1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
|
/** @file
* @brief Xapian::PL2PlusWeight class - the PL2+ weighting scheme of the DFR framework.
*/
/* Copyright (C) 2013 Aarsh Shah
* Copyright (C) 2013,2014,2016,2017,2024 Olly Betts
* Copyright (C) 2016 Vivek Pal
*
* 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., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include <config.h>
#include "xapian/weight.h"
#include "common/log2.h"
#include "serialise-double.h"
#include "xapian/error.h"
#include <algorithm>
using namespace std;
namespace Xapian {
PL2PlusWeight::PL2PlusWeight(double c, double delta)
: param_c(c), param_delta(delta)
{
if (param_c <= 0)
throw Xapian::InvalidArgumentError("Parameter c is invalid");
if (param_delta <= 0)
throw Xapian::InvalidArgumentError("Parameter delta is invalid");
need_stat(AVERAGE_LENGTH);
need_stat(DOC_LENGTH);
need_stat(DOC_LENGTH_MIN);
need_stat(DOC_LENGTH_MAX);
need_stat(COLLECTION_SIZE);
need_stat(COLLECTION_FREQ);
need_stat(WDF);
need_stat(WDF_MAX);
need_stat(WQF);
}
PL2PlusWeight *
PL2PlusWeight::clone() const
{
return new PL2PlusWeight(param_c, param_delta);
}
void
PL2PlusWeight::init(double factor_)
{
if (factor_ == 0.0) {
// This object is for the term-independent contribution, and that's
// always zero for this scheme.
return;
}
factor = factor_ * get_wqf();
auto wdf_upper_bound = get_wdf_upper_bound();
mean = double(get_collection_freq()) / get_collection_size();
if (rare(wdf_upper_bound == 0 || mean > 1)) {
// PL2+ is based on a modified PL2 which "essentially ignores
// non-discriminative query terms".
upper_bound = 0;
return;
}
double base_change(1.0 / log(2.0));
P1 = mean * base_change + 0.5 * log2(2.0 * M_PI);
P2 = log2(mean) + base_change;
cl = param_c * get_average_length();
double wdfn_lower = log2(1 + cl / get_doclength_upper_bound());
double divisior = max(get_wdf_upper_bound(), get_doclength_lower_bound());
double wdfn_upper = get_wdf_upper_bound() * log2(1 + cl / divisior);
double P_delta = P1 + (param_delta + 0.5) * log2(param_delta) - P2 * param_delta;
dw = P_delta / (param_delta + 1.0);
// Calculate an upper bound on the weights which get_sumpart() can return.
//
// We consider the equation for P as the sum of two parts which we
// maximise individually:
//
// (a) (wdfn + 0.5) / (wdfn + 1) * log2(wdfn)
// (b) (P1 - P2 * wdfn) / (wdfn + 1)
//
// To maximise (a), the fractional part is always positive (since wdfn>0)
// and is maximised by maximising wdfn - clearer when rewritten as:
// (1 - 0.5 / (wdfn + 1))
//
// The log part of (a) is clearly also maximised by maximising wdfn,
// so we want to evaluate (a) at wdfn=wdfn_upper.
double P_max2a = (wdfn_upper + 0.5) * log2(wdfn_upper) / (wdfn_upper + 1.0);
// To maximise (b) substitute x=wdfn+1 (so x>1) and we get:
//
// (P1 + P2)/x - P2
//
// Differentiating wrt x gives:
//
// -(P1 + P2)/x²
//
// So there are no local minima or maxima, and the function is continuous
// in the range of interest, so the sign of this differential tells us
// whether we want to maximise or minimise wdfn, and the denominator is
// always positive so we can just consider the sign of: (P1 + P2)
//
// Commonly P1 + P2 > 0, in which case we evaluate P at wdfn=wdfn_upper
// giving us a bound that can't be bettered if wdfn_upper is tight.
double wdfn_optb = P1 + P2 > 0 ? wdfn_upper : wdfn_lower;
double P_max2b = (P1 - P2 * wdfn_optb) / (wdfn_optb + 1.0);
upper_bound = factor * (P_max2a + P_max2b + dw);
if (rare(upper_bound < 0)) upper_bound = 0;
}
string
PL2PlusWeight::name() const
{
return "Xapian::PL2PlusWeight";
}
string
PL2PlusWeight::serialise() const
{
string result = serialise_double(param_c);
result += serialise_double(param_delta);
return result;
}
PL2PlusWeight *
PL2PlusWeight::unserialise(const string & s) const
{
const char *ptr = s.data();
const char *end = ptr + s.size();
double c = unserialise_double(&ptr, end);
double delta = unserialise_double(&ptr, end);
if (rare(ptr != end))
throw Xapian::SerialisationError("Extra data in PL2PlusWeight::unserialise()");
return new PL2PlusWeight(c, delta);
}
double
PL2PlusWeight::get_sumpart(Xapian::termcount wdf, Xapian::termcount len,
Xapian::termcount) const
{
// Note: lambda_t in the paper is 1/mean.
if (wdf == 0 || mean > 1) {
// PL2+ is based on a modified PL2 which "essentially ignores
// non-discriminative query terms".
return 0.0;
}
double wdfn = wdf * log2(1 + cl / len);
double P = P1 + (wdfn + 0.5) * log2(wdfn) - P2 * wdfn;
double wt = (P / (wdfn + 1.0)) + dw;
// FIXME: Is a negative wt possible here? It is with vanilla PL2, but for
// PL2+ we've added on dw, and bailed out early if mean < 1.
if (rare(wt <= 0)) return 0.0;
return factor * wt;
}
double
PL2PlusWeight::get_maxpart() const
{
return upper_bound;
}
double
PL2PlusWeight::get_sumextra(Xapian::termcount, Xapian::termcount) const
{
return 0;
}
double
PL2PlusWeight::get_maxextra() const
{
return 0;
}
}
|
