「反向合意模式」（陳學志、鄭昭明、卓淑玲，2001）是當代幽默理論中，對幽默理解歷程描述最為具體的理論，然而除了閱讀完整則笑話後的主觀評定外，仍缺乏更直接的實徵證據。本研究旨在檢驗反向合意模式的「逆溯推論距離」與「解困與否」對幽默閱讀在眼動軌跡上的差異。當讀者閱讀逆溯推論較多的幽默語句時，會產生較多往返於發笑語句和營造語句的眼動歷程。故而，「回視」和「總凝視時間」會有增加的現象。實驗一將發笑語句操弄成「複雜逆溯推論」、「簡單逆溯推論」和「提供逆溯推論」三種條件，結果發現，「複雜逆溯推論」笑話比「提供逆溯推論」笑話有較多的「回視次數」與「總凝視時間」，表示參與者會經歷「逆溯推論歷程」，且逆溯推論步驟多寡會影響「回視次數」與「總凝視時間」。實驗二操弄發笑語句為「失諧解困」組與「失諧未解困」組。結果發現「失諧未解困」組比「失諧解困」組有較多的「總凝視時間」與「回視次數」，表示參與者會先進行逆溯推論，但並非無止盡的回視，通常在回視一定次數後會選擇放棄。本研究除了支持「反向合意模式」對「逆溯推論歷程」的假設外，更透過眼動歷程表徵不同的幽默處理階段，提供幽默理解的眼動模式。

All jokes contain incongruity between set-up sentences and punch lines. People who read jokes that require more inferences will spend more time resolving the incongruity by making connections between the set-up sentences and the punch line. The opposition-coherence theory of humor (Chen, Cheng, & Cho, 2001) proposes backward inference, defined as the inference directed to parts of text preceding the inference, as an important component of humor processes. The present study examined how backward inference processing affects humor comprehension in two eye-tracking experiments with participants reading jokes at difficult levels. In experiment 1, we examined effects of each of three difficulty levels on humor comprehension. We compared eye movements with provided inference, simple inference, and complex inference jokes with differing punch lines and of varying difficulty. The participants spent a longer time for total viewing and made more regressions for the complex condition than for the provided condition. Results are consistent with the prediction regarding backward inference, suggesting that readers literally revisit aspects of the prior context while apprehending jokes. In experiment 2, we examined effects of resolution on humor comprehension by investigating whether the punch line can resolve the incongruity in incongruity-resolution and incongruity-irresolution conditions. Results showed that participants made more regressions and needed longer total viewing time for the incongruity-irresolution condition than for the incongruity-resolution condition. This finding suggested that participants searched for and took 4 regressions before terminating comprehension. The results mentioned above support the process of backward inferences and provide information about eye movement patterns of joke comprehension.