An Information Bottleneck Perspective for Effective Noise Filtering on Retrieval-Augmented Generation

The information bottleneck principle Tishby et al. (1999) has been a great concept in finding a compression $\tilde{X}$ for signals $X$ that preserves the maximum information relevant to signals $Y$. Given a joint probability distribution $p(X,Y)$ between a random variable $X$ and an observed relevant variable $Y$ (with support sets $x\in\mathcal{X}$ and $y\in\mathcal{Y}$), the amount of information about $Y$ in compressed representation $\tilde{X}$ ($\tilde{x}\in\tilde{\mathcal{X}}$) is given by mutual information:

$$I(\tilde{X};Y)=\int_{\tilde{\mathcal{X}}}\int_{\mathcal{Y}}p(\tilde{x},y)\log\frac{p(\tilde{x},y)}{p(\tilde{x})p(y)}\mathrm{d}\tilde{x}\mathrm{d}y,$$

(1)

where $I(\tilde{X};Y)\leq I(X;Y)$ since compressed data can not convey more information than original ones. The information bottleneck is obtained through

$$\min\mathcal{L}_{\text{IB}}=I(\tilde{X};X)-\beta I(\tilde{X};Y),$$

(2)

where $\beta$ is the Lagrange multiplier for the trade-off between preserving meaningful information and compressing at various resolutions.

Retrieval-augmented generation involves in generating contents conditioned on input queries $q\in Q$, where relevant passages $x\in X$ are retrieved to advance the content generation. As illustrated in Figure 2A, recent noise filtering approaches for retrieval-augmented generation learn to compress retrieved passages $\tilde{x}\in\tilde{X}\subseteq X$ for large language models with a log likelihood objective $\minus\log p_{\text{LM}}(y|[q,\tilde{x}])$ to ground outputs $y\in Y$ Liu et al. (2023); Yang et al. (2023a); Xu et al. (2023); Wang et al. (2023), which are special cases of conditional mutual information $I(\tilde{X},Y|Q)$ and still unable to avoid irrelevant information.

Now we introduce our information bottleneck approach for retrieval-augmented generation and we derive the information bottleneck between retrieved passages $X$ and ground outputs $Y$ conditioned on queries $Q$ given Equation 2. As demonstrated in Figure 2B, the noise filter is required to simultaneously maximizing the mutual information of its compression with ground outputs while minimizing the mutual information with retrieved passages:

$$\min\mathcal{L}_{\text{IB}}=\underbrace{I(\tilde{X},X|Q)}_{\textit{conciseness}}-\beta\underbrace{I(\tilde{X};Y|Q)}_{\textit{correctness}}.$$

(3)

The former term $I(\tilde{X};X|Q)$ serves not only to enhance efficiency, which is a common application, but also to promote conciseness by minimizing the inclusion of irrelevant information. As the filtered information becomes increasingly precise, language models are able to reduce the computational resources allocated to input extraction, thereby enhancing their capacity to concentrate on producing higher-quality contents. It’s worth noting that this term guarantees that the filtered information will be rendered null when the retrieved content is entirely irrelevant to the output.

Then we provide the details of each term in the information bottleneck. The conciseness is:

		$\displaystyle I(\tilde{X};X|Q)=$		(4)
		$\displaystyle\quad\,\,\mathbb{E}_{q}\bigg{[}\int p(\tilde{x},x|q)\log\frac{p(\tilde{x},x|q)}{p(\tilde{x}|q)p(x|q)}\mathrm{d}\tilde{x}\mathrm{d}x\bigg{]}$
		$\displaystyle=\mathbb{E}_{q}\bigg{[}\int p(x|q)p(\tilde{x}|x,q)$
		$\displaystyle\qquad\qquad\log\frac{p(x|\tilde{x},q)\bcancel{p(\tilde{x}|q)}}{p(x|q)\bcancel{p(\tilde{x}|q)}}\mathrm{d}\tilde{x}\mathrm{d}x\bigg{]}$
		$\displaystyle\leq\mathbb{E}_{q}\bigg{[}\int p(x|q)\log\frac{\int p(x|\tilde{x},q)p(\tilde{x}|x,q)\mathrm{d}\tilde{x}}{p(x|q)}\mathrm{d}x\bigg{]}$
		$\displaystyle=-\mathbb{E}_{q}\bigg{[}D_{\text{KL}}\Big{[}p(x|q)||\mathbb{E}_{\tilde{x}\sim p(\tilde{x}|x,q)}p(x|\tilde{x},q)\Big{]}\bigg{]},$

where we get an upper bound of conciseness based on Jensen’s inequality. Therefore, $I(\tilde{X};X,Q)$ can be converted to the form of the Kullback–Leibler divergence between the retrieval probability distribution $p(x|q)$ and the expectation of the probability to recover retrieved passages from compression $p(x|\tilde{x},q)$, where $\tilde{x}$ is required to be integrated over the representation space of noise filter $p(\tilde{x}|x,q)$.

In the scenario of offline retrievers, where the retrieved passages and queries are jointly sampled from training datasets $\{(q,x,y)\}$, $p(x|q)$ becomes a constant number and we can simplify:

$$\min I(\tilde{X};X|Q)\simeq\min\mathbb{E}_{(q,x,\tilde{x})}\big{[}p(x|\tilde{x},q)\big{]}.$$

(5)

Hence, in cases where training the retriever jointly is not necessary, minimizing the conditional mutual information $I(\tilde{X};X|Q)$ can be elucidated as the process of selectively filtering out information to such an extent that it becomes unfeasible to reconstruct the original content, regardless of the strength of generative language models employed.

Next, the correctness is derived as:

		$\displaystyle I(\tilde{X};Y|Q)=H(Y|Q)-H(Y|X,Q)$		(6)
		$\displaystyle=-\int p(y,q)\log p(y|q)\mathrm{d}y\mathrm{d}q-H(Y|X,Q)$

where the former term $\mathbb{E}_{(q,y)}\big{[}\log p(y|q)\big{]}$ is considered as a constant independent of the noise filter. We can ignore this term and simplify as:

		$\displaystyle I(\tilde{X};Y|Q)\simeq-H(Y|\tilde{X},Q)$		(7)
		$\displaystyle=\int p(y,\tilde{x},q)\log p(y|\tilde{x},q)\mathrm{d}y\mathrm{d}\tilde{x}\mathrm{d}q$
		$\displaystyle=\mathbb{E}_{(q,\tilde{x},y)}\big{[}\log p(y|\tilde{x},q)\big{]}.$

When generative language models are fixed, query-answer pairs $\{(q,y)\}$ are sampled from datasets and $\tilde{x}$ is pre-obtained with noise filter. Therefore, maximizing $I(\tilde{X};Y|Q)$ approximates maximizing log likelihood $\log p(y|\tilde{x},q)$, which is explained as the process of selectively retaining as much useful information as possible to enable the language model generate target outputs.

Besides, recent studies on the information bottleneck Fischer (2020); Federici et al. (2020); Lee et al. (2021); Kawaguchi et al. (2023b) suggest to replace $I(X;\tilde{X})$ with $I(X;\tilde{X}|Y)$, because $I(X;\tilde{X})$ can not be zero while maintaining the target-relevant information. Therefore, we follow Federici et al. (2020) and decompose $I(\tilde{X};X|Q)$ into two components by using the chain rule as:

$$I(\tilde{X};X|Q)=I(\tilde{X};X|Y;Q)+I(\tilde{X};Y|Q).$$

(8)

Finally, our information bottleneck for noise filtering in retrieval-augmented generation is:

		$\displaystyle\mathcal{L}_{\text{IB}}=I(\tilde{X};X|Y;Q)-(\beta-1)I(\tilde{X};Y|Q)$		(9)
		$\displaystyle\quad\simeq\mathbb{E}_{(q,x,\tilde{x},y)}\big{[}p(x|\tilde{x},q,y)\big{]}$
		$\displaystyle\qquad-(\beta-1)\ \mathbb{E}_{(q,\tilde{x},y)}\big{[}\log p(y|\tilde{x},q)\big{]},$

where the Lagrange multiplier $\beta-1>0$.

The information bottleneck represents not merely a methodological approach, but a fundamental principle to be applied in retrieval-augmented generation. In this section, we will delineate three distinct applications of information bottlenecks, encompassing the establishment of an evaluation metric for noise filtering, the creation of supervised-fine-tuning (Sft) training datasets, and the formulation of reward functions in reinforcement learning.

Table 1 presents the experimental results on NQ and TriviaQA datasets. Initially, we demonstrate the performance of the generator (Llama2 13B) without retrieval, where a low Em score implies the limitation of large language models facing open-domain question answering task. When the generation is directly augmented with retrieved documents, the result is even worse with a drop of $9.2$ Em score on TriviaQA at most, reflecting the presence of a large amount of noise in the retrieved content and the vulnerability of the generator to noise interference. Besides, the Tfr reveals that noise in retrieved passages will cause over $50\%$ of the generated answers wrong that could have been answered correctly on the NQ dataset.

Extractive noise filters including RankGPT and LongLLMLingua can compensate for the performance losses in Em, F1, and IB scores caused by retrieval, via noise filtering. However, they can hardly outperform the non-retrieval generator, because they enable the generator to answer more questions correctly with $10.3$ of Ffr, but also make more mistakes with $51.0$ of Tfr. Also as an extractive compression method, our approach significantly improves the Em scores on on NQ and TriviaQA by $5.3$ and $2.2$, compared with the non-retrieval generator. For F1 scores, the advancement are $4.5$ and $1.5$ respectively. Our information bottleneck objectives can achieve considerable Ffr improvement while minimizing Tfr performance degradation. For instance, on NQ dataset, we effectively minimize Tfr ($51.0\rightarrow 17.6$) to alleviate the noise interference caused by retrieval, and maintain a comparable Ffr ($10.3\rightarrow 10.2$) to minimize the loss of effective information as much as possible. It’s worth noting that our compression rate achieves $2.5\%$ and $2.6\%$ for these datasets, which impressively reduces irrelevant noise and computational cost. Compared to Llama2-7B, a large language model with knowledge stored in parameters during pretraining stage where we consider it as a powerful abstractive summarization method, our method still outperforms it in Em and F1 scores.

In addition, our proposed information bottleneck metric IB assesses these models from a comprehensive perspective, which takes both conciseness and correctness of compression into account. The IB score of non-retrieval is the basic score for null compression $\tilde{X}=\phi$, which reflects the capability of the language model $p_{\text{LM}}(\cdot)$. The IB score generally aligns with the model’s performance, where models exhibiting lower compression rates and reduced accuracy tend to yield lower scores. Our approach, with the best IB score, achieves the Pareto optimum between compression ratio and performance. We provide case study in Appendix C.

Filtering models encounter more significant hurdles when dealing with multi-hop problems, as solving these requires not only auxiliary information but multiple rounds of reasoning and analysis. As demonstrated in Table 3, the results obtained are marginally improved compared to those achieved through supervised training alone.

We utilize ablation experiments to exhibit the significance of the conciseness term in the information bottleneck theory on NQ. As shown in Table 4, integrating the information bottleneck approach that combines both conciseness and correctness leads to superior outcomes compared to solely utilizing correctness. The filtered outcomes from the former approach are more concise and of superior quality.

To visually demonstrate that applying the information bottleneck to select training data can enhance the upper bound, we conducted experiments on the validation sets of NQ and HotpotQA. Table 5 lists two basic filtering methods: exact search and greedy search. The application of IB selection is not limited to just two filtering methods and it can be generalized to other answer-guided silver mining strategies, such as the leave-one-out evidentiality mining strategy Asai et al. (2022) and Cxmi Wang et al. (2023). Our approach is not fixated on finding the optimal solution in the initial stages, but rather focuses on gradually approaching the optimal filter model through iterative training. Here we choose the two easiest methods, which can greatly reduce computational costs on the construction of training data

The goal of exact search is to find the paragraphs or sentences containing the ground answers. Greedy search Nallapati et al. (2017) is one of the most popular heuristic method by far used in extractive summarization. This algorithm extracts oracle labels with the highest Rouge Lin (2004) scores compared to human-annotated abstracts. We considered two silver summaries, one that concatenates the query and answer, and the other that focuses solely on the answer itself. The former can cover more information, while the latter focuses more on the answer itself. Specially, the answer in intermediate state, supporting facts, are incorporated for multi-hop questions.

By using information bottleneck for selecting among the four filtering results, the obtained outcome is superior to any of them, which reveals that existing simple annotation methods only yield solutions that are far from optimal.

Due to the adoption of extractive filtering method, the content retained after filtering is sourced from the original text. This leads to the hypothesis that there may be a correlation between the compression rate and the conciseness mutual information $I(\tilde{X};X|Y;Q)$. We verify the relationship between compression rate and mutual information on NQ and HotpotQA with a toy experiment based on §5.1. For each query-answer pair with four filtering candidates, we calculate their corresponding length and conciseness $I(\tilde{X};X|Y;Q)$. As the statistical information for different samples is independently and identically distributed, we convert the values of length and conciseness in each sample into their rankings among various compression methods. Then we calculate the Pearson correlation coefficient between length $Rank_{L}$ and conciseness $Rank_{C}$, which is $0.953$, indicating a significant correlation at the 0.01 level (two-tailed).