When Giant Language Brains Just Aren't Enough! Domain Pizzazz with Knowledge Sparkle Dust

Minh-Tien Nguyen, Duy-Hung Nguyen, Shahab Sabahi, Hung Le, Jeff Yang, Hajime HottaarXiv 2023 · OpenReview

TL;DR

Large Language Models (LLMs) like GPT-3.5 show impressive capabilities but struggle with specialized business domains. This paper demonstrates that augmenting LLMs with domain-specific knowledge, such as insurance policy rulebooks and external knowledge bases like DBpedia, significantly improves their accuracy and reasoning abilities in complex question-answering tasks.

What Problem Does This Paper Address?

While LLMs have revolutionized natural language processing, adapting them to real-world business scenarios, especially those requiring deep domain expertise and complex reasoning, remains a significant challenge. The authors focus on the insurance industry, a domain characterized by intricate policies, regulations, and legal jargon. Standard LLMs, trained on general data, often lack the nuanced understanding needed to accurately answer specific insurance-related questions, leading to generic responses or even hallucinations. This gap hinders their practical application in customer-facing roles where precision and reliability are paramount.

Methods

The core of the proposed solution involves enhancing LLMs with domain-specific knowledge. The authors introduce a model that leverages LLMs (specifically GPT-3.5-Turbo) by feeding them with relevant context extracted from two sources:

  1. Insurance Policy Rulebooks: These documents contain the definitions, terms, policies, and regulations specific to insurance. The raw text from these rulebooks is preprocessed, segmented into manageable chunks, and converted into vectors for information retrieval.
  2. DBpedia: This is a public knowledge graph that provides general, external knowledge. It's used to enrich the context when questions involve concepts or entities not explicitly found in the rulebooks.

The process involves a pipeline where user questions are first processed by an "adapter" that retrieves relevant context from both the rulebooks and DBpedia. This retrieved context is then combined with the original question to form a prompt that is fed into the LLM. The authors also highlight the importance of table processing, where tabular data within rulebooks is converted into a sequence format to improve LLM comprehension.

Datasets

The paper uses two datasets for evaluation:

  1. Expert-created dataset: This dataset consists of 104 gold question-answer pairs directly from an insurance company, covering various aspects of insurance policies and benefits.
  2. Synthesized dataset: This dataset was created by combining commonsense knowledge from DBpedia with domain-specific knowledge from rulebooks. It was generated by GPT itself, with human checking to ensure quality. This dataset contains 87 question-answer pairs and aims to capture questions that might not explicitly mention concepts found in rulebooks.

Key Insights and Ideas

  • LLMs are not standalone solutions: For complex, domain-specific tasks, LLMs require augmentation with external knowledge to perform effectively.
  • Domain-specific knowledge is crucial: Insurance policy rulebooks provide essential context that general LLM training data lacks, leading to more accurate answers.
  • External knowledge bases complement domain data: DBpedia helps bridge gaps when rulebooks are insufficient, especially for concepts not explicitly defined within the insurance context.
  • Table processing matters: Converting tabular data into a sequence format significantly improves LLM performance on questions related to structured information.
  • Information Retrieval (IR) is key: The effectiveness of the knowledge enhancement heavily relies on the quality of the context retrieved from the knowledge sources.

Key Results

  • Significant Accuracy Improvement: Knowledge enhancement dramatically improved GPT-3.5's performance. On the expert-created dataset, accuracy rose from 9.60% (context-agnostic) to 61.42% (GPT+RB) and 65.40% (GPT+RB+DBpedia). On the synthesized dataset, accuracy increased from 25.30% to 55.42% (GPT+RB) and a remarkable 83.13% (GPT+RB+DBpedia).
  • External Knowledge Impact: DBpedia played a more critical role in the synthesized dataset (83.13% accuracy) compared to the expert-created dataset (65.40% accuracy), suggesting its value in handling broader or less explicitly defined queries.
  • Table Processing Efficacy: Using table processing improved GPT's accuracy on questions involving tabular data from 7 correct answers to 16 correct answers (80%) on the expert-created dataset.
  • Reasoning Capability: The GPT+RB+DBpedia model demonstrated a better ability to answer questions requiring multi-step reasoning and understanding of specific insurance terms.

Limitations and Future Work

  • Information Retrieval (IR) Limitations: The current IR module uses a simple Cosine similarity method, which can lead to incorrect context retrieval. Future work should focus on improving IR with more sophisticated similarity and vectorization methods.
  • Ambiguity and Complexity: The expert-created dataset presented challenges due to ambiguity and complex questions requiring multi-step reasoning, which even enhanced LLMs struggled with.
  • Synthesized Dataset Creation: The synthesized dataset, while comprehensive, was generated by GPT, potentially introducing biases or simplifying reasoning processes.
  • Future Directions: The authors suggest exploring knowledge graphs for both open and closed domains, using chain-of-thought prompts to enhance reasoning, and further improving the IR component.

Summary updated

Abstract

Large language models (LLMs) have significantly advanced the field of natural language processing, with GPT models at the forefront. While their remarkable performance spans a range of tasks, adapting LLMs for real-world business scenarios still poses challenges warranting further investigation. This paper presents an empirical analysis aimed at bridging the gap in adapting LLMs to practical use cases. To do that, we select the question answering (QA) task of insurance as a case study due to its challenge of reasoning. Based on the task we design a new model relied on LLMs which are empowered by additional knowledge extracted from insurance policy rulebooks and DBpedia. The additional knowledge helps LLMs to understand new concepts of insurance for domain adaptation. Preliminary results on two QA datasets show that knowledge enhancement significantly improves the reasoning ability of GPT-3.5 (55.80% and 57.83% in terms of accuracy). The analysis also indicates that existing public knowledge bases, e.g., DBPedia is beneficial for knowledge enhancement. Our findings reveal that the inherent complexity of business scenarios often necessitates the incorporation of domain-specific knowledge and external resources for effective problem-solving.
BibTeX
@misc{nguyen2023when,
  author = {Minh-Tien Nguyen and Duy-Hung Nguyen and Shahab Sabahi and Hung Le and Jeff Yang and Hajime Hotta},
  title = {{When Giant Language Brains Just Aren't Enough! Domain Pizzazz with Knowledge Sparkle Dust}},
  year = {2023},
  eprint = {2305.07230},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2305.07230},
}
Figure