Lift Yourself Up: Retrieval-augmented Text Generation with Self Memory

With direct access to human-written reference as memory, retrieval-augmented generation has achieved much progress in a wide range of text generation tasks. Since better memory would typically prompt better generation (we define this as primal problem). The traditional approach for memory retrieval involves selecting memory that exhibits the highest similarity to the input. However, this method is constrained by the quality of the fixed corpus from which memory is retrieved. In this paper, by exploring the duality of the primal problem: better generation also prompts better memory, we propose a novel framework, selfmem, which addresses this limitation by iteratively employing a retrieval-augmented generator to create an unbounded memory pool and using a memory selector to choose one output as memory for the subsequent generation round. This enables the model to leverage its own output, referred to as self-memory, for improved generation. We evaluate the effectiveness of selfmem on three distinct text generation tasks: neural machine translation, abstractive text summarization, and dialogue generation, under two generation paradigms: fine-tuned small model and few-shot LLM. Our approach achieves state-of-the-art results in four directions in JRC-Acquis, XSum (50.3 ROUGE-1), and BigPatent (62.9 ROUGE-1), demonstrating the potential of self-memory in enhancing retrieval-augmented generation models. Furthermore, we conduct thorough analyses of each component in the selfmem framework to identify bottlenecks and provide insights for future research.

Two Important Things

1. Addressing Corpus Quality

Note: I did not read this paper too closely, as I found it written in a rather confusing manner. It is likely that some/much of my interpretation of its points may not be faithful to the author’s intentions.

RAG pipelines suffer from a fundamental problem where the quality of generation is limited by the quality of the RAG dataset (which the paper refers to as “memory”).

The observation by the authors is that a RAG model can improve its output quality by first synthesizing the retrieved context to generate new content, that is added back into its knowledge base. The generated outputs are more in-distribution with what the model might see at query time, therefore resulting in better retrieval performance and final outputs.

They define a primal and dual formulation for RAG:

Primal problem: better memory prompts better generation. This means better data and retrieved context leads to better outputs.
Dual problem: better generation also prompts better memory. This means that when RAG systems can produce good outputs, it can also augment its own knowledge store with higher-quality data.

To utilize this insight, they developed a system called Selfmem

2. `Selfmem`

The framework looks like the following:

The two components are the retrieval-augmented generator, and the memory selector.

To my undersatnding, the algorithm works as follows:

Get query
Retrieve relevant chunks in dataset with a retriever
Use LLM to generate pool of synthetic candidates based on the chunks
Train a selector on the candidate pool using a selection metric, so the selector is now good at selecting good chunks to optimize the metric
Using the trained selector, iteratively sample from the candidate pool, and use it to generate further candidates which are added back into the pool, until when metric converges for a validation set
Generate final output with the best data chunk

Most Glaring Deficiency

The paper was not really written very clearly, and I personally found it really difficult to understand even though I believe the main points that the authors were trying to make are not that complicated.

Conclusions for Future Work

A RAG system can benefit from higher-quality output by bootstrapping itself by generating more high-quality data using data synthesized from its own data store.