Microsoft Word turney-littman-acm doc

Neighbourhood Size
A
c
c
u
ra
c
y
100% Threshold
75% Threshold
50% Threshold
25% Threshold
Figure 6. SO-PMI with varying neighbourhoods, using TASA and the GI lexicon. 
With AltaVista, we can use the AND operator instead of the NEAR operator, which 
increases the neighbourhood size from ten words to a whole document. Figure 7 is a 
graph of accuracy as a function of the percentage of the test set that is classified 
(threshold), using AV-ENG and the GI lexicon. With the whole test set, NEAR is clearly 
superior to AND, but the gap closes as the threshold decreases. This is not surprising, 
since, as the threshold decreases, the selected testing words have increasingly high 
confidences. That is, the absolute values of the semantic orientations of the remaining 
words grow increasingly large. The words with a very strong semantic orientation (high 
absolute value) do not need the extra sensitivity of NEAR; they are easily classified using 
the less sensitive AND operator. 

---

23
0
10
20
30
40
50
60
70
80
90
100
0
10
20
30
40
50
60
70
80
90
100
Threshold
A
c
c
u
ra
c
y
NEAR
AND
Figure 7. AND versus NEAR with AV-ENG and the GI lexicon. 
Figure 8, with AV-CA (2 × 10
9
words), displays the same general pattern as Figure 7, 
with AV-ENG (1 × 10
11
words). However, on the smaller corpus, AND is superior to 
NEAR for the words with the strongest semantic orientation (threshold below 10%). The 
smaller corpus shows more clearly the trade-off between the greater sensitivity of a small 
neighbourhood (with NEAR) and the greater resistance to noise of a large neighbourhood 
(with AND).
0
10
20
30
40
50
60
70
80
90
100
0
10
20
30
40
50
60
70
80
90
100