C
⁢

S
⁡
(

a
,
b

)

=

<

a
⁢

❘
"[LeftBracketingBar]"

b
>

❘
"[LeftBracketingBar]"

a
⁢

❘
"[LeftBracketingBar]"

❘
"[LeftBracketingBar]"

b

❘
"[RightBracketingBar]"

=

cos
⁢

(

θ

a
-
b

)

(
3
)

Since the basis of the text vectors represent the unigrams and bigrams of the actual text, mathematical similarity between two text vectors indicates textual similarity. Numbers close to 0 indicate that the two text vectors are aligned and thus, have overlapping text tokens. Numbers close to 1 indicate that the two text vectors are perpendicular and thus, have different tokens. Several examples of the comparison to ‘the Earth revolves around the Sun’ are given in Table 2.

TABLE 2

Sample Text and Quantitative Similarity Measurements

Text
Similarity

The Earth revolves around the Sun
0 (Same)

The Earth does not orbit the Sun
0.5 (Medium Similarity)

Orange is my favorite color
1 (No Similarity)

The values included in Table 2 represent the behavior of a cosine similarity operator. As seen in Table 2, identical text has a value of 0 while text that is somewhat similar has a more intermediate value. Of interest, are values that are close to 1. In these cases, the text has no similarity, which indicates the discussion of two different topics as seen in the final entry in Table 2. Additional studies of the training data indicated that dissenting text falls within intermediate similarity values. In order to adjust the behavior of the similarity operator to better model this factual disagreement, the output of the cosine similarity metric between a predefined true text (T) and a given piece of text (t) is used as the input of a Gaussian distribution 540, which gives a Scaled Cosine Similarity (SCS) as seen in Equation 4. This distribution then scales a similarity metric 520 most when it is near the mean of the distribution and the least when it is far away from the mean, as seen in 540.

S
⁢
C
⁢

S
⁡
(

t
,
T

)

=

β
⁢

exp
⁢

(

-

(

C
⁢

S
⁡
(

t
,
T

)

-
μ

σ

)

2

)

(
4
)

The mean and standard deviation of this Gaussian distribution 540 serve as a way to adjust the similarity metric's behavior. It is desirable to adjust the mean of the Gaussian distribution 540 such that it coincides with the cosine similarity metric values of text which contains misinformation. The standard deviation serves as a way to build in a tolerance window for misinformation that has a cosine similarity metric 540 that does not directly coincide with the mean of the Gaussian distribution. For instance, a wider standard deviation would result in a wider range of cosine similarity measurements that indicate misinformation. Due to the topic-dependent nature of misinformation risk assessment, the mean and standard deviation of this distribution are fit to a set of training data—this process will be discussed later. As seen at the rightmost region of the Gaussian distribution 540, text vectors with completely different meaning may or may not be scaled the least so as to not penalize unrelated text.

Similarly, a sentiment value or sentiment values 510 of the text vectors or batch of text vectors are scaled 530 using a linear function as seen in Equation 5, which scales negative sentiment values 510 to a degree corresponding linearly to the magnitude of the sentiment. For instance, strongly negative text has been shown to have a higher chance of containing misinformation than weakly negative text. This relationship is expressed in the sentiment weighting distribution (SWD) 530 seen in Equation 5.