열린 도메인 환각
Open-domain hallucination is a hallucination where a model confidently generates false information about the world without being anchored to any specific input context.1
Open-domain hallucinations can be classified further:2
- Entity-error Hallucination: This type of hallucination refers to the situations where the generated text of LLMs contains erroneous entities, such as person, date, location, and object, that contradict with the world knowledge. As shown in Table 1, when inquired about “the announcement date of Nokia 3510 phone”, the model generates an erroneous date “October 2002”, in conflict with the real fact “March, 2002”.
- Relation-error Hallucination: This type of hallucination refers to instances where the generated text of LLMs contains wrong relations between entities such as quantitative and chronological relation. As shown in Table 1, the model generates “Among the chemical elements that make up the human body, calcium is more than oxygen” (wrong quantitative relation) and “Aaron Gillespie was born before Nathan Leone” (wrong chronological relation).
- Incompleteness Hallucination: LLMs might exhibit incomplete output when generating lengthy or listed responses. This hallucination arises when LLMs are asked about aggregated facts and they fail to reserve the factual completeness. For example, as presented in Table 1, when inquired “list ten book titles on social cognitive theory”, the model only generates eight book titles; and the statement “in an ecosystem organisms include consumers and producers” is factually incomplete.
- Outdatedness Hallucination: This type of hallucination refers to situations where the generated content of LLMs is outdated for the present moment, but was correct at some point in the past. This phenomenon arises primarily due to the fact that most LLMs were trained on time-limited corpora. For example, when asked about “the present president of the United States”, the model trained on corpora before 2021 will generate “Donald Trump” instead of the latest fact “Joe Biden”.
- Overclaim Hallucination: This type of hallucination refers to cases where the statement expressed in the generated text of LLMs is beyond the scale of factual knowledge (Schlichtkrull et al., 2023). For example, as shown in Table 1, the model generates overclaimed statements “the only way to lose weight is to exercise” and “you can grow taller just by drinking milk”.
- Unverifiability Hallucination: In some cases, the information generated by LLMs cannot be verified by available information sources. For example, LLMs might generate plausible but non-existent academic reading lists. As demonstrated in Table 1, it cannot find any information about the book “Cognitive Foundations of Social Behavior” from existing sources.
Footnotes
-
The dawn after the dark - An empirical study on factuality hallucination in large language models ↩