Information Technology Reference
In-Depth Information
terms threshold, English has three metrics with the best results, the one already
mentioned and
Least Median Tf-Idf and Phi-Square
.
As pointed above, Rvar and MI metrics alone were unable to discriminate the top
5, 10 or 20 best ranked terms. This probably explains the need to use the Least and
Median operators proposed by [1,2]. Precision for the Rvar and MI derived metrics is
shown in table 3. It shows clearly that Tf-Idf and Phi Square based metrics, in table 2,
are much better than those based on Rvar and MI. They get for the best metrics,
values a bit higher than 0.50, and generally all bellow 0.50 which makes the average
precision for these metrics rather poor.
Table 2.
Average precision values for the 5, 10 and 20 best terms using the best metrics, and
average for each threshold
Czech
Metric
Prec. (5)
Prec. (10)
Prec. (20)
Tf-Idf
0.90
0.86
0.66
L Tf-Idf
0.75
0.70
0.61
LM Tf-Idf
0.70
0.65
0.59
LB Tf-Idf
0.80
0.68
0.65
LBM Tf-Idf
0.65
0.68
0.66
0.70
0.70
0.61
ϕ2
0.70
0.60
0.58
L ϕ2
LM
ϕ
2
0.70
0.60
0.58
0.55
0.63
0.55
LB ϕ2
0.55
0.65
0.59
LBM ϕ2
Average
0.72
0.68
0.61
English
Metric
Prec. (5)
Prec. (10)
Prec. (20)
Tf-Idf
0.84
0.74
0.67
L Tf-Idf
0.78
0.66
0.68
LM Tf-Idf
0.81
0.78
0.66
LB Tf-Idf
0.85
0.66
0.65
LBM Tf-Idf
0.82
0.69
0.62
0.84
0.78
0.68
ϕ2
0.83
0.76
0.69
L ϕ2
LM
ϕ
2
0.87
0.78
0.70
0.83
0.74
0.62
LB ϕ2
0.80
0.74
0.65
LBM ϕ2
Average
0.83
0.73
0.66
Portuguese
Metric
Prec. (5)
Prec. (10)
Prec. (20)
Tf-Idf
0.69
0.70
0.66
L Tf-Idf
0.64
0.66
0.65
LM Tf-Idf
0.68
0.63
0.64
LB Tf-Idf
0.86
0.71
0.65
LBM Tf-Idf
0.83
0.70
0.68
0.73
0.73
0.62
ϕ2
L
ϕ
2
0.68
0.64
0.59
0.61
0.64
0.59
LM
ϕ
2
0.60
0.65
0.65
LB ϕ2
0.62
0.61
0.62
LBM ϕ2
Average
0.70
0.67
0.63
