Entity2Intent：一個用於歧義感知意圖分類的實體引導推理框架

Abstract

大型語言模型 (LLM) 憑藉其泛化能力，在意圖分類任務中展現出強大的表現。然而，在語義模糊的環境中，它們的有效性會下降，因為許多意圖標籤的含義緊密相關或重疊。這些挑戰在細粒度分類任務中尤其明顯，因為意圖之間的細微差別會導致邊界模糊和頻繁的誤分類。我們還觀察到，增加意圖類別的數量會導致準確率下降，這是由緊密相關的意圖標籤之間的語義重疊和邊界模糊造成的。在這種模糊場景下，基於提示的通用方法往往無法發揮作用，因為它們往往依賴淺層的詞彙線索，難以消除緊密相關的意圖的歧義。 為了克服這些限制，我們提出了 Entity2Intent，這是一種新穎的實體引導推理框架，它引入了對使用者查詢的結構化解釋，從而能夠在語義重疊的標籤空間中實現更準確的意圖分類。 我們在三個廣泛使用的意圖分類基準資料集（Banking77、Clinc150 和 LIU54）上進行了實驗，以評估我們提出的 Entity2Intent 框架的有效性。與 Zero-Shot、Few-Shot 和 PC-CoT 等強基線相比，我們的方法始終保持更高的準確率，在每種設定下，與最佳基線相比，平均提升高達 +0.51。

Large Language Models (LLMs) have demonstrated strong performance in intent classification tasks due to their generalization capabilities. However, their effectiveness declines in semantically ambiguous settings, where many intent labels are closely related or overlapping in meaning. These challenges are especially pronounced in fine-grained classification tasks, where subtle distinctions between intents give rise to boundary ambiguity and frequent misclassification. We also observed that increasing the number of intent classes leads to lower accuracy, driven by semantic overlap and boundary ambiguity among closely related intent labels. In such ambiguous scenarios, general prompt-based methods often fall short, as they tend to rely on shallow lexical cues and struggle to disambiguate closely related intents. To overcome these limitations, we propose Entity2Intent, a novel entity-guided reasoning framework that introduces structured interpretation of user queries, enabling more accurate intent classification in semantically overlapping label spaces. We conducted experiments on three widely used intent classification benchmark datasets, Banking77, Clinc150, and LIU54, to evaluate the effectiveness of our proposed Entity2Intent framework. Compared to strong baselines such as Zero-Shot, Few-Shot, and PC-CoT, our method consistently achieves higher accuracy, with average improvements of up to +0.51% (GPT-3.5-Turbo) and +3.56% (LLaMA2-70B-Chat) over the best baseline in each setting.