A data science team is implementing a Natural Language Generation (NLG) system to create summaries of complex legal documents. They are using a state-of-the-art transformer-based large language model. A critical issue they face is "hallucination," where the model generates plausible but factually incorrect information not present in the source text. To address this issue while maintaining high linguistic quality, which of the following strategies is the most appropriate to implement?
Implementing a retrieval-augmented generation (RAG) architecture to ground the model's output in the source documents.
Switching from subword tokenization to a character-level model to capture finer grammatical details.
Increasing the generation temperature to encourage more novel sentence structures.
Fine-tuning the model on a much larger, general-purpose text corpus to improve its world knowledge.
The correct answer is to implement a retrieval-augmented generation (RAG) architecture. RAG is specifically designed to reduce hallucination by grounding the language model's output in external knowledge. It works by first retrieving relevant information from a trusted source (in this case, the legal documents) and then using that retrieved context to generate the summary, which significantly improves factual accuracy.
Incorrect options are:
Increasing the generation temperature would make the output more random and creative, which is likely to increase the frequency of hallucinations rather than reduce them.
Fine-tuning the model on a large, general-purpose text corpus would not solve the problem of factuality related to specific legal documents and could even introduce more non-domain-specific information, potentially worsening hallucinations.
Switching from subword tokenization to a character-level model addresses how text is broken down for the model to process, not the semantic problem of ensuring the generated content is factually grounded. This architectural change is irrelevant to solving the hallucination issue.
Ask Bash
Bash is our AI bot, trained to help you pass your exam. AI Generated Content may display inaccurate information, always double-check anything important.
What is a Retrieval-Augmented Generation (RAG) architecture?
Open an interactive chat with Bash
Why does increasing the generation temperature increase hallucinations?
Open an interactive chat with Bash
How does fine-tuning on a domain-specific corpus reduce hallucinations?