LMの出力の低確信度は知識の欠如を示す傾向がある
Language Models (Mostly) Know What They Know (Kadavath et al., 2022)
● GPT-3は自身の知識の程度をある程度判別できる
How Can We Know When Language Models Know? On the Calibration of Language Models for Question
Answering (Jiang et al., 2021)
● (モデルによっては)確信度と正答率の間に相関が見られる
Detecting Hallucinated Content in Conditional Neural Sequence Generation (Zhou et al., 2021)
● ハルシネーションの内容は入力から予測が困難であると考えられる(という言及)
Can Open-Domain QA Reader Utilize External Knowledge Efficiently like Humans? (Varshney et al., 2022)
● 検索を使わないモデルで低確信度の出力が得られた際、検索を使うモデルに切り替えて生成すると好成績
検索を行うタイミングはLMが生成したトークンの確信度が低い時とすればよさそう
FLARE(direct)
6
Slide 7
Slide 7 text
Joe Biden attended the University of Pennsylvania, where he earned a law degree.
赤い部分の確信度が閾値θを下回っている→この文をクエリにして検索
Joe Biden attended [MASK], where he earned [MASK].
クエリとして使う際、確信度が低い (<β)トークンをマスクしてクエリとする。
マスクしたことがわかるトークンに置換している、と思っていたが、
Fig. 3 を見ると単にそのトークンを削除しているっぽい?
FLARE(direct)
7
Skill 1のinstruction例
Question: But what are the risks during production of nanomaterials?
Answer (with Search): [Search(nanomaterial production risks)] Some
nanomaterials may give rise to various kinds of lung damage.
Question: The colors on the flag of Ghana have the following
meanings.
Answer (with Search): Red is for [Search(Ghana flag red meaning)]
the blood of martyrs, green for forests, and gold for mineral wealth.
Skill 2のinstruction例
Question: When did the director of film Hypocrite (Film) die?
Answer (with step-by-step): The film Hypocrite was directed by Miguel
Morayta. Miguel Morayta died on 19 June 2013. So the answer is 19
June 2013.
FLARE(instruct)
10