Page 37 - CyberSecurityEssentialsEnglish

A model does not always treat a whole word as one unit. It may divide a word into smaller
pieces, then analyze their relationship to the context to estimate what comes next.
Tokenization differs across languages, including Arabic and English, which is one reason

performance can vary from one language to another.

If the text reads, “Before sending the report, we should…,” the model may predict words
such as “review” or “verify.” After choosing one token, it predicts the token that follows, and
the process continues until a sentence and paragraph are formed. This is why the answer

may vary from one attempt to another while retaining a similar overall meaning [7].
Answers may vary because generation is probabilistic and because the application may
change the selection method or the instructions surrounding the question. Changing these

settings, however, does not add new knowledge or turn incorrect information into fact.

FIGURE 1-5: HOW AN ANSWER IS GENERATED STEP BY STEP

What to notice: Words are selected according to their linguistic fit with the context. This

does not mean that every claim in the answer is correct or supported by an authoritative
source.
Source: Prepared by the author based on an introductory explanation of language models

[7].

Context and Constraints
An answer depends on the context: the words, instructions, and documents available to the
model during the task. A long document does not guarantee that every detail will receive
equal attention. The model may overlook a condition, confuse two sources, or favor familiar

wording over precise information.

36