// langId is language of the query
long long getSynBaseHash64 ( char *qstr , uint8_t langId ) {
Words ww;
ww.set3 ( qstr );
long nw = ww.getNumWords();
long long *wids = ww.getWordIds();
//char **wptrs = ww.getWords();
//long *wlens = ww.getWordLens();
long long baseHash64 = 0LL;
Synonyms syn;
// assume english if unknown to fix 'pandora's tower'
// vs 'pandoras tower' where both words are in both
// english and german so langid is unknown
if ( langId == langUnknown ) langId = langEnglish;
// . store re-written query into here then hash that string
// . this way we can get rid of spaces
//char rebuf[1024];
//char *p = rebuf;
//if ( strstr(qstr,"cheatcodes") )
// log("hey");
// for deduping
HashTableX dups;
if ( ! dups.set ( 8,0,1024,NULL,0,false,0,"qhddup") ) return false;
// scan the words
for ( long i = 0 ; i < nw ; i++ ) {
// skip if not alnum
if ( ! wids[i] ) continue;
// get its synonyms into tmpBuf
char tmpBuf[TMPSYNBUFSIZE];
// . assume niceness of 0 for now
// . make sure to get all synsets!! ('love' has two synsets)
long naids = syn.getSynonyms (&ww,i,langId,tmpBuf,0);
// term freq algo
//long pop = g_speller.getPhrasePopularity(NULL,
// wids[i],
// true,
// langId);
// is it a queryStopWord like "the" or "and"?
bool isQueryStop = ::isQueryStopWord(NULL,0,wids[i]);
// a more restrictive list
bool isStop = ::isStopWord(NULL,0,wids[i]);
if ( ::isCommonQueryWordInEnglish(wids[i]) ) isStop = true;
// find the smallest one
unsigned long long min = wids[i];
//char *minWordPtr = wptrs[i];
//long minWordLen = wlens[i];
// declare up here since we have a goto below
long j;
// add to table too
if ( dups.isInTable ( &min ) ) goto gotdup;
// add to it
if ( ! dups.addKey ( &min ) ) return false;
// now scan the synonyms, they do not include "min" in them
for ( j = 0 ; j < naids ; j++ ) {
// get it
unsigned long long aid64;
aid64 = (unsigned long long)syn.m_aids[j];
// if any syn already hashed then skip it and count
// as a repeated term. we have to do it this way
// rather than just getting the minimum synonym
// word id, because 'love' has two synsets and
// 'like', a synonym of 'love' only has one synset
// and they end up having different minimum synonym
// word ids!!!
if ( dups.isInTable ( &aid64 ) ) break;
// add it. this could fail!
if ( ! dups.addKey ( &aid64 ) ) return false;
// set it?
if ( aid64 >= min ) continue;
// got a new min
min = aid64;
//minWordPtr = syn.m_termPtrs[j];
//minWordLen = syn.m_termLens[j];
// get largest term freq of all synonyms
//long pop2 = g_speller.getPhrasePopularity(NULL,aid64,
// true,langId);
//if ( pop2 > pop ) pop = pop2;
}
// early break out means a hit in dups table
if ( j < naids ) {
gotdup:
// do not count as repeat if query stop word
// because they often repeat
if ( isQueryStop ) continue;
// count # of repeated word forms
//nrwf++;
continue;
}
// hash that now
// do not include stop words in synbasehash so
// 'search the web' != 'search web'
if ( ! isStop ) {
// no! make it order independent so 'search the web'
// equals 'web the search' and 'engine search'
// equals 'search engine'
//baseHash64 <<= 1LL;
baseHash64 ^= min;
}
// count it, but only if not a query stop word like "and"
// or "the" or "a". # of unique word forms.
//if ( ! isQueryStop ) nuwf++;
// get term freq
//if ( pop > maxPop ) maxPop = pop;
// control word?
//if ( wids[i] == cw1 ) ncwf++;
}
return baseHash64;
}
bool sendTurkPageReply ( State60 *st ) {
XmlDoc *xd = &st->m_xd;
//char *content = xd->ptr_utf8Content;
//int32_t contentLen = xd->size_utf8Content - 1;
// count the total number of EventDesc classes for all evids
//char *evd = xd->ptr_eventData;
//EventDisplay *ed = (EventDisplay *)evd;
//char *addr = evd + (int32_t)ed->m_addr;
//char timeZoneOffset = getTimeZoneFromAddr ( addr );
// in case getSections() block come right back in
xd->setCallback ( st , xdcallback );
// . set niceness to 1 so all this processing doesn't slow queries down
// . however, g_niceness should still be zero... hmmm...
xd->m_niceness = 1;
// default to 1 niceness
st->m_niceness = 1;
// now set the sections class
Sections *ss = xd->getSections();
// now for each section with alnum text, telescope up as far as
// possible without containing anymore alnum text than what it
// contained. set SEC_CONTROL bit. such sections will have the
// 2 green/blue dots, that are used for turning on/off title/desc.
// but really the indians will only turn off sections that should
// not have a title/desc.
for ( Section *si = ss->m_rootSection ; si ; si = si->m_next ) {
// breathe
QUICKPOLL(st->m_niceness);
// skip if does not have text
if ( si->m_firstWordPos < 0 ) continue;
// otherwise, find biggest parent that contains just that text
Section *p = si->m_parent;
Section *last = si;
for ( ; p ; p = p->m_parent ) {
if ( p->m_firstWordPos != si->m_firstWordPos ) break;
if ( p->m_lastWordPos != si->m_lastWordPos ) break;
last = p;
}
// set that bit then
last->m_flags |= SEC_CONTROL;
// and speed up the loop
si = last;
}
// * now each SEC_CONTROL sections have a fence activated by a turker
// * an event title or description can not span a fence. it must be
// confined within a fence. however, it is allowed to include
// title or description from a "title section".
// * hold shift down to designate as title section when clicking it
// * show the raw text of each event changing as you fence
// sections in or out. show in a right frame.
// * show list of events on page in the top frame. can toggle them
// all individually.
// * and remove no-display from all tags so we can see everything.
// * highlight addresses, not just dates.
// * each section hash has its own unique bg color when activated
// * with a single click, completely reject an event because:
// contains bad time, address, title or desc. specify which so
// we can improve our algo.
// * when selecting an individual event, scroll to its tod...
// * remove all color from webpage that we can so our colors show up
// * remove all imgs. just src them to dev null.
// * allow for entering a custom title for an event or all events
// that are or will ever appear on the page.
// * when displaying the text of the events, use hyphens to
// delineate the section topology. strike out text as a section
// fence is activated.
// * when a section is activated is it easier to just redownload
// the whole text of the page? maybe just the text frame?
// * clicking on an individual sentence section should just remove
// that sentence. that is kinda a special content hash removal
// tag. like "Click here for video."
// * when an event id is selected i guess activate its bgcolor to
// be light blue for all sentences currently in the event that
// are not in activated sections. (make exception for designated
// title sections). so we need multiple tags for each events
// sentence div section. if sentence is split use multiple div tags
// then to keep the order. so each event sentence would have
// <div ev1=1 ev2=1 ev10=1>...</div> if it is in event ids 1,2 and
// 10. that way we can activate it when one of those event ids is
// activated.
SafeBuf sb;
// int16_tcuts
if ( ! xd->m_wordsValid ) { char *xx=NULL;*xx=0; }
Words *words = &xd->m_words;
int32_t nw = words->getNumWords();
char **wptrs = words->getWords();
int32_t *wlens = words->getWordLens();
nodeid_t *tids = words->getTagIds();
// a special array for printing </div> tags
char *endCounts = (char *)mcalloc ( nw ,"endcounts");
if ( ! endCounts ) return sendErrorReply ( st , g_errno );
//
// now loop over all the words. if word starts a section that has
// SEC_CONTROL bit set, and print out the section hash and a color
// tag to be activated if the turkey activates us.
// CAUTION: word may start multiple sections.
//
for ( int32_t i = 0 ; i < nw ; i++ ) {
// get section ptr
Section *sj = ss->m_sectionPtrs[i];
// sanity check. sj must be first section ptr that starts @ a
if ( sj && sj->m_a==i && sj->m_prev && sj->m_prev->m_a==i ) {
char *xx=NULL;*xx=0; }
// . does word #i start a section?
// . if section is control, print out the control
while ( sj && sj->m_a == i ) {
// print this section's hash
if ( sj->m_flags & SEC_CONTROL) {
// after the turkeys have made all the edits
// they need to submit the changes they made.
// how can we get that data sent back to the
// back end? we need to send back the colors
// of the sections that have been activated
// i guess. just do a loop over them.
sb.safePrintf("<div nobreak gbsecid=%"UINT32" "
"bgcolor=#%"XINT32" "
"onclick=gbtogglecolor()>",
(uint32_t)sj->m_tagHash,
(uint32_t)sj->m_tagHash);
// sanity check
if ( sj->m_b < 0 ) { char *xx=NULL;*xx=0; }
if ( sj->m_b > nw ) { char *xx=NULL;*xx=0; }
// and inc the /div count for that word
endCounts[sj->m_b-1]++;
}
// try next section too
sj = sj->m_next;
}
// if this is a tag, remove any coloring
if ( tids[i] ) {
}
// print the word, be it a tag, alnum, punct
sb.safeMemcpy ( wptrs[i] , wlens[i] );
// end a div tag?
if ( ! endCounts[i] ) continue;
// might be many so loop it
for ( int32_t j = 0 ; j < endCounts[i] ; j++ )
sb.safePrintf("</div>");
}
return false;
}
// . return the score of the highest-scoring window containing match #m
// . window is defined by the half-open interval [a,b) where a and b are
// word #'s in the Words array indicated by match #m
// . return -1 and set g_errno on error
int64_t Summary::getBestWindow ( Matches *matches, int32_t mm, int32_t *lasta,
int32_t *besta, int32_t *bestb, char *gotIt,
char *retired, int32_t maxExcerptLen ) {
// get the window around match #mm
Match *m = &matches->m_matches[mm];
// what is the word # of match #mm?
int32_t matchWordNum = m->m_wordNum;
// what Words/Pos/Bits classes is this match in?
Words *words = m->m_words;
Section **sp = NULL;
int32_t *pos = m->m_pos->m_pos;
// use "m_swbits" not "m_bits", that is what Bits::setForSummary() uses
const swbit_t *bb = m->m_bits->m_swbits;
// shortcut
if ( m->m_sections ) {
sp = m->m_sections->m_sectionPtrs;
}
int32_t nw = words->getNumWords();
int64_t *wids = words->getWordIds();
nodeid_t *tids = words->getTagIds();
// . sanity check
// . this prevents a core i've seen
if ( matchWordNum >= nw ) {
log("summary: got overflow condition for q=%s",m_q->m_orig);
// assume no best window
*besta = -1;
*bestb = -1;
*lasta = matchWordNum;
return 0;
}
// . we NULLify the section ptrs if we already used the word in another summary.
int32_t badFlags = SEC_SCRIPT|SEC_STYLE|SEC_SELECT|SEC_IN_TITLE;
if ( (bb[matchWordNum] & D_USED) || ( sp && (sp[matchWordNum]->m_flags & badFlags) ) ) {
// assume no best window
*besta = -1;
*bestb = -1;
*lasta = matchWordNum;
return 0;
}
// . "a" is the left fence post of the window (it is a word # in Words)
// . go to the left as far as we can
// . thus we decrement "a"
int32_t a = matchWordNum;
// "posa" is the character position of the END of word #a
int32_t posa = pos[a+1];
int32_t firstFrag = -1;
bool startOnQuote = false;
bool goodStart = false;
int32_t wordCount = 0;
// . decrease "a" as int32_t as we stay within maxNumCharsPerLine
// . avoid duplicating windows by using "lasta", the last "a" of the
// previous call to getBestWindow(). This can happen if our last
// central query term was close to this one.
for ( ; a > 0 && posa - pos[a-1] < maxExcerptLen && a > *lasta; a-- ) {
// . don't include any "dead zone",
// . dead zones have already been used for the summary, and
// we are getting a second/third/... excerpt here now then
// stop if its the start of a sentence, too
// stop before title word
if ( (bb[a-1] & D_USED) || (bb[a] & D_STARTS_SENTENCE) || ( bb[a-1] & D_IN_TITLE )) {
goodStart = true;
break;
}
// don't go beyond an LI, TR, P tag
if ( tids && ( tids[a-1] == TAG_LI ||
tids[a-1] == TAG_TR ||
tids[a-1] == TAG_P ||
tids[a-1] == TAG_DIV ) ) {
goodStart = true;
break;
}
// stop if its the start of a quoted sentence
if ( a+1<nw && (bb[a+1] & D_IN_QUOTES) &&
words->getWord(a)[0] == '\"' ){
startOnQuote = true;
goodStart = true;
break;
}
// find out the first instance of a fragment (comma, etc)
// watch out! because frag also means 's' in there's
if ( ( bb[a] & D_STARTS_FRAG ) && !(bb[a-1] & D_IS_STRONG_CONNECTOR) && firstFrag == -1 ) {
firstFrag = a;
}
if ( wids[a] ) {
wordCount++;
}
}
// if didn't find a good start, then start at the start of the frag
if ( !goodStart && firstFrag != -1 ) {
a = firstFrag;
}
// don't let punct or tag word start a line, unless a quote
if ( a < matchWordNum && !wids[a] && words->getWord(a)[0] != '\"' ){
while ( a < matchWordNum && !wids[a] ) a++;
// do not break right after a "strong connector", like
// apostrophe
while ( a < matchWordNum && a > 0 &&
( bb[a-1] & D_IS_STRONG_CONNECTOR ) )
a++;
// don't let punct or tag word start a line
while ( a < matchWordNum && !wids[a] ) a++;
}
// remember, b is not included in the summary, the summary is [a,b-1]
// remember to include all words in a matched phrase
int32_t b = matchWordNum + m->m_numWords ;
int32_t endQuoteWordNum = -1;
int32_t numTagsCrossed = 0;
for ( ; b <= nw; b++ ) {
if ( b == nw ) {
break;
}
if ( pos[b+1] - pos[a] >= maxExcerptLen ) {
break;
}
if ( startOnQuote && words->getWord(b)[0] == '\"' ) {
endQuoteWordNum = b;
}
// don't include any dead zone, those are already-used samples
if ( bb[b] & D_USED ) {
break;
}
// stop on a title word
if ( bb[b] & D_IN_TITLE ) {
break;
}
if ( wids[b] ) {
wordCount++;
}
// don't go beyond an LI or TR backtag
if ( tids && ( tids[b] == (BACKBIT|TAG_LI) ||
tids[b] == (BACKBIT|TAG_TR) ) ) {
numTagsCrossed++;
// try to have atleast 10 words in the summary
if ( wordCount > 10 ) {
break;
}
}
// go beyond a P or DIV backtag in case the earlier char is a
// ':'. This came from a special case for wikipedia pages
// eg. http://en.wikipedia.org/wiki/Flyover
if ( tids && ( tids[b] == (BACKBIT|TAG_P) ||
tids[b] == (BACKBIT|TAG_DIV) )) {
numTagsCrossed++;
// try to have atleast 10 words in the summary
if ( wordCount > 10 && words->getWord(b-1)[0] != ':' ) {
break;
}
}
}
// don't end on a lot of punct words
if ( b > matchWordNum && !wids[b-1]){
// remove more than one punct words. if we're ending on a quote
// keep it
while ( b > matchWordNum && !wids[b-2] && endQuoteWordNum != -1 && b > endQuoteWordNum ) {
b--;
}
// do not break right after a "strong connector", like apostrophe
while ( b > matchWordNum && (bb[b-2] & D_IS_STRONG_CONNECTOR) ) {
b--;
}
}
Match *ms = matches->m_matches;
// make m_matches.m_matches[mi] the first match in our [a,b) window
int32_t mi ;
// . the match at the center of the window is match #"mm", so that
// matches->m_matches[mm] is the Match class
// . set "mi" to it and back up "mi" as int32_t as >= a
for ( mi = mm ; mi > 0 && ms[mi-1].m_wordNum >=a ; mi-- )
;
// now get the score of this excerpt. Also mark all the represented
// query words. Mark the represented query words in the array that
// comes to us. also mark how many times the same word is repeated in
// this summary.
int64_t score = 0LL;
// is a url contained in the summary, that looks bad! punish!
bool hasUrl = false;
// the word count we did above was just an approximate. count it right
wordCount = 0;
// for debug
//char buf[5000];
//char *xp = buf;
SafeBuf xp;
// wtf?
if ( b > nw ) {
b = nw;
}
// first score from the starting match down to a, including match
for ( int32_t i = a ; i < b ; i++ ) {
// debug print out
if ( g_conf.m_logDebugSummary ) {
int32_t len = words->getWordLen(i);
char cs;
for (int32_t k=0;k<len; k+=cs ) {
const char *c = words->getWord(i)+k;
cs = getUtf8CharSize(c);
if ( is_binary_utf8 ( c ) ) {
continue;
}
xp.safeMemcpy ( c , cs );
xp.nullTerm();
}
}
// skip if in bad section, marquee, select, script, style
if ( sp && (sp[i]->m_flags & badFlags) ) {
continue;
}
// don't count just numeric words
if ( words->isNum(i) ) {
continue;
}
// check if there is a url. best way to check for '://'
if ( wids && !wids[i] ) {
const char *wrd = words->getWord(i);
int32_t wrdLen = words->getWordLen(i);
if ( wrdLen == 3 && wrd[0] == ':' && wrd[1] == '/' && wrd[2] == '/' ) {
hasUrl = true;
}
}
// skip if not wid
if ( ! wids[i] ) {
continue;
}
// just make every word 100 pts
int32_t t = 100;
// penalize it if in one of these sections
if ( bb[i] & ( D_IN_PARENS | D_IN_SUP | D_IN_LIST ) ) {
t /= 2;
}
// boost it if in bold or italics
if ( bb[i] & D_IN_BOLDORITALICS ) {
t *= 2;
}
// add the score for this word
score += t;
// print the score, "t"
if ( g_conf.m_logDebugSummary ) {
xp.safePrintf("(%" PRId32")",t);
}
// count the alpha words we got
wordCount++;
// if no matches left, skip
if ( mi >= matches->m_numMatches ) {
continue;
}
// get the match
Match *next = &ms[mi];
// skip if not a match
if ( i != next->m_wordNum ) {
continue;
}
// must be a match in this class
if ( next->m_words != words ) {
continue;
}
// advance it
mi++;
// which query word # does it match
int32_t qwn = next->m_qwordNum;
if ( qwn < 0 || qwn >= m_q->m_numWords ){g_process.shutdownAbort(true);}
// undo old score
score -= t;
// add 100000 per match
t = 100000;
// weight based on tf, goes from 0.1 to 1.0
t = (int32_t)((float)t * m_wordWeights [ qwn ]);
// if it is a query stop word, make it 10000 pts
if ( m_q->m_qwords[qwn].m_isQueryStopWord ) {
t = 0;//10000;
}
// penalize it if in one of these sections
if ( bb[i] & ( D_IN_PARENS | D_IN_SUP | D_IN_LIST ) ) {
t /= 2;
}
if ( gotIt[qwn] > 0 ) {
// have we matched it in this [a,b) already?
if ( gotIt[qwn] == 1 ) {
t /= 15;
} else {
// if we have more than 2 matches in the same window,
// it may not give a good summary. give a heavy penalty
t -= 200000;
}
} else if ( retired [qwn] > 0 ) {
// have we matched it already in a winning window?
t /= 12;
}
// add it back
score += t;
if ( g_conf.m_logDebugSummary ) {
xp.safePrintf ("[%" PRId32"]{qwn=%" PRId32",ww=%f}",t,qwn,
m_wordWeights[qwn]);
}
// inc the query word count for this window
if ( gotIt[qwn] < 100 ) {
gotIt[qwn]++;
}
}
int32_t oldScore = score;
// apply the bonus if it starts or a sentence
// only apply if the score is positive and if the wordcount is decent
if ( score > 0 && wordCount > 7 ){
// a match can give us 10k to 100k pts based on the tf weights
// so we don't want to overwhelm that too much, so let's make
// this a 20k bonus if it starts a sentence
if ( bb[a] & D_STARTS_SENTENCE ) {
score += 8000;
} else if ( bb[a] & D_STARTS_FRAG ) {
// likewise, a fragment, like after a comma
score += 4000;
}
// 1k if the match word is very close to the
// start of a sentence, lets say 3 alphawords
if ( matchWordNum - a < 7 ) {
score += 1000;
}
}
// a summary isn't really a summary if its less than 7 words.
// reduce the score, but still give it a decent score.
// minus 5M.
if ( wordCount < 7 ) {
score -= 20000;
}
// summaries that cross a lot of tags are usually bad, penalize them
if ( numTagsCrossed > 1 ) {
score -= (numTagsCrossed * 20000);
}
if ( hasUrl ) {
score -= 8000;
}
// show it
if ( g_conf.m_logDebugSummary ) {
log(LOG_DEBUG, "sum: score=%08" PRId32" prescore=%08" PRId32" a=%05" PRId32" b=%05" PRId32" %s",
(int32_t)score,oldScore,(int32_t)a,(int32_t)b,
xp.getBufStart());
}
// set lasta, besta, bestb
*lasta = a;
*besta = a;
*bestb = b;
return score;
}
// returns false and sets g_errno on error
bool Title::setTitle ( Xml *xml, Words *words, int32_t maxTitleLen, Query *query,
LinkInfo *linkInfo, Url *firstUrl, const char *filteredRootTitleBuf, int32_t filteredRootTitleBufSize,
uint8_t contentType, uint8_t langId, int32_t niceness ) {
// make Msg20.cpp faster if it is just has
// Msg20Request::m_setForLinkInfo set to true, no need to extricate a title.
if ( maxTitleLen <= 0 ) {
return true;
}
m_niceness = niceness;
m_maxTitleLen = maxTitleLen;
// if this is too big the "first line" algo can be huge!!!
// and really slow everything way down with a huge title candidate
int32_t maxTitleWords = 128;
// assume no title
reset();
int32_t NW = words->getNumWords();
//
// now get all the candidates
//
// . allow up to 100 title CANDIDATES
// . "as" is the word # of the first word in the candidate
// . "bs" is the word # of the last word IN the candidate PLUS ONE
int32_t n = 0;
int32_t as[MAX_TIT_CANDIDATES];
int32_t bs[MAX_TIT_CANDIDATES];
float scores[MAX_TIT_CANDIDATES];
Words *cptrs[MAX_TIT_CANDIDATES];
int32_t types[MAX_TIT_CANDIDATES];
int32_t parent[MAX_TIT_CANDIDATES];
// record the scoring algos effects
float baseScore [MAX_TIT_CANDIDATES];
float noCapsBoost [MAX_TIT_CANDIDATES];
float qtermsBoost [MAX_TIT_CANDIDATES];
float inCommonCandBoost[MAX_TIT_CANDIDATES];
// reset these
for ( int32_t i = 0 ; i < MAX_TIT_CANDIDATES ; i++ ) {
// assume no parent
parent[i] = -1;
}
// xml and words class for each link info, rss item
Xml tx[MAX_TIT_CANDIDATES];
Words tw[MAX_TIT_CANDIDATES];
int32_t ti = 0;
// restrict how many link texts and rss blobs we check for titles
// because title recs like www.google.com have hundreds and can
// really slow things down to like 50ms for title generation
int32_t kcount = 0;
int32_t rcount = 0;
//int64_t x = gettimeofdayInMilliseconds();
// . get every link text
// . TODO: repeat for linkInfo2, the imported link text
for ( Inlink *k = NULL; linkInfo && (k = linkInfo->getNextInlink(k)) ; ) {
// breathe
QUICKPOLL(m_niceness);
// fast skip check for link text
if ( k->size_linkText >= 3 && ++kcount >= 20 ) continue;
// fast skip check for rss item
if ( k->size_rssItem > 10 && ++rcount >= 20 ) continue;
// set Url
Url u;
u.set( k->getUrl(), k->size_urlBuf );
// is it the same host as us?
bool sh = true;
// skip if not from same host and should be
if ( firstUrl->getHostLen() != u.getHostLen() ) {
sh = false;
}
// skip if not from same host and should be
if ( strncmp( firstUrl->getHost(), u.getHost(), u.getHostLen() ) ) {
sh = false;
}
// get the link text
if ( k->size_linkText >= 3 ) {
char *p = k->getLinkText();
int32_t plen = k->size_linkText - 1;
if ( ! verifyUtf8 ( p , plen ) ) {
log("title: set4 bad link text from url=%s", k->getUrl());
continue;
}
// now the words.
if ( !tw[ti].set( k->getLinkText(), k->size_linkText - 1, true, 0 ) ) {
return false;
}
// set the bookends, it is the whole thing
cptrs [n] = &tw[ti];
as [n] = 0;
bs [n] = tw[ti].getNumWords();
// score higher if same host
if ( sh ) scores[n] = 1.05;
// do not count so high if remote!
else scores[n] = 0.80;
// set the type
if ( sh ) types [n] = TT_LINKTEXTLOCAL;
else types [n] = TT_LINKTEXTREMOTE;
// another candidate
n++;
// use xml and words
ti++;
// break out if too many already. save some for below.
if ( n + 30 >= MAX_TIT_CANDIDATES ) break;
}
// get the rss item
if ( k->size_rssItem <= 10 ) continue;
// . returns false and sets g_errno on error
// . use a 0 for niceness
if ( ! k->setXmlFromRSS ( &tx[ti] , 0 ) ) return false;
// get the word range
int32_t tslen;
bool isHtmlEnc;
char *ts = tx[ti].getRSSTitle ( &tslen , &isHtmlEnc );
// skip if not in the rss
if ( ! ts ) continue;
// skip if empty
if ( tslen <= 0 ) continue;
// now set words to that
if ( !tw[ti].set( ts, tslen, true, 0 ) ) {
return false;
}
// point to that
cptrs [n] = &tw[ti];
as [n] = 0;
bs [n] = tw[ti].getNumWords();
// increment since we are using it
ti++;
// base score for rss title
if ( sh ) scores[n] = 5.0;
// if not same host, treat like link text
else scores[n] = 2.0;
// set the type
if ( sh ) types [n] = TT_RSSITEMLOCAL;
else types [n] = TT_RSSITEMREMOTE;
// advance
n++;
// break out if too many already. save some for below.
if ( n + 30 >= MAX_TIT_CANDIDATES ) break;
}
//logf(LOG_DEBUG,"title: took1=%" PRId64,gettimeofdayInMilliseconds()-x);
//x = gettimeofdayInMilliseconds();
// . set the flags array
// . indicates what words are in title candidates already, but
// that is set below
// . up here we set words that are not allowed to be in candidates,
// like words that are in a link that is not a self link
// . alloc for it
char *flags = NULL;
char localBuf[10000];
int32_t need = words->getNumWords();
if ( need <= 10000 ) {
flags = (char *)localBuf;
} else {
flags = (char *)mmalloc(need,"TITLEflags");
}
if ( ! flags ) {
return false;
}
// clear it
memset ( flags , 0 , need );
// check tags in body
nodeid_t *tids = words->getTagIds();
// scan to set link text flags
// loop over all "words" in the html body
char inLink = false;
char selfLink = false;
for ( int32_t i = 0 ; i < NW ; i++ ) {
// breathe
QUICKPOLL(m_niceness);
// if in a link that is not self link, cannot be in a candidate
if ( inLink && ! selfLink ) {
flags[i] |= 0x02;
}
// out of a link
if ( tids[i] == (TAG_A | BACKBIT) ) {
inLink = false;
}
// if not start of <a> tag, skip it
if ( tids[i] != TAG_A ) {
continue;
}
// flag it
inLink = true;
// get the node in the xml
int32_t xn = words->getNodes()[i];
// is it a self link?
int32_t len;
char *link = xml->getString(xn,"href",&len);
// . set the url class to this
// . TODO: use the base url in the doc
Url u;
u.set( link, len, true, false );
// compare
selfLink = u.equals ( firstUrl );
// skip if not selfLink
if ( ! selfLink ) {
continue;
}
// if it is a selflink , check for an "onClick" tag in the
// anchor tag to fix that Mixx issue for:
// http://www.npr.org/templates/story/story.php?storyId=5417137
int32_t oclen;
char *oc = xml->getString(xn,"onclick",&oclen);
if ( ! oc ) {
oc = xml->getString(xn,"onClick",&oclen);
}
// assume not a self link if we see that...
if ( oc ) {
selfLink = false;
}
// if this <a href> link has a "title" attribute, use that
// instead! that thing is solid gold.
int32_t atlen;
char *atitle = xml->getString(xn,"title",&atlen);
// stop and use that, this thing is gold!
if ( ! atitle || atlen <= 0 ) {
continue;
}
// craziness? ignore it...
if ( atlen > 400 ) {
continue;
}
// if it contains permanent, permalink or share, ignore it!
if ( strncasestr ( atitle, "permalink", atlen ) ||
strncasestr ( atitle,"permanent", atlen) ||
strncasestr ( atitle,"share", atlen) ) {
continue;
}
// do not count the link text as viable
selfLink = false;
// aw, dammit
if ( ti >= MAX_TIT_CANDIDATES ) {
continue;
}
// other dammit
if ( n >= MAX_TIT_CANDIDATES ) {
break;
}
// ok, process it
if ( ! tw[ti].set ( atitle, atlen, true, 0 )) {
return false;
}
// set the bookends, it is the whole thing
cptrs [n] = &tw[ti];
as [n] = 0;
bs [n] = tw[ti].getNumWords();
scores [n] = 3.0; // not ALWAYS solid gold!
types [n] = TT_TITLEATT;
// we are using the words class
ti++;
// advance
n++;
// break out if too many already. save some for below.
if ( n + 20 >= MAX_TIT_CANDIDATES ) {
break;
}
}
//logf(LOG_DEBUG,"title: took2=%" PRId64,gettimeofdayInMilliseconds()-x);
//x = gettimeofdayInMilliseconds();
//int64_t *wids = WW->getWordIds();
// . find the last positive scoring guy
// . do not consider title candidates after "r" if "r" is non-zero
// . FIXES http://larvatusprodeo.net/2009/01/07/partisanship-politics-and-participation/
// the candidate # of the title tag
int32_t tti = -1;
// allow up to 4 tags from each type
char table[512];
// sanity check
if ( getNumXmlNodes() > 512 ) { char *xx=NULL;*xx=0; }
// clear table counts
memset ( table , 0 , 512 );
// the first word
char *wstart = NULL;
if ( NW > 0 ) {
wstart = words->getWord(0);
}
// loop over all "words" in the html body
for ( int32_t i = 0 ; i < NW ; i++ ) {
// come back up here if we encounter another "title-ish" tag
// within our first alleged "title-ish" tag
subloop:
// stop after 30k of text
if ( words->getWord(i) - wstart > 200000 ) {
break; // 1106
}
// get the tag id minus the back tag bit
nodeid_t tid = tids[i] & BACKBITCOMP;
// pen up and pen down for these comment like tags
if ( tid == TAG_SCRIPT || tid == TAG_STYLE ) {
// ignore "titles" in script or style tags
if ( ! (tids[i] & BACKBIT) ) {
continue;
}
}
/// @todo ALC we should allow more tags than just title/link
// skip if not a good tag.
if (tid != TAG_TITLE && tid != TAG_A) {
continue;
}
// must NOT be a back tag
if ( tids[i] & BACKBIT ) {
continue;
}
// skip if we hit our limit
if ( table[tid] >= 4 ) {
continue;
}
// skip over tag/word #i
i++;
// no words in links, unless it is a self link
if ( i < NW && (flags[i] & 0x02) ) {
continue;
}
// the start should be here
int32_t start = -1;
// do not go too far
int32_t max = i + 200;
// find the corresponding back tag for it
for ( ; i < NW && i < max ; i++ ) {
// hey we got it, BUT we got no alnum word first
// so the thing was empty, so loop back to subloop
if ( (tids[i] & BACKBITCOMP) == tid &&
(tids[i] & BACKBIT ) &&
start == -1 ) {
goto subloop;
}
// if we hit another title-ish tag, loop back up
if ( (tids[i] & BACKBITCOMP) == TAG_TITLE || (tids[i] & BACKBITCOMP) == TAG_A ) {
// if no alnum text, restart at the top
if ( start == -1 ) {
goto subloop;
}
// otherwise, break out and see if title works
break;
}
// if we hit a breaking tag...
if ( isBreakingTagId ( tids[i] & BACKBITCOMP ) &&
// do not consider <span> tags breaking for
// our purposes. i saw a <h1><span> setup before.
tids[i] != TAG_SPAN ) {
break;
}
// skip if not alnum word
if ( ! words->isAlnum(i) ) {
continue;
}
// if we hit an alnum word, break out
if ( start == -1 ) {
start = i;
}
}
// if no start was found, must have had a 0 score in there
if ( start == -1 ) {
continue;
}
// if we exhausted the doc, we are done
if ( i >= NW ) {
break;
}
// skip if way too big!
if ( i >= max ) {
continue;
}
// if was too long do not consider a title
if ( i - start > 300 ) {
continue;
}
// . skip if too many bytes
// . this does not include the length of word #i, but #(i-1)
if ( words->getStringSize ( start , i ) > 1000 ) {
continue;
}
// when using pdftohtml, the title tag is the filename when PDF property does not have title tag
if ( tid == TAG_TITLE && contentType == CT_PDF ) {
// skip if title == '/in.[0-9]*'
char* title_start = words->getWord(start);
char* title_end = words->getWord(i);
size_t title_size = title_end - title_start;
const char* result = strnstr( title_start, "/in.", title_size );
if (result != NULL) {
char* endp = NULL;
// do some further verification to avoid screwing up title
if ((strtoll(result + 4, &endp, 10) > 0) && (endp == title_end)) {
continue;
}
}
}
// count it
table[tid]++;
// max it out if we are positive scoring. stop after the
// first positive scoring guy in a section. this might
// hurt the "Hamlet" thing though...
// store a point to the title tag guy. Msg20.cpp needs this
// because the zak's proximity algo uses it in Summary.cpp
// and in Msg20.cpp
// only get the first one! often the 2nd on is in an iframe!! which we now expand into here.
if ( tid == TAG_TITLE && m_titleTagStart == -1 ) {
m_titleTagStart = start;
m_titleTagEnd = i;
// save the candidate # because we always use this
// as the title if we are a root
if ( tti < 0 ) {
tti = n;
}
}
// point to words class of the body that was passed in to us
cptrs[n] = words;
as[n] = start;
bs[n] = i;
if ( tid == TAG_B ) {
types[n] = TT_BOLDTAG;
scores[n] = 1.0;
} else if ( tid == TAG_H1 ) {
types[n] = TT_HTAG;
scores[n] = 1.8;
} else if ( tid == TAG_H2 ) {
types[n] = TT_HTAG;
scores[n] = 1.7;
} else if ( tid == TAG_H3 ) {
types[n] = TT_HTAG;
scores[n] = 1.6;
} else if ( tid == TAG_TITLE ) {
types[n] = TT_TITLETAG;
scores[n] = 3.0;
} else if ( tid == TAG_DIV ) {
types[n] = TT_DIVTAG;
scores[n] = 1.0;
} else if ( tid == TAG_TD ) {
types[n] = TT_TDTAG;
scores[n] = 1.0;
} else if ( tid == TAG_P ) {
types[n] = TT_PTAG;
scores[n] = 1.0;
} else if ( tid == TAG_FONT ) {
types[n] = TT_FONTTAG;
scores[n] = 1.0;
} else if ( tid == TAG_A ) {
types[n] = TT_ATAG;
// . self link is very powerful BUT
// http://www.npr.org/templates/story/story.php?storyId=5417137
// doesn't use it right! so use
// 1.3 instead of 3.0. that has an "onClick" thing in the
// <a> tag, so check for that!
// this was bad for
// http://www.spiritualwoman.net/?cat=191
// so i am demoting from 3.0 to 1.5
scores[n] = 1.5;
}
// count it
n++;
// start loop over at tag #i, for loop does an i++, so negate
// that so this will work
i--;
// break out if too many already. save some for below.
if ( n + 10 >= MAX_TIT_CANDIDATES ) {
break;
}
}
//logf(LOG_DEBUG,"title: took3=%" PRId64,gettimeofdayInMilliseconds()-x);
//x = gettimeofdayInMilliseconds();
// to handle text documents, throw in the first line of text
// as a title candidate, just make the score really low
bool textDoc = (contentType == CT_UNKNOWN || contentType == CT_TEXT);
if (textDoc) {
// make "i" point to first alphabetical word in the document
int32_t i ;
for ( i = 0 ; i < NW && !words->isAlpha(i) ; i++);
// if we got a first alphabetical word, then assume that to be the start of our title
if ( i < NW && n < MAX_TIT_CANDIDATES ) {
// first word in title is "t0"
int32_t t0 = i;
// find end of first line
int32_t numWords = 0;
// set i to the end now. we MUST find a \n to terminate the
// title, otherwise we will not have a valid title
while (i < NW && numWords < maxTitleWords && (words->isAlnum(i) || !words->hasChar(i, '\n'))) {
if(words->isAlnum(i)) {
numWords++;
}
++i;
}
// "t1" is the end
int32_t t1 = -1;
// we must have found our \n in order to set "t1"
if (i <= NW && numWords < maxTitleWords ) {
t1 = i;
}
// set the ptrs
cptrs [n] = words;
// this is the last resort i guess...
scores [n] = 0.5;
types [n] = TT_FIRSTLINE;
as [n] = t0;
bs [n] = t1;
// add it as a candidate if t0 and t1 were valid
if (t0 >= 0 && t1 > t0) {
n++;
}
}
}
//logf(LOG_DEBUG,"title: took4=%" PRId64,gettimeofdayInMilliseconds()-x);
//x = gettimeofdayInMilliseconds();
{
// now add the last url path to contain underscores or hyphens
char *pstart = firstUrl->getPath();
// get first url
Url *fu = firstUrl;
// start at the end
char *p = fu->getUrl() + fu->getUrlLen();
// end pointer
char *pend = NULL;
// come up here for each path component
while ( p >= pstart ) {
// save end
pend = p;
// skip over /
if ( *p == '/' ) {
p--;
}
// now go back to next /
int32_t count = 0;
for ( ; p >= pstart && *p !='/' ; p-- ) {
if ( *p == '_' || *p == '-' ) {
count++;
}
}
// did we get it?
if ( count > 0 ) {
break;
}
}
// did we get any?
if ( p > pstart && n < MAX_TIT_CANDIDATES ) {
// now set words to that
if ( ! tw[ti].set ( p, (pend - p), true, 0 )) {
return false;
}
// point to that
cptrs [n] = &tw[ti];
as [n] = 0;
bs [n] = tw[ti].getNumWords();
scores [n] = 1.0;
types [n] = TT_URLPATH;
// increment since we are using it
ti++;
// advance
n++;
}
}
// save old n
int32_t oldn = n;
// . do not split titles if we are a root url maps.yahoo.com was getting "Maps" for the title
if ( firstUrl->isRoot() ) {
oldn = -2;
}
// point to list of \0 separated titles
const char *rootTitleBuf = NULL;
const char *rootTitleBufEnd = NULL;
// get the root title if we are not root!
if (filteredRootTitleBuf) {
#ifdef _VALGRIND_
VALGRIND_CHECK_MEM_IS_DEFINED(filteredRootTitleBuf,filteredRootTitleBufSize);
#endif
// point to list of \0 separated titles
rootTitleBuf = filteredRootTitleBuf;
rootTitleBufEnd = filteredRootTitleBuf + filteredRootTitleBufSize;
}
{
Matches m;
if ( rootTitleBuf && query ) {
m.setQuery ( query );
}
// convert into an array
int32_t nr = 0;
const char *pr = rootTitleBuf;
const char *rootTitles[20];
int32_t rootTitleLens[20];
// loop over each root title segment
for ( ; pr && pr < rootTitleBufEnd ; pr += strnlen(pr,rootTitleBufEnd-pr) + 1 ) {
// if we had a query...
if ( query ) {
// reset it
m.reset();
// see if root title segment has query terms in it
m.addMatches ( const_cast<char*>(pr), strnlen(pr,rootTitleBufEnd-pr), MF_TITLEGEN, m_niceness );
// if matches query, do NOT add it, we only add it for
// removing from the title of the page...
if ( m.getNumMatches() ) {
continue;
}
}
// point to it. it should start with an alnum already
// since it is the "filtered" list of root titles...
// if not, fix it in xmldoc then.
rootTitles [nr] = pr;
rootTitleLens[nr] = gbstrlen(pr);
// advance
nr++;
// no breaching
if ( nr >= 20 ) break;
}
// now split up candidates in children candidates by tokenizing
// using :, | and - as delimters.
// the hyphen must have a space on at least one side, so "cd-rom" does
// not create a pair of tokens...
// FIX: for the title:
// Best Careers 2009: Librarian - US News and World Report
// we need to recognize "Best Careers 2009: Librarian" as a subtitle
// otherwise we don't get it as the title. so my question is are we
// going to have to do all the permutations at some point? for now
// let's just add in pairs...
for ( int32_t i = 0 ; i < oldn && n + 3 < MAX_TIT_CANDIDATES ; i++ ) {
// stop if no root title segments
if ( nr <= 0 ) break;
// get the word info
Words *w = cptrs[i];
int32_t a = as[i];
int32_t b = bs[i];
// init
int32_t lasta = a;
char prev = false;
// char length in bytes
//int32_t charlen = 1;
// see how many we add
int32_t added = 0;
char *skipTo = NULL;
bool qualified = true;
// . scan the words looking for a token
// . sometimes the candidates end in ": " so put in "k < b-1"
// . made this from k<b-1 to k<b to fix
// "Hot Tub Time Machine (2010) - IMDb" to strip IMDb
for ( int32_t k = a ; k < b && n + 3 < MAX_TIT_CANDIDATES; k++){
// get word
char *wp = w->getWord(k);
// skip if not alnum
if ( ! w->isAlnum(k) ) {
// in order for next alnum word to
// qualify for "clipping" if it matches
// the root title, there has to be more
// than just spaces here, some punct.
// otherwise title
// "T. D. Jakes: Biography from Answers.com"
// becomes
// "T. D. Jakes: Biography from"
qualified=isWordQualified(wp,w->getWordLen(k));
continue;
}
// gotta be qualified!
if ( ! qualified ) continue;
// skip if in root title
if ( skipTo && wp < skipTo ) continue;
// does this match any root page title segments?
int32_t j;
for ( j = 0 ; j < nr ; j++ ) {
// . compare to root title
// . break out if we matched!
if ( ! strncmp( wp, rootTitles[j], rootTitleLens[j] ) ) {
break;
}
}
// if we did not match a root title segment,
// keep on chugging
if ( j >= nr ) continue;
// . we got a root title match!
// . skip over
skipTo = wp + rootTitleLens[j];
// must land on qualified punct then!!
int32_t e = k+1;
for ( ; e<b && w->getWord(e)<skipTo ; e++ );
// ok, word #e must be a qualified punct
if ( e<b &&
! isWordQualified(w->getWord(e),w->getWordLen(e)))
// assume no match then!!
continue;
// if we had a previous guy, reset the end of the
// previous candidate
if ( prev ) {
bs[n-2] = k;
bs[n-1] = k;
}
// . ok, we got two more candidates
// . well, only one more if this is not the 1st time
if ( ! prev ) {
cptrs [n] = cptrs [i];
scores [n] = scores [i];
types [n] = types [i];
as [n] = lasta;
bs [n] = k;
parent [n] = i;
n++;
added++;
}
// the 2nd one
cptrs [n] = cptrs [i];
scores [n] = scores [i];
types [n] = types [i];
as [n] = e + 1;
bs [n] = bs [i];
parent [n] = i;
n++;
added++;
// now add in the last pair as a whole token
cptrs [n] = cptrs [i];
scores [n] = scores [i];
types [n] = types [i];
as [n] = lasta;
bs [n] = bs [i];
parent [n] = i;
n++;
added++;
// nuke the current candidate then since it got
// split up to not contain the root title...
//cptrs[i] = NULL;
// update this
lasta = k+1;
// if we encounter another delimeter we will have to revise bs[n-1], so note that
prev = true;
}
// nuke the current candidate then since it got
// split up to not contain the root title...
if ( added ) {
scores[i] = 0.001;
//cptrs[i] = NULL;
}
// erase the pair if that there was only one token
if ( added == 3 ) n--;
}
}
for ( int32_t i = 0 ; i < n ; i++ ) baseScore[i] = scores[i];
//
// . now punish by 0.85 for every lower case non-stop word it has
// . reward by 1.1 if has a non-stopword in the query
//
for ( int32_t i = 0 ; i < n ; i++ ) {
// point to the words
Words *w = cptrs[i];
// skip if got nuked above
if ( ! w ) {
continue;
}
// the word ptrs
char **wptrs = w->getWordPtrs();
// skip if empty
if ( w->getNumWords() <= 0 ) {
continue;
}
// get the word boundaries
int32_t a = as[i];
int32_t b = bs[i];
// record the boosts
float ncb = 1.0;
float qtb = 1.0;
// a flag
char uncapped = false;
// scan the words in this title candidate
for ( int32_t j = a ; j < b ; j++ ) {
// skip stop words
if ( w->isQueryStopWord( j, langId ) ) {
continue;
}
// punish if uncapitalized non-stopword
if ( ! w->isCapitalized(j) ) {
uncapped = true;
}
// skip if no query
if ( ! query ) {
continue;
}
int64_t wid = w->getWordId(j);
// reward if in the query
if ( query->getWordNum(wid) >= 0 ) {
qtb *= 1.5;
scores[i] *= 1.5;
}
}
// . only punish once if missing a capitalized word hurts us for:
// http://content-uk.cricinfo.com/ausvrsa2008_09/engine/current/match/351682.html
if ( uncapped ) {
ncb *= 1.00;
scores[i] *= 1.00;
}
// punish if a http:// title thingy
char *s = wptrs[a];
int32_t size = w->getStringSize(a,b);
if ( size > 9 && memcmp("http://", s, 7) == 0 ) {
ncb *= .10;
}
if ( size > 14 && memcmp("h\0t\0t\0p\0:\0/\0/", s, 14) == 0 ) {
ncb *= .10;
}
// set these guys
scores[i] *= ncb;
noCapsBoost[i] = ncb;
qtermsBoost[i] = qtb;
}
// . now compare each candidate to the other candidates
// . give a boost if matches
for ( int32_t i = 0 ; i < n ; i++ ) {
// point to the words
Words *w1 = cptrs[i];
// skip if got nuked above
if ( ! w1 ) {
continue;
}
int32_t a1 = as[i];
int32_t b1 = bs[i];
// reset some flags
char localFlag1 = 0;
char localFlag2 = 0;
// record the boost
float iccb = 1.0;
// total boost
float total = 1.0;
// to each other candidate
for ( int32_t j = 0 ; j < n ; j++ ) {
// not to ourselves
if ( j == i ) {
continue;
}
// or our derivatives
if ( parent[j] == i ) {
continue;
}
// or derivates to their parent
if ( parent[i] == j ) {
continue;
}
// only check parents now. do not check kids.
// this was only for when doing percent contained
// not getSimilarity() per se
//if ( parent[j] != -1 ) continue;
// TODO: do not accumulate boosts from a parent
// and its kids, subtitles...
//
// do not compare type X to type Y
if ( types[i] == TT_TITLETAG ) {
if ( types[j] == TT_TITLETAG ) {
continue;
}
}
// do not compare a div candidate to another div cand
// http://friendfeed.com/foxiewire?start=30
// likewise, a TD to another TD
// http://content-uk.cricinfo.com/ausvrsa2008_09/engine/match/351681.html
// ... etc.
if ( types[i] == TT_BOLDTAG ||
types[i] == TT_HTAG ||
types[i] == TT_DIVTAG ||
types[i] == TT_TDTAG ||
types[i] == TT_FONTTAG ) {
if ( types[j] == types[i] ) continue;
}
// . do not compare one kid to another kid
// . i.e. if we got "x | y" as a title and "x | z"
// as a link text, it will emphasize "x" too much
// http://content-uk.cricinfo.com/ausvrsa2008_09/engine/current/match/351682.html
if ( parent[j] != -1 && parent[i] != -1 ) continue;
// . body type tags are mostly mutually exclusive
// . for the legacy.com url mentioned below, we have
// good stuff in <td> tags, so this hurts us...
// . but for the sake of
// http://larvatusprodeo.net/2009/01/07/partisanship-politics-and-participation/
// i put bold tags back
if ( types[i] == TT_LINKTEXTLOCAL ) {
if ( types[j] == TT_LINKTEXTLOCAL ) continue;
}
if ( types[i] == TT_RSSITEMLOCAL ) {
if ( types[j] == TT_RSSITEMLOCAL ) continue;
}
// only compare to one local link text for each i
if ( types[j] == TT_LINKTEXTLOCAL && localFlag1 ) {
continue;
}
if ( types[j] == TT_RSSITEMLOCAL && localFlag2 ) {
continue;
}
if ( types[j] == TT_LINKTEXTLOCAL ) {
localFlag1 = 1;
}
if ( types[j] == TT_RSSITEMLOCAL ) {
localFlag2 = 1;
}
// not link title attr to link title attr either
// fixes http://www.spiritualwoman.net/?cat=191
if ( types[i] == TT_TITLEATT &&
types[j] == TT_TITLEATT )
continue;
// get our words
Words *w2 = cptrs[j];
// skip if got nuked above
if ( ! w2 ) continue;
int32_t a2 = as [j];
int32_t b2 = bs [j];
// how similar is title #i to title #j ?
float fp = getSimilarity ( w2 , a2 , b2 , w1 , a1 , b1 );
// error?
if ( fp == -1.0 ) return false;
// custom boosting...
float boost = 1.0;
if ( fp >= .95 ) boost = 3.0;
else if ( fp >= .90 ) boost = 2.0;
else if ( fp >= .85 ) boost = 1.5;
else if ( fp >= .80 ) boost = 1.4;
else if ( fp >= .75 ) boost = 1.3;
else if ( fp >= .70 ) boost = 1.2;
else if ( fp >= .60 ) boost = 1.1;
else if ( fp >= .50 ) boost = 1.08;
else if ( fp >= .40 ) boost = 1.04;
// limit total
total *= boost;
if ( total > 100.0 ) break;
// if you are matching the url path, that is pretty
// good so give more!
// actually, that would hurt:
// http://michellemalkin.com/2008/12/29/gag-worthy/
// custom boosting!
if ( fp > 0.0 && g_conf.m_logDebugTitle )
logf(LOG_DEBUG,"title: i=%" PRId32" j=%" PRId32" fp=%.02f "
"b=%.02f", i,j,fp,boost);
// apply it
scores[i] *= boost;
iccb *= boost;
}
inCommonCandBoost[i] = iccb;
}
//logf(LOG_DEBUG,"title: took7=%" PRId64,gettimeofdayInMilliseconds()-x);
//x = gettimeofdayInMilliseconds();
// loop over all n candidates
for ( int32_t i = 0 ; i < n ; i++ ) {
// skip if not in the document body
if ( cptrs[i] != words ) continue;
// point to the words
int32_t a1 = as [i];
int32_t b1 = bs [i];
// . loop through this candidates words
// . TODO: use memset here?
for ( int32_t j = a1 ; j <= b1 && j < NW ; j++ ) {
// flag it
flags[j] |= 0x01;
}
}
// free our stuff
if ( flags!=localBuf ) {
mfree (flags, need, "TITLEflags");
}
// now get the highest scoring candidate title
float max = -1.0;
int32_t winner = -1;
for ( int32_t i = 0 ; i < n ; i++ ) {
// skip if got nuked
if ( ! cptrs[i] ) {
continue;
}
if ( winner != -1 && scores[i] <= max ) {
continue;
}
// url path's cannot be titles in and of themselves
if ( types[i] == TT_URLPATH ) {
continue;
}
// skip if empty basically, like if title was exact
// copy of root, then the whole thing got nuked and
// some empty string added, where a > b
if ( as[i] >= bs[i] ) {
continue;
}
// got one
max = scores[i];
// save it
winner = i;
}
// if we are a root, always pick the title tag as the title
if ( oldn == -2 && tti >= 0 ) {
winner = tti;
}
// if no winner, all done. no title
if ( winner == -1 ) {
// last resort use file name
if ((contentType == CT_PDF) && (firstUrl->getFilenameLen() != 0)) {
Words w;
w.set(firstUrl->getFilename(), firstUrl->getFilenameLen(), true);
if (!copyTitle(&w, 0, w.getNumWords())) {
return false;
}
}
return true;
}
// point to the words class of the winner
Words *w = cptrs[winner];
// skip if got nuked above
if ( ! w ) { char *xx=NULL;*xx=0; }
// need to make our own Pos class if title not from body
Pos tp;
if ( w != words ) {
// set "Scores" ptr to NULL. we assume all are positive scores
if ( ! tp.set ( w ) ) {
return false;
}
}
// the string ranges from word #a up to and including word #b
int32_t a = as[winner];
int32_t b = bs[winner];
// sanity check
if ( a < 0 || b > w->getNumWords() ) { char*xx=NULL;*xx=0; }
// save the title
if ( ! copyTitle(w, a, b) ) {
return false;
}
/*
// debug logging
SafeBuf sb;
SafeBuf *pbuf = &sb;
log("title: candidates for %s",xd->getFirstUrl()->getUrl() );
pbuf->safePrintf("<div stype=\"border:1px solid black\">");
pbuf->safePrintf("<b>***Finding Title***</b><br>\n");
pbuf->safePrintf("<table cellpadding=5 border=2><tr>"
"<td colspan=20><center><b>Title Generation</b>"
"</center></td>"
"</tr>\n<tr>"
"<td>#</td>"
"<td>type</td>"
"<td>parent</td>"
"<td>base score</td>"
"<td>format penalty</td>"
"<td>query term boost</td>"
"<td>candidate intersection boost</td>"
"<td>FINAL SCORE</td>"
"<td>title</td>"
"</tr>\n" );
// print out all candidates
for ( int32_t i = 0 ; i < n ; i++ ) {
char *ts = "unknown";
if ( types[i] == TT_LINKTEXTLOCAL ) ts = "local inlink text";
if ( types[i] == TT_LINKTEXTREMOTE ) ts = "remote inlink text";
if ( types[i] == TT_RSSITEMLOCAL ) ts = "local rss title";
if ( types[i] == TT_RSSITEMREMOTE ) ts = "remote rss title";
if ( types[i] == TT_BOLDTAG ) ts = "bold tag";
if ( types[i] == TT_HTAG ) ts = "header tag";
if ( types[i] == TT_TITLETAG ) ts = "title tag";
if ( types[i] == TT_FIRSTLINE ) ts = "first line in text";
if ( types[i] == TT_FONTTAG ) ts = "font tag";
if ( types[i] == TT_ATAG ) ts = "anchor tag";
if ( types[i] == TT_DIVTAG ) ts = "div tag";
if ( types[i] == TT_TDTAG ) ts = "td tag";
if ( types[i] == TT_PTAG ) ts = "p tag";
if ( types[i] == TT_URLPATH ) ts = "url path";
if ( types[i] == TT_TITLEATT ) ts = "title attribute";
// get the title
pbuf->safePrintf(
"<tr>"
"<td>#%" PRId32"</td>"
"<td><nobr>%s</nobr></td>"
"<td>%" PRId32"</td>"
"<td>%0.2f</td>" // baseScore
"<td>%0.2f</td>"
"<td>%0.2f</td>"
"<td>%0.2f</td>"
"<td>%0.2f</td>"
"<td>",
i,
ts ,
parent[i],
baseScore[i],
noCapsBoost[i],
qtermsBoost[i],
inCommonCandBoost[i],
scores[i]);
// ptrs
Words *w = cptrs[i];
int32_t a = as[i];
int32_t b = bs[i];
// skip if no words
if ( w->getNumWords() <= 0 ) continue;
// the word ptrs
char **wptrs = w->getWordPtrs();
// string ptrs
char *ptr = wptrs[a];//w->getWord(a);
int32_t size = w->getStringSize(a,b);
// it is utf8
pbuf->safeMemcpy ( ptr , size );
// end the line
pbuf->safePrintf("</td></tr>\n");
}
pbuf->safePrintf("</table>\n<br>\n");
// log these for now
log("title: %s",sb.getBufStart());
*/
return true;
}
// returns false and sets g_errno on error
bool Summary::setSummary ( Xml *xml, Words *words, Sections *sections, Pos *pos, Query *q, int32_t maxSummaryLen,
int32_t maxNumLines, int32_t numDisplayLines, int32_t maxNumCharsPerLine, Url *f,
Matches *matches, char *titleBuf, int32_t titleBufLen ) {
m_numDisplayLines = numDisplayLines;
m_displayLen = 0;
// assume we got maxnumlines of summary
if ( (maxNumCharsPerLine + 6) * maxNumLines > maxSummaryLen ) {
if ( maxNumCharsPerLine < 10 ) {
maxNumCharsPerLine = 10;
}
static char s_flag = 1;
if ( s_flag ) {
s_flag = 0;
log("query: Warning. "
"Max summary excerpt length decreased to "
"%" PRId32" chars because max summary excerpts and "
"max summary length are too big.",
maxNumCharsPerLine);
}
}
// . sanity check
// . summary must fit in m_summary[]
// . leave room for tailing \0
if ( maxSummaryLen >= MAX_SUMMARY_LEN ) {
g_errno = EBUFTOOSMALL;
return log("query: Summary too big to hold in buffer of %" PRId32" bytes.",(int32_t)MAX_SUMMARY_LEN);
}
// do not overrun the final*[] buffers
if ( maxNumLines > 256 ) {
g_errno = EBUFTOOSMALL;
return log("query: More than 256 summary lines requested.");
}
// Nothing to match...print beginning of content as summary
if ( matches->m_numMatches == 0 && maxNumLines > 0 ) {
return getDefaultSummary ( xml, words, sections, pos, maxSummaryLen );
}
int32_t need1 = q->m_numWords * sizeof(float);
m_wordWeightSize = need1;
if ( need1 < 128 ) {
m_wordWeights = (float *)m_tmpWordWeightsBuf;
} else {
m_wordWeights = (float *)mmalloc ( need1 , "wwsum" );
}
if ( ! m_wordWeights ) {
return false;
}
/// @todo ALC fix word weights
/// non-working logic is removed in commit 5eacee9063861e859b54ec62035a600aa8af25df
// . compute our word weights wrt each query. words which are more rare
// have a higher weight. We use this to weight the terms importance
// when generating the summary.
// . used by the proximity algo
// . used in setSummaryScores() for scoring summaries
for ( int32_t i = 0 ; i < q->m_numWords; i++ ) {
m_wordWeights[i] = 1.0;
}
// convenience
m_maxNumCharsPerLine = maxNumCharsPerLine;
m_q = q;
// set the max excerpt len to the max summary excerpt len
int32_t maxExcerptLen = m_maxNumCharsPerLine;
int32_t lastNumFinal = 0;
int32_t maxLoops = 1024;
// if just computing absScore2...
if ( maxNumLines <= 0 ) {
return true;
}
char *p = m_summary;
char *pend = m_summary + maxSummaryLen;
m_numExcerpts = 0;
int32_t need2 = (1+1+1) * m_q->m_numWords;
m_buf4Size = need2;
if ( need2 < 128 ) {
m_buf4 = m_tmpBuf4;
} else {
m_buf4 = (char *)mmalloc ( need2 , "stkbuf" );
}
if ( ! m_buf4 ) {
return false;
}
char *x = m_buf4;
char *retired = x;
x += m_q->m_numWords;
char *maxGotIt = x;
x += m_q->m_numWords;
char *gotIt = x;
// . the "maxGotIt" count vector accumulates into "retired"
// . that is how we keep track of what query words we used for previous
// summary excerpts so we try to get diversified excerpts with
// different query terms/words in them
//char retired [ MAX_QUERY_WORDS ];
memset ( retired, 0, m_q->m_numWords * sizeof(char) );
// some query words are already matched in the title
for ( int32_t i = 0 ; i < m_q->m_numWords ; i++ ) {
if ( matches->m_qwordFlags[i] & MF_TITLEGEN ) {
retired [ i ] = 1;
}
}
bool hadEllipsis = false;
//
// Loop over all words that match a query term. The matching words
// could be from any one of the 3 Words arrays above. Find the
// highest scoring window around each term. And then find the highest
// of those over all the matching terms.
//
int32_t numFinal;
for ( numFinal = 0; numFinal < maxNumLines; numFinal++ ) {
if ( numFinal == m_numDisplayLines ) {
m_displayLen = p - m_summary;
}
// reset these at the top of each loop
Match *maxm;
int64_t maxScore = 0;
int32_t maxa = 0;
int32_t maxb = 0;
int32_t maxi = -1;
int32_t lasta = -1;
if(lastNumFinal == numFinal) {
if(maxLoops-- <= 0) {
log(LOG_WARN, "query: got infinite loop bug, query is %s url is %s", m_q->m_orig, f->getUrl());
break;
}
}
lastNumFinal = numFinal;
// loop through all the matches and see which is best
for ( int32_t i = 0 ; i < matches->m_numMatches ; i++ ) {
int32_t a , b;
// reset lasta if we changed words class
if ( i > 0 && matches->m_matches[i-1].m_words != matches->m_matches[i].m_words ) {
lasta = -1;
}
// only use matches in title, etc.
mf_t flags = matches->m_matches[i].m_flags;
bool skip = true;
if ( flags & MF_METASUMM ) {
skip = false;
}
if ( flags & MF_METADESC ) {
skip = false;
}
if ( flags & MF_BODY ) {
skip = false;
}
if ( flags & MF_RSSDESC ) {
skip = false;
}
if ( skip ) {
continue;
}
// ask him for the query words he matched
//char gotIt [ MAX_QUERY_WORDS ];
// clear it for him
memset ( gotIt, 0, m_q->m_numWords * sizeof(char) );
// . get score of best window around this match
// . do not allow left post of window to be <= lasta to
// avoid repeating the same window.
int64_t score = getBestWindow (matches, i, &lasta, &a, &b, gotIt, retired, maxExcerptLen);
// USE THIS BUF BELOW TO DEBUG THE ABOVE CODE.
// PRINTS OUT THE SUMMARY
/*
//if ( score >=12000 ) {
char buf[10*1024];
char *xp = buf;
if ( i == 0 )
log (LOG_WARN,"=-=-=-=-=-=-=-=-=-=-=-=-=-=-=");
sprintf(xp, "score=%08" PRId32" a=%05" PRId32" b=%05" PRId32" ",
(int32_t)score,(int32_t)a,(int32_t)b);
xp += strlen(xp);
for ( int32_t j = a; j < b; j++ ){
//int32_t s = scores->m_scores[j];
int32_t s = 0;
if ( s < 0 ) continue;
char e = 1;
int32_t len = words->getWordLen(j);
for(int32_t k=0;k<len;k +=e){
char c = words->m_words[j][k];
//if ( is_binary( c ) ) continue;
*xp = c;
xp++;
}
//p += strlen(p);
if ( s == 0 ) continue;
sprintf ( xp ,"(%" PRId32")",s);
xp += strlen(xp);
}
log (LOG_WARN,"query: summary: %s", buf);
//}
*/
// prints out the best window with the score
/*
char buf[MAX_SUMMARY_LEN];
char *bufPtr = buf;
char *bufPtrEnd = p + MAX_SUMMARY_LEN;
if ( i == 0 )
log (LOG_WARN,"=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=");
int32_t len = 0;
Words *ww = matches->m_matches[i].m_words;
//Sections *ss = matches->m_matches[i].m_sections;
//if ( ss->m_numSections <= 0 ) ss = NULL;
//len=pos->filter(bufPtr, bufPtrEnd, ww, a, b, NULL);
//log(LOG_WARN,"summary: %" PRId32") %s - %" PRId64,i,bufPtr,
//score);
log(LOG_WARN,"summary: %" PRId32") %s - %" PRId64,i,bufPtr,
score);
*/
// skip if was in title or something
if ( score <= 0 ) {
continue;
}
// skip if not a winner
if ( maxi >= 0 && score <= maxScore ) {
continue;
}
// we got a new winner
maxi = i;
maxa = a;
maxb = b;
maxScore = score;
// save this too
gbmemcpy ( maxGotIt , gotIt , m_q->m_numWords );
}
// retire the query words in the winning summary
//log( LOG_WARN,"summary: took %" PRId64" ms to finish getbestwindo",
// gettimeofdayInMilliseconds() - stget );
// all done if no winner was made
if ( maxi == -1 || maxa == -1 || maxb == -1) {
break;
}
// who is the winning match?
maxm = &matches->m_matches[maxi];
Words *ww = maxm->m_words;
// we now use "m_swbits" for the summary bits since they are
// of size sizeof(swbit_t), a int16_t at this point
swbit_t *bb = maxm->m_bits->m_swbits;
// this should be impossible
if ( maxa > ww->getNumWords() || maxb > ww->getNumWords() ) {
log ( LOG_WARN,"query: summary starts or ends after "
"document is over! maxa=%" PRId32" maxb=%" PRId32" nw=%" PRId32,
maxa, maxb, ww->getNumWords() );
maxa = ww->getNumWords() - 1;
maxb = ww->getNumWords();
}
// assume we do not preceed with ellipsis "..."
bool needEllipsis = true;
const char *c = ww->getWord(maxa)+0;
// rule of thumb, don't use ellipsis if the first letter is capital, or a non letter
// is punct word before us pair acrossable? if so then we probably are not the start of a sentence.
// or if into the sample and previous excerpt had an ellipsis do not bother using one for us.
if ( !is_alpha_utf8(c) || is_upper_utf8(c) ||
(bb[maxa] & D_STARTS_SENTENCE) ||
(p > m_summary && hadEllipsis)) {
needEllipsis = false;
}
if ( needEllipsis ) {
// break out if no room for "..."
if ( p + 4 + 2 > pend ) {
break;
}
// space first?
if ( p > m_summary ) {
*p++ = ' ';
}
memcpy ( p, "\342\200\246 ", 4 ); //horizontal ellipsis, code point 0x2026
p += 4;
}
// separate summary excerpts with a single space.
if ( p > m_summary ) {
if ( p + 2 > pend ) {
break;
}
*p++ = ' ';
}
// assume we need a trailing ellipsis
needEllipsis = true;
// so next excerpt does not need to have an ellipsis if we
// have one at the end of this excerpt
hadEllipsis = needEllipsis;
// start with quote?
if ( (bb[maxa] & D_IN_QUOTES) && p + 1 < pend ) {
// preceed with quote
*p++ = '\"';
}
// . filter the words into p
// . removes back to back spaces
// . converts html entities
// . filters in stores words in [a,b) interval
int32_t len = pos->filter( ww, maxa, maxb, false, p, pend, xml->getVersion() );
// break out if did not fit
if ( len == 0 ) {
break;
}
// don't consider it if it is a substring of the title
if ( len == titleBufLen && strncasestr(titleBuf, p, titleBufLen, len) ) {
// don't consider this one
numFinal--;
goto skip;
}
// don't consider it if the length wasn't anything nice
if ( len < 5 ){
numFinal--;
goto skip;
}
// otherwise, keep going
p += len;
// now we just indicate which query terms we got
for ( int32_t i = 0 ; i < m_q->m_numWords ; i++ ) {
// do not breach
if ( retired[i] >= 100 ) {
continue;
}
retired [ i ] += maxGotIt [ i ];
}
// add all the scores of the excerpts to the doc summary score.
// zero out scores of the winning sample so we don't get them
// again. use negative one billion to ensure that we don't get
// them again
for ( int32_t j = maxa ; j < maxb ; j++ ) {
// mark it as used
bb[j] |= D_USED;
}
// if we ended on punct that can be paired across we need
// to add an ellipsis
if ( needEllipsis ) {
if ( p + 4 + 2 > pend ) {
break;
}
memcpy ( p, " \342\200\246", 4 ); //horizontal ellipsis, code point 0x2026
p += 4;
}
// try to put in a small summary excerpt if we have atleast
// half of the normal excerpt length left
if ( maxExcerptLen == m_maxNumCharsPerLine && len <= ( m_maxNumCharsPerLine / 2 + 1 ) ) {
maxExcerptLen = m_maxNumCharsPerLine / 2;
// don't count it in the finals since we try to get a small excerpt
numFinal--;
} else if ( m_numExcerpts < MAX_SUMMARY_EXCERPTS && m_numExcerpts >= 0 ) {
m_summaryExcerptLen[m_numExcerpts] = p - m_summary;
m_numExcerpts++;
// also reset maxExcerptLen
maxExcerptLen = m_maxNumCharsPerLine;
}
skip:
// zero out the scores so they will not be used in others
for ( int32_t j = maxa ; j < maxb ; j++ ) {
// mark it
bb[j] |= D_USED;
}
}
if ( numFinal <= m_numDisplayLines ) {
m_displayLen = p - m_summary;
}
// free the mem we used if we allocated it
if ( m_buf4 && m_buf4 != m_tmpBuf4 ) {
mfree ( m_buf4 , m_buf4Size , "ssstkb" );
m_buf4 = NULL;
}
// If we still didn't find a summary, get the default summary
if ( p == m_summary ) {
bool status = getDefaultSummary ( xml, words, sections, pos, maxSummaryLen );
if ( m_numDisplayLines > 0 ) {
m_displayLen = m_summaryLen;
}
return status;
}
// if we don't find a summary, theres no need to NULL terminate
*p++ = '\0';
// set length
m_summaryLen = p - m_summary;
if ( m_summaryLen > 50000 ) { g_process.shutdownAbort(true); }
return true;
}