TLDR: This research investigates how Large Language Models (LLMs) handle ambiguous Chinese text. Researchers created a benchmark dataset of 900 ambiguous sentences categorized into lexical, syntactic, and semantic-pragmatic types. Experiments showed that LLMs struggle to reliably detect ambiguity, often overconfidently interpret ambiguous text as having a single meaning, and “overthink” when trying to understand multiple meanings. The study highlights a fundamental limitation in current LLMs, especially for real-world applications where linguistic ambiguity is common, and suggests that methods like Retrieval-Augmented Generation (RAG) can help improve performance.
Large Language Models (LLMs) have become incredibly powerful, demonstrating impressive language understanding capabilities and being widely used in various real-world applications, from processing complex instructions in multi-turn dialogues to acting as AI agents. However, despite their advancements, LLMs still face significant challenges, particularly concerning their trustworthiness. Issues like hallucinations, misunderstandings, and misalignments are critical, especially in safety-sensitive scenarios.
A recent study delves into a crucial aspect of LLM trustworthiness: how these models behave when encountering ambiguous narrative text, with a specific focus on Chinese textual ambiguity. Ambiguity is a common and inherent part of human language, frequently appearing in everyday interactions, including those with AI systems. For instance, an instruction like “return the phone and computer accessories I purchased last month” can have multiple interpretations: does it mean returning the phone and the computer’s accessories, or accessories for both the phone and the computer? An intelligent agent should be able to resolve such ambiguities rather than proceeding with a single, potentially incorrect, interpretation.
To investigate this, researchers developed a new benchmark dataset specifically for ambiguity detection and interpretation in Chinese text. This dataset comprises 900 ambiguous sentences collected and generated from real-world contexts. Each ambiguous sentence is meticulously annotated by native Chinese speakers with all plausible interpretations and corresponding disambiguated versions, where each rewritten sentence clearly reflects one possible meaning. The annotated examples are systematically categorized into three main types: lexical, syntactic, and semantic-pragmatic, with nine further subcategories.
Experiments conducted with a range of open-weight LLMs, including models from the Qwen3 series, Gemma2 series, and DeepSeek-R1, revealed significant fragility in their ability to handle ambiguity. The findings highlight behavior that substantially differs from human understanding. Specifically, LLMs struggle to reliably distinguish ambiguous text from unambiguous text, often showing overconfidence in interpreting ambiguous text as having only a single meaning rather than multiple possibilities. Furthermore, when prompted to understand various possible meanings, they sometimes exhibit “overthinking,” producing unnecessarily complex or speculative explanations.
The study explored three core experimental tasks: ambiguity detection (binary classification to determine if a sentence is ambiguous), ambiguity understanding (identifying ambiguity sources, generating multiple interpretations, and creating disambiguated sentences), and an end-to-end task combining both detection and understanding. Evaluation metrics included accuracy, precision, recall, and F1 score, with a particular emphasis on F1 and recall due to the imbalanced distribution of ambiguous sentences in real-world data.
Different prompting strategies were tested to see their impact on LLM performance. These included direct prompting, few-shot prompting (providing examples), knowledge-enhanced prompting (incorporating linguistic background), Chain-of-Thought (CoT) prompting (step-by-step analysis), and combinations of these. Notably, Retrieval-Augmented Generation (RAG) combined with few-shot prompting proved to be the most effective approach for improving both ambiguity detection and understanding. RAG helps models by retrieving relevant examples to guide their reasoning, addressing issues like selecting only one interpretation or over-interpreting due to lack of context.
The research underscores that while larger models generally perform better, and reasoning-enhanced models show improved capabilities, the RAG method significantly boosts sensitivity to Chinese ambiguity, especially for medium-sized non-reasoning models. For models with strong inherent reasoning, RAG provides modest improvements, as they rely more on internal logic. The study also found that perplexity scores, a measure of a language model’s certainty, may not be a reliable indicator of an LLM’s ambiguity understanding ability.
Also Read:
- Evaluating LLMs: Why Different Voices Matter in Benchmarking
- Unlocking Smarter AI Responses: The SemRAG Approach to Knowledge Integration
This work provides a novel perspective on the trustworthiness of LLMs and serves as a call for the research community to address this inherent limitation. The findings have significant implications for the deployment of LLMs in real-world applications where linguistic ambiguity is common, urging caution and the development of improved approaches to handle uncertainty in language understanding. The dataset and code from this research are publicly available, fostering further advancements in this critical area. You can find more details in the full research paper: Uncovering the Fragility of Trustworthy LLMs through Chinese Textual Ambiguity.
