/* Copyright (C) 2004 J.F.Dockes
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program 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 General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the
* Free Software Foundation, Inc.,
* 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
*/
#include "autoconfig.h"
#include <math.h>
#include <map>
#include <tr1/unordered_set>
using std::tr1::unordered_set;
using namespace std;
#include "debuglog.h"
#include "rcldb.h"
#include "rcldb_p.h"
#include "rclquery.h"
#include "rclquery_p.h"
#include "textsplit.h"
#include "searchdata.h"
#include "utf8iter.h"
#include "hldata.h"
namespace Rcl {
// This is used as a marker inside the abstract frag lists, but
// normally doesn't remain in final output (which is built with a
// custom sep. by our caller).
static const string cstr_ellipsis("...");
// This is used to mark positions overlapped by a multi-word match term
static const string occupiedmarker("?");
#undef DEBUGABSTRACT
#ifdef DEBUGABSTRACT
#define LOGABS LOGDEB
static void listList(const string& what, const vector<string>&l)
{
string a;
for (vector<string>::const_iterator it = l.begin(); it != l.end(); it++) {
a = a + *it + " ";
}
LOGDEB(("%s: %s\n", what.c_str(), a.c_str()));
}
#else
#define LOGABS LOGDEB2
static void listList(const string&, const vector<string>&)
{
}
#endif
// Unprefix terms.
static void noPrefixList(const vector<string>& in, vector<string>& out)
{
for (vector<string>::const_iterator qit = in.begin();
qit != in.end(); qit++) {
out.push_back(strip_prefix(*qit));
}
sort(out.begin(), out.end());
vector<string>::iterator it = unique(out.begin(), out.end());
out.resize(it - out.begin());
}
// Retrieve db-wide frequencies for the query terms and store them in
// the query object. This is done at most once for a query, and the data is used
// while computing abstracts for the different result documents.
void Query::Native::setDbWideQTermsFreqs()
{
// Do it once only for a given query.
if (!termfreqs.empty())
return;
vector<string> qterms;
{
vector<string> iqterms;
m_q->getQueryTerms(iqterms);
noPrefixList(iqterms, qterms);
}
// listList("Query terms: ", qterms);
Xapian::Database &xrdb = m_q->m_db->m_ndb->xrdb;
double doccnt = xrdb.get_doccount();
if (doccnt == 0)
doccnt = 1;
for (vector<string>::const_iterator qit = qterms.begin();
qit != qterms.end(); qit++) {
termfreqs[*qit] = xrdb.get_termfreq(*qit) / doccnt;
LOGABS(("setDbWideQTermFreqs: [%s] db freq %.1e\n", qit->c_str(),
termfreqs[*qit]));
}
}
// Compute matched terms quality coefficients for a matched document by
// retrieving the Within Document Frequencies and multiplying by
// overal term frequency, then using log-based thresholds.
// 2012: it's not too clear to me why exactly we do the log thresholds thing.
// Preferring terms wich are rare either or both in the db and the document
// seems reasonable though
// To avoid setting a high quality for a low frequency expansion of a
// common stem, which seems wrong, we group the terms by
// root, compute a frequency for the group from the sum of member
// occurrences, and let the frequency for each group member be the
// aggregated frequency.
double Query::Native::qualityTerms(Xapian::docid docid,
const vector<string>& terms,
multimap<double, vector<string> >& byQ)
{
LOGABS(("qualityTerms\n"));
setDbWideQTermsFreqs();
map<string, double> termQcoefs;
double totalweight = 0;
Xapian::Database &xrdb = m_q->m_db->m_ndb->xrdb;
double doclen = xrdb.get_doclength(docid);
if (doclen == 0)
doclen = 1;
HighlightData hld;
if (!m_q->m_sd.isNull()) {
m_q->m_sd->getTerms(hld);
}
#ifdef DEBUGABSTRACT
{
string deb;
hld.toString(deb);
LOGABS(("qualityTerms: hld: %s\n", deb.c_str()));
}
#endif
// Group the input terms by the user term they were possibly expanded from
map<string, vector<string> > byRoot;
for (vector<string>::const_iterator qit = terms.begin();
qit != terms.end(); qit++) {
bool found = false;
map<string, string>::const_iterator eit = hld.terms.find(*qit);
if (eit != hld.terms.end()) {
byRoot[eit->second].push_back(*qit);
} else {
LOGDEB0(("qualityTerms: [%s] not found in hld\n", (*qit).c_str()));
byRoot[*qit].push_back(*qit);
}
}
#ifdef DEBUGABSTRACT
{
string byRootstr;
for (map<string, vector<string> >::const_iterator debit =
byRoot.begin(); debit != byRoot.end(); debit++) {
byRootstr.append("[").append(debit->first).append("]->");
for (vector<string>::const_iterator it = debit->second.begin();
it != debit->second.end(); it++) {
byRootstr.append("[").append(*it).append("] ");
}
byRootstr.append("\n");
}
LOGABS(("\nqualityTerms: uterms to terms: %s\n", byRootstr.c_str()));
}
#endif
// Compute in-document and global frequencies for the groups.
map<string, double> grpwdfs;
map<string, double> grptfreqs;
for (map<string, vector<string> >::const_iterator git = byRoot.begin();
git != byRoot.end(); git++) {
for (vector<string>::const_iterator qit = git->second.begin();
qit != git->second.end(); qit++) {
Xapian::TermIterator term = xrdb.termlist_begin(docid);
term.skip_to(*qit);
if (term != xrdb.termlist_end(docid) && *term == *qit) {
if (grpwdfs.find(git->first) != grpwdfs.end()) {
grpwdfs[git->first] = term.get_wdf() / doclen;
grptfreqs[git->first] = termfreqs[*qit];
} else {
grpwdfs[git->first] += term.get_wdf() / doclen;
grptfreqs[git->first] += termfreqs[*qit];
}
}
}
}
// Build a sorted by quality container for the groups
for (map<string, vector<string> >::const_iterator git = byRoot.begin();
git != byRoot.end(); git++) {
double q = (grpwdfs[git->first]) * grptfreqs[git->first];
q = -log10(q);
if (q < 3) {
q = 0.05;
} else if (q < 4) {
q = 0.3;
} else if (q < 5) {
q = 0.7;
} else if (q < 6) {
q = 0.8;
} else {
q = 1;
}
totalweight += q;
byQ.insert(pair<double, vector<string> >(q, git->second));
}
#ifdef DEBUGABSTRACT
for (multimap<double, vector<string> >::reverse_iterator mit= byQ.rbegin();
mit != byQ.rend(); mit++) {
LOGABS(("qualityTerms: group\n"));
for (vector<string>::const_iterator qit = mit->second.begin();
qit != mit->second.end(); qit++) {
LOGABS(("%.1e->[%s]\n", mit->first, qit->c_str()));
}
}
#endif
return totalweight;
}
// Return page number for first match of "significant" term.
int Query::Native::getFirstMatchPage(Xapian::docid docid, string& term)
{
if (!m_q|| !m_q->m_db || !m_q->m_db->m_ndb || !m_q->m_db->m_ndb->m_isopen) {
LOGERR(("Query::getFirstMatchPage: no db\n"));
return -1;
}
Rcl::Db::Native *ndb(m_q->m_db->m_ndb);
Xapian::Database& xrdb(ndb->xrdb);
vector<string> terms;
{
vector<string> iterms;
m_q->getMatchTerms(docid, iterms);
noPrefixList(iterms, terms);
}
if (terms.empty()) {
LOGDEB(("getFirstMatchPage: empty match term list (field match?)\n"));
return -1;
}
vector<int> pagepos;
ndb->getPagePositions(docid, pagepos);
if (pagepos.empty())
return -1;
setDbWideQTermsFreqs();
// We try to use a page which matches the "best" term. Get a sorted list
multimap<double, vector<string> > byQ;
double totalweight = qualityTerms(docid, terms, byQ);
for (map<double, vector<string> >::reverse_iterator mit = byQ.rbegin();
mit != byQ.rend(); mit++) {
for (vector<string>::const_iterator qit = mit->second.begin();
qit != mit->second.end(); qit++) {
string qterm = *qit;
Xapian::PositionIterator pos;
string emptys;
try {
for (pos = xrdb.positionlist_begin(docid, qterm);
pos != xrdb.positionlist_end(docid, qterm); pos++) {
int pagenum = ndb->getPageNumberForPosition(pagepos, *pos);
if (pagenum > 0) {
term = qterm;
return pagenum;
}
}
} catch (...) {
// Term does not occur. No problem.
}
}
}
return -1;
}
// Build a document abstract by extracting text chunks around the query terms
// This uses the db termlists, not the original document.
//
// DatabaseModified and other general exceptions are catched and
// possibly retried by our caller
abstract_result Query::Native::makeAbstract(Xapian::docid docid,
vector<Snippet>& vabs,
int imaxoccs, int ictxwords)
{
Chrono chron;
LOGABS(("makeAbstract: docid %l imaxoccs %d ictxwords %d\n",
long(docid), imaxoccs, ictxwords));
// The (unprefixed) terms matched by this document
vector<string> matchedTerms;
{
vector<string> iterms;
m_q->getMatchTerms(docid, iterms);
noPrefixList(iterms, matchedTerms);
if (matchedTerms.empty()) {
LOGDEB(("makeAbstract::Empty term list\n"));
return ABSRES_ERROR;
}
}
listList("Match terms: ", matchedTerms);
// Retrieve the term freqencies for the query terms. This is
// actually computed only once for a query, and for all terms in
// the query (not only the matches for this doc)
setDbWideQTermsFreqs();
// Build a sorted by quality container for the match terms We are
// going to try and show text around the less common search terms.
// TOBEDONE: terms issued from an original one by stem expansion
// should be somehow aggregated here, else, it may happen that
// such a group prevents displaying matches for other terms (by
// removing its meaning from the maximum occurrences per term test
// used while walking the list below)
multimap<double, vector<string> > byQ;
double totalweight = qualityTerms(docid, matchedTerms, byQ);
LOGABS(("makeAbstract:%d: computed Qcoefs.\n", chron.ms()));
// This can't happen, but would crash us
if (totalweight == 0.0) {
LOGERR(("makeAbstract: totalweight == 0.0 !\n"));
return ABSRES_ERROR;
}
Rcl::Db::Native *ndb(m_q->m_db->m_ndb);
Xapian::Database& xrdb(ndb->xrdb);
///////////////////
// For each of the query terms, ask xapian for its positions list
// in the document. For each position entry, insert it and its
// neighbours in the set of 'interesting' positions
// The terms 'array' that we partially populate with the document
// terms, at their positions around the search terms positions:
map<unsigned int, string> sparseDoc;
// Also remember apart the search term positions so that we can list
// them with their snippets.
unordered_set<unsigned int> searchTermPositions;
// Remember max position. Used to stop walking positions lists while
// populating the adjacent slots.
unsigned int maxpos = 0;
// Total number of occurences for all terms. We stop when we have too much
unsigned int totaloccs = 0;
// Total number of slots we populate. The 7 is taken as
// average word size. It was a mistake to have the user max
// abstract size parameter in characters, we basically only deal
// with words. We used to limit the character size at the end, but
// this damaged our careful selection of terms
const unsigned int maxtotaloccs = imaxoccs > 0 ? imaxoccs :
m_q->m_db->getAbsLen() /(7 * (m_q->m_db->getAbsCtxLen() + 1));
int ctxwords = ictxwords == -1 ? m_q->m_db->getAbsCtxLen() : ictxwords;
LOGABS(("makeAbstract:%d: mxttloccs %d ctxwords %d\n",
chron.ms(), maxtotaloccs, ctxwords));
abstract_result ret = ABSRES_OK;
// Let's go populate
for (map<double, vector<string> >::reverse_iterator mit = byQ.rbegin();
mit != byQ.rend(); mit++) {
unsigned int maxgrpoccs;
float q;
if (byQ.size() == 1) {
maxgrpoccs = maxtotaloccs;
q = 1.0;
} else {
// We give more slots to the better term groups
q = mit->first / totalweight;
maxgrpoccs = int(ceil(maxtotaloccs * q));
}
unsigned int grpoccs = 0;
for (vector<string>::const_iterator qit = mit->second.begin();
qit != mit->second.end(); qit++) {
// Group done ?
if (grpoccs >= maxgrpoccs)
break;
string qterm = *qit;
LOGABS(("makeAbstract: [%s] %d max grp occs (coef %.2f)\n",
qterm.c_str(), maxgrpoccs, q));
// The match term may span several words
int qtrmwrdcnt =
TextSplit::countWords(qterm, TextSplit::TXTS_NOSPANS);
Xapian::PositionIterator pos;
// There may be query terms not in this doc. This raises an
// exception when requesting the position list, we catch it ??
// Not clear how this can happen because we are walking the
// match list returned by Xapian. Maybe something with the
// fields?
string emptys;
try {
for (pos = xrdb.positionlist_begin(docid, qterm);
pos != xrdb.positionlist_end(docid, qterm); pos++) {
int ipos = *pos;
if (ipos < int(baseTextPosition)) // Not in text body
continue;
LOGABS(("makeAbstract: [%s] at pos %d grpoccs %d maxgrpoccs %d\n",
qterm.c_str(), ipos, grpoccs, maxgrpoccs));
totaloccs++;
grpoccs++;
// Add adjacent slots to the set to populate at next
// step by inserting empty strings. Special provisions
// for adding ellipsis and for positions overlapped by
// the match term.
unsigned int sta = MAX(0, ipos - ctxwords);
unsigned int sto = ipos + qtrmwrdcnt-1 +
m_q->m_db->getAbsCtxLen();
for (unsigned int ii = sta; ii <= sto; ii++) {
if (ii == (unsigned int)ipos) {
sparseDoc[ii] = qterm;
searchTermPositions.insert(ii);
if (ii > maxpos)
maxpos = ii;
} else if (ii > (unsigned int)ipos &&
ii < (unsigned int)ipos + qtrmwrdcnt) {
sparseDoc[ii] = occupiedmarker;
} else if (!sparseDoc[ii].compare(cstr_ellipsis)) {
// For an empty slot, the test has a side
// effect of inserting an empty string which
// is what we want
sparseDoc[ii] = emptys;
}
}
// Add ellipsis at the end. This may be replaced later by
// an overlapping extract. Take care not to replace an
// empty string here, we really want an empty slot,
// use find()
if (sparseDoc.find(sto+1) == sparseDoc.end()) {
sparseDoc[sto+1] = cstr_ellipsis;
}
// Group done ?
if (grpoccs >= maxgrpoccs)
break;
// Global done ?
if (totaloccs >= maxtotaloccs) {
ret = ABSRES_TRUNC;
LOGABS(("Db::makeAbstract: max occurrences cutoff\n"));
break;
}
}
} catch (...) {
// Term does not occur. No problem.
}
if (totaloccs >= maxtotaloccs) {
ret = ABSRES_TRUNC;
LOGABS(("Db::makeAbstract: max1 occurrences cutoff\n"));
break;
}
}
}
maxpos += ctxwords + 1;
LOGABS(("makeAbstract:%d:chosen number of positions %d\n",
chron.millis(), totaloccs));
// This can happen if there are term occurences in the keywords
// etc. but not elsewhere ?
if (totaloccs == 0) {
LOGDEB(("makeAbstract: no occurrences\n"));
return ABSRES_OK;
}
// Walk all document's terms position lists and populate slots
// around the query terms. We arbitrarily truncate the list to
// avoid taking forever. If we do cutoff, the abstract may be
// inconsistant (missing words, potentially altering meaning),
// which is bad.
{
Xapian::TermIterator term;
int cutoff = m_q->m_snipMaxPosWalk;
for (term = xrdb.termlist_begin(docid);
term != xrdb.termlist_end(docid); term++) {
// Ignore prefixed terms
if (has_prefix(*term))
continue;
if (m_q->m_snipMaxPosWalk > 0 && cutoff-- < 0) {
ret = ABSRES_TERMMISS;
LOGDEB0(("makeAbstract: max term count cutoff %d\n",
m_q->m_snipMaxPosWalk));
break;
}
map<unsigned int, string>::iterator vit;
Xapian::PositionIterator pos;
for (pos = xrdb.positionlist_begin(docid, *term);
pos != xrdb.positionlist_end(docid, *term); pos++) {
if (m_q->m_snipMaxPosWalk > 0 && cutoff-- < 0) {
ret = ABSRES_TERMMISS;
LOGDEB0(("makeAbstract: max term count cutoff %d\n",
m_q->m_snipMaxPosWalk));
break;
}
// If we are beyond the max possible position, stop
// for this term
if (*pos > maxpos) {
break;
}
if ((vit = sparseDoc.find(*pos)) != sparseDoc.end()) {
// Don't replace a term: the terms list is in
// alphabetic order, and we may have several terms
// at the same position, we want to keep only the
// first one (ie: dockes and dockes@wanadoo.fr)
if (vit->second.empty()) {
LOGDEB2(("makeAbstract: populating: [%s] at %d\n",
(*term).c_str(), *pos));
sparseDoc[*pos] = *term;
}
}
}
}
}
#if 0
// Debug only: output the full term[position] vector
bool epty = false;
int ipos = 0;
for (map<unsigned int, string>::iterator it = sparseDoc.begin();
it != sparseDoc.end();
it++, ipos++) {
if (it->empty()) {
if (!epty)
LOGDEB(("makeAbstract:vec[%d]: [%s]\n", ipos, it->c_str()));
epty=true;
} else {
epty = false;
LOGDEB(("makeAbstract:vec[%d]: [%s]\n", ipos, it->c_str()));
}
}
#endif
vector<int> vpbreaks;
ndb->getPagePositions(docid, vpbreaks);
LOGABS(("makeAbstract:%d: extracting. Got %u pages\n", chron.millis(),
vpbreaks.size()));
// Finally build the abstract by walking the map (in order of position)
vabs.clear();
string chunk;
bool incjk = false;
int page = 0;
string term;
for (map<unsigned int, string>::const_iterator it = sparseDoc.begin();
it != sparseDoc.end(); it++) {
LOGDEB2(("Abtract:output %u -> [%s]\n", it->first,it->second.c_str()));
if (!occupiedmarker.compare(it->second))
continue;
if (chunk.empty() && !vpbreaks.empty()) {
page = ndb->getPageNumberForPosition(vpbreaks, it->first);
if (page < 0)
page = 0;
term.clear();
}
Utf8Iter uit(it->second);
bool newcjk = false;
if (TextSplit::isCJK(*uit))
newcjk = true;
if (!incjk || (incjk && !newcjk))
chunk += " ";
incjk = newcjk;
if (searchTermPositions.find(it->first) != searchTermPositions.end())
term = it->second;
if (it->second == cstr_ellipsis) {
vabs.push_back(Snippet(page, chunk).setTerm(term));
chunk.clear();
} else {
if (it->second.compare(end_of_field_term) &&
it->second.compare(start_of_field_term))
chunk += it->second;
}
}
if (!chunk.empty())
vabs.push_back(Snippet(page, chunk).setTerm(term));
LOGDEB2(("makeAbtract: done in %d mS\n", chron.millis()));
return ret;
}
}