JellyBell at TextGraphs-17 Shared Task: Fusing Large Language Models with External Knowledge for Enhanced Question Answering

Julia Belikova, Evegeniy Beliakin, Vasily KonovalovACL 2024

TL;DR

This paper presents JellyBell, a system that achieved 2nd place in the TextGraphs-17 shared task by fusing Large Language Models (LLMs) with Knowledge Graphs (KGs) for question answering. They demonstrate that augmenting LLMs with external knowledge, particularly from Wikidata and web searches, significantly improves performance, especially for less common entities. Their approach also explores effective strategies for combining multiple LLMs with diverse knowledge sources.

What Problem Does This Paper Address?

Answering factual questions accurately can be challenging for Large Language Models (LLMs) alone, as their internal knowledge might be incomplete or outdated. Knowledge Graphs (KGs), on the other hand, provide structured, factual information. The core problem this paper addresses is how to effectively fuse the generative capabilities of LLMs with the structured knowledge of KGs to enhance Question Answering (QA) performance, specifically in the context of selecting the correct KG entity that answers a given question. This is particularly relevant for the TextGraphs-17 shared task, which focuses on this LLM-KG integration.

Methods

The JellyBell system is built upon Meta's Llama 3 LLM, utilizing both 8 billion (8B) and 70 billion (70B) parameter versions. Their approach employs Retrieval-Augmented Generation (RAG) to incorporate external knowledge. This external knowledge comes in three forms:

  • Wikidata ID Description: Simple textual descriptions of answer candidates from Wikidata.
  • Web-search Results: Information retrieved from DuckDuckGo based on the question.
  • Textualized KG: The sub-graphs provided by the task organizers, which represent shortest paths between entities, are converted into a textual format.

A key aspect of their method is Uncertainty Estimation (UE), specifically using "white-box" methods like maximum probability and margin probability. This helps in selecting the most confident answer when combining outputs from different LLM configurations. They also explore ensemble strategies, particularly the "margin probability approach," which selects an answer where the difference between the top two predicted probabilities is maximized, as a way to combine outputs from multiple LLMs.

Datasets

The primary dataset used is the KGQA dataset designed for the TextGraphs-17 shared task. This dataset is based on Wikidata and consists of:

  • Textual questions.
  • Candidate answer options, each being a distinct Wikidata entity.
  • A sub-graph extracted from Wikidata, showing shortest paths connecting entities from the question to the answer candidates.

The training set contains 37,672 samples with 3,535 unique questions, while the test set has 10,961 samples with 1,000 unique questions.

Key Insights and Ideas

  • LLMs partially "know" Wikidata: The experiments suggest that LLMs like Llama 3 have some inherent knowledge about Wikidata, but this knowledge is stronger for well-known entities.
  • External knowledge is crucial, especially for obscure entities: Augmenting LLMs with external data (Wikidata descriptions, web search) significantly boosts QA performance. This benefit is more pronounced when dealing with entities that are not widely known.
  • RAG can sometimes hurt performance: For very popular entities, relying on RAG can be detrimental if the retrieved context is misleading.
  • Model size matters: The larger Llama 3 70B model consistently outperforms the 8B version, even when external knowledge is provided, suggesting that parametric knowledge and reasoning abilities are still critical.
  • Combining knowledge sources requires care: While integrating multiple external knowledge sources generally improves performance, simply adding all of them can sometimes lead to confusion and a decrease in accuracy. The textualized KG, in particular, seemed to have a negative impact when combined with web search and Wikidata descriptions for the Llama 3 70B model.
  • Ensembling with margin probability is effective: The margin probability approach for combining outputs from multiple LLM configurations proved to be the most effective aggregation strategy.

Key Results

  • The JellyBell system achieved 2nd place in the post-evaluation phase of the TextGraphs-17 shared task.
  • When equipped with both web-search results and Wikidata descriptions, Llama 3 models outperformed other combinations of external knowledge sources.
  • The Llama 3 70B model, augmented with web-search results and Wikidata descriptions, showed the best performance.
  • The margin probability ensemble strategy significantly outperformed other aggregation methods when combining outputs from three Llama 3 70B models, each with different knowledge sources.
  • The proposed method significantly outperformed a ChatGPT-based baseline, though it still lagged behind the top-performing system.

Limitations and Future Work

  • Potential for confusion with too much knowledge: The authors observed that incorporating too many diverse knowledge sources could confuse the model and decrease performance.
  • Entity prediction from ID is challenging: Predicting an entity from its Wikidata ID proved to be less accurate than generating an ID from an entity.
  • Future work suggestions:
    • Testing various LLMs at different scales.
  • Exploring different methods for integrating external knowledge.
  • Comparing open LLMs with proprietary LLMs.
  • Examining the multilabel capability of their solution.
  • Conducting a comprehensive error analysis.

Summary updated

BibTeX
@inproceedings{belikova2024jellybell,
  author = {Julia Belikova and Evegeniy Beliakin and Vasily Konovalov},
  title = {{JellyBell at TextGraphs-17 Shared Task: Fusing Large Language Models with External Knowledge for Enhanced Question Answering}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
}
Figure