVF part of the VPC_BARREL.  Sum together the
     counts and weights for individual documents, grabbing only the
     training documents. */
  for (wi = 0; wi < max_wi; wi++)
    {
      dv = bow_wi2dvf_dv (doc_barrel->wi2dvf, wi);
      if (!dv)
	continue;
      for (dvi = 0; dvi < dv->length; dvi++)
	{
	  di = dv->entry[dvi].di;
	  cdoc = bow_array_entry_at_index (doc_barrel->cdocs, di);
	  if (cdoc->type == bow_doc_train ||
	      cdoc->type == bow_doc_unlabeled)
	    {
	      float weight;
	      
	      /* The old version of bow_wi2dvf_add_di_text_fp() initialized
		 the dv WEIGHT to 0 instead of the word count.  If the weight 
		 is zero, then use the count instead.  Note, however, that
		 the TFIDF method might have set the weight, so we don't
		 want to use the count all the time. */
	      if (dv->entry[dvi].weight)
		weight = dv->entry[dvi].weight;
	      else
		weight = dv->entry[dvi].count;

	      for (ci = 0; ci < num_classes; ci++) 
		{

		  /* do the right thing based on the event model */
		  if (bow_event_model == bow_event_document)
		    {
		      assert (dv->entry[dvi].count);
		      bow_wi2dvf_add_wi_di_count_weight (&(vpc_barrel->wi2dvf), 
							 wi, ci, 1, 
							 cdoc->class_probs[ci]);
		    }
		  else if (bow_event_model == bow_event_document_then_word)
		    {
		      bow_wi2dvf_add_wi_di_count_weight
			(&(vpc_barrel->wi2dvf), wi, ci, 1,
			 (bow_event_document_then_word_document_length
			  * weight * cdoc->class_probs[ci] / cdoc->word_count));
		    }
		  else
		    {
		      bow_wi2dvf_add_wi_di_count_weight (&(vpc_barrel->wi2dvf), 
							 wi, ci, 
							 1,
							 weight * cdoc->class_probs[ci]);
		    }
		}
	    }
	}
      /* Set the IDF of the class's wi2dvf directly from the doc's wi2dvf */
      vpc_dv = bow_wi2dvf_dv (vpc_barrel->wi2dvf, wi);
      if (vpc_dv) 
	vpc_dv->idf = dv->idf;
      if (max_wi - wi % 100 == 0)
	bow_verbosify (bow_verbose, "\b\b\b\b\b\b%6d", max_wi - wi);
    }

  bow_verbosify (bow_verbose, "\b\b\b\b\b\b\n");

  /* Initialize the CDOCS and CLASSNAMES parts of the VPC_BARREL.
     Create BOW_CDOC structures for each class, and append them to the
     VPC->CDOCS array. */
  for (ci = 0; ci < num_classes; ci++)
    {
      bow_cdoc cdoc;
      const char *classname = NULL;

      cdoc.type = bow_doc_train;
      cdoc.normalizer = num_docs_per_ci[ci];
      /* Make WORD_COUNT be the number of documents in the class.
         This is for the document event model.*/
      cdoc.word_count = rint (num_docs_per_ci[ci]);
      if (doc_barrel->classnames)
	{
	  classname = bow_barrel_classname_at_index (doc_barrel, ci);
	  cdoc.filename = strdup (classname);
	  if (!cdoc.filename)
	    bow_error ("Memory exhausted.");
	}
      else
	{
	  cdoc.filename = NULL;
	}
      cdoc.class_probs = NULL;
      cdoc.class = ci;
      bow_verbosify (bow_verbose, "%20f model documents in class `%s'\n",
		     num_docs_per_ci[ci], cdoc.filename);
      /* Add a CDOC for this class to the VPC_BARREL */
      bow_array_append (vpc_barrel->cdocs, &cdoc);
      /* Add an entry for this class into the VPC_BARREL->CLASSNAMES map. */
      bow_str2int (vpc_barrel->classnames, classname);
    }

  if (doc_barrel->method->vpc_set_priors)
    {
      /* Set the prior probabilities on classes, if we're doing
	 NaiveBayes or something else that needs them.  */
      (*doc_barrel->method->vpc_set_priors) (vpc_barrel, doc_barrel);
    }
  else
    {
      /* We don't need priors, so set them to obviously bogus values,
	 so we'll notice if they accidently get used. */
      for (ci = 0; ci < num_classes; ci++)
	{
	  bow_cdoc *cdoc;
	  cdoc = bow_array_entry_at_index (vpc_barrel->cdocs, ci);
	  cdoc->prior = -1;
	}
    }

  return vpc_barrel;
}

/* Set the class prior probabilities by doing a weighted (by class
   membership) count of the number of labeled and unlabeled documents
   in each class.  This uses class_probs to determine class
   memberships of the documents. */
void
bow_barrel_set_vpc_priors_using_class_probs (bow_barrel *vpc_barrel,
					     bow_barrel *doc_barrel)
     
{
  float prior_sum = 0;
  int ci;
  int max_ci = vpc_barrel->cdocs->length;
  int di;

  /* Zero them. */
  for (ci = 0; ci < max_ci; ci++)
    {
      bow_cdoc *cdoc;
      cdoc = bow_array_entry_at_index (vpc_barrel->cdocs, ci);
      cdoc->prior = 0;
    }

  /* Count each document for each class according to the
     class_probs. */
  for (di = 0; di < doc_barrel->cdocs->length; di++)
    {
      bow_cdoc *doc_cdoc = bow_array_entry_at_index (doc_barrel->cdocs, di);
      bow_cdoc *vpc_cdoc;
      
      if (doc_cdoc->type == bow_doc_train ||
	  doc_cdoc->type == bow_doc_unlabeled)
	{
	  for (ci = 0; ci < max_ci; ci++)
	    {
	      vpc_cdoc = bow_array_entry_at_index (vpc_barrel->cdocs, ci); 
	      vpc_cdoc->prior += doc_cdoc->class_probs[ci];
	    }
	}
    }
  
  /* Sum them all. */
  for (ci = 0; ci < max_ci; ci++)
    {
      bow_cdoc *cdoc;
      cdoc = bow_array_entry_at_index (vpc_barrel->cdocs, ci);
      prior_sum += cdoc->prior;
    }

  /* Normalize to set the prior. */
  for (ci = 0; ci < max_ci; ci++)
    {
      bow_cdoc *cdoc;
      cdoc = bow_array_entry_at_index (vpc_barrel->cdocs, ci);
      cdoc->prior /= prior_sum;
      if (cdoc->prior == 0)
	bow_verbosify (bow_progress, 
		       "WARNING: class `%s' has zero prior\n",
		       cdoc->filename);
      assert (cdoc->prior >= 0.0 && cdoc->prior <= 1.0);
    }
}