# Chain of Natural Language Inference for Reducing Large Language Model Ungrounded Hallucinations
[paper](https://arxiv.org/pdf/2310.03951)
Microsoft, 2023

## Introduction

<center>


</center>

## Methodology

<center>


</center>

### Detection agent

#### Step1 : Split and select

**Raw Response Segmentation**

- **Raw responses** are segmented into **individual sentences** using the NLTK sentence splitter.

**Noise Removal**

- Sentences **lacking factual information** or **considered noise** are **purged** from the segmented responses.

**Exception for Short-Generated Responses**

- **Short-generated responses** that can be directly formulated as hypotheses are **exempt** from the purging process for benchmark comparison purposes.

**Future Work**

- There's a plan to develop an **advanced hypothesis selector** as future work.

**Resulting Set**

- After the **segmentation** and **filtering** steps, a set of selected hypotheses, referred to as **Hselected**, is obtained.

#### Step2 : Sentence-level detection

**Judge each hypothesis independently against the corresponding premise**

- $X$ denote the input source text
- $Y_{raw}$ denote the output raw response
- $hyp_{i}$ denote selected hypotheses from $Y_{raw}$
- $Entailment: X \Rightarrow hyp_{i}$
- $Contradiction: X \Rightarrow ¬hyp_{i}$
- $Neutral: X \nRightarrow hyp_{i}$

They utilize **CoT prompting**, **guiding** the LLM to locate relevant passages in the **source text X** and allow it to **reason** and then make a **conclusion**.

**Detection agent prompt**

- **System instruction**
- Definition of noun
- ex : hypothesis, Entailment
- How to think step by step for each judgement
- ex : Repeat hypothesis you are judging
- Rules
- ex : You may assume that today is March 24th, 2023
- **First few shot example**
- User
- Single Premise
- Multiple Hypothesis
- Assistant
- Answer with **reason** and **conclusion** for each Hypothesis
- **Second few shot example**
- User
- Single Premise
- Multiple Hypothesis with **tagged words**
- Assistant
- Answer with **reason** and **conclusion** for each Hypothesis
- **Current request**
- Single Premise(**Source Text**)
- Multiple Hypothesis(**$hyp_{i}$**)

#### Step3 : Entity-level detection

- Leverage an **entity recognition model (NER)** to find entities in the nonhallucinated hypothesis
- This is because that if a hypothesis contains **abundant factual details** or **some details require complex reasoning** against the source text, sentence-level detection may reach false negative conclusions.

##### Entity category definition

- **PERSON**: Names of people.
- **PERSONTYPE**: Job types or roles held by a person
- **LOCATION**: Natural and human-made landmarks, structures, geographical features, and geopolitical entities.
- **EVENT**: Historical, social, and naturally occurring events.
- **SKILL**: A capability, skill, or expertise.
- **DATETIME-DATERANGE**: Date ranges.
- **DATETIME-DURATION**: Durations.
- **QUANTITY-NUMBER**: Numbers.
- **QUANTITY-CURRENCY**: Currencies

#### Step4 : Merging

- A hypothesis will be judged as **non-hallucination** only if overall **sentence judgment** and **tagged entities judgments** all vote for non-hallucination.

### Mitigation agent

- Now, we have **reason** and **conclusion** of **Hallucinations Hypothesis**, and also **Source Texts** and **Raw Response**.
- We use these as **instructions** to **rewrite Raw Response**.

#### Mitigation agent prompt

- **System instruction**
- Explain the task
- **Current request**
- Source Text
- Raw Response
- Reason and Conclusion
- Explain writing detail

## Experiment

### Hallucination detection experiments

#### Datasets

<center>


</center>

- Datasets with **synthetic hallucination** generated on **ground truth response text**
- Datasets with hallucination annotated **manually** on real state-of-theart (SOTA) **NLG model output** response text

#### Result

##### Synthetic hallucination dataset results

<center>


</center>

- Dark green : No.1
- Green : No.2
- We can see that CoNLI achieves good result

##### Annotated hallucination dataset results

<center>


</center>

- Dark green : No.1
- Green : No.2
- We can see that CoNLI achieves good result

##### Ablation study

<center>


</center>

- Using both **Sentence-level detection** and **Entity-level detection** get better result on most dataset

### Hallucination reduction experiments

#### Result

<center>


</center>

- **Refined Response is better** than Raw Response on most setting

## Conclusion

- **Detect** and **reduce** ungrounded hallucinations in a **plug-and-play** manner
- Propose a simple yet effective **LLM-based** framework that formulates **hallucination detection** into a chain of **NLI tasks**
- Importantly, its **interpretable output** can also be **leveraged for hallucination reduction**.