《2020年度最佳碩士論文獎》

林宜儒 （國立台灣大學語言學研究所）

論文題目 「以計算與認知觀點預測分析梗圖裡的黑色幽默」

指導教授: 謝舒凱教授

中文摘要

本論文主要核心內容為運用神經網路模型 (neural network)來預測網路迷因(internet memes) 的黑色幽默程度，並藉用認知語言學的觀點來解釋模型結果。根據本研究的定義，地獄哏 (hellish gags) 梗圖常用來傳達“黑色幽默”的效果，指的是以他人或是自身的殘疾、悲劇或其他足以引人悲傷、憤怒的事情做為幽默、笑點的梗圖。不過，卻也時常因文本呈現的手法太過獨特，使閱聽者「忍不住覺得好笑」，游移於「道德」與「不道德」間的模糊地帶。本研究將黑色幽默視為一個連續體 (continuum) 現象，最極致的一端為已失去幽默效果的仇恨言論，是純粹以攻擊個人或團體為目的的言論，例如:性別、人種、宗教、族群、殘疾或性取向等。透過這樣的標準，我們運用神經網路模型來預測一個梗圖裡的黑色幽默程度，若我們的模型顯示黑色幽默程度過高，極有可能暗示它是已失去幽默效果的冒犯性梗圖 (offensive meme)。

研究結果顯示，透過運用梗圖裡的字幕、圖片、及模板名稱，我們的模型可以達到 64% 的預測效果。並且，多模態 (multimodal) 預測模型也顯示字幕及模板名稱對於預測黑色幽默有最好的效果。此外，由於幽默涉及人類認知經驗，因此本研究借用認知語言學文獻所提出理解梗圖所需的四種能力，比較人類與機器在理解梗圖歷程上的差異。若能將人類理解梗圖幽默所運用的知識轉化為機器學習的特徵，相信會對機器在理解幽默上有所幫助。總結而言，本研究主要貢獻有二: 一是透過計算模型來自動篩選出冒犯性梗圖; 二是結合認知上的觀點來解釋、分析神經網路模型的結果，希望藉此增進對於黑色幽默梗圖的認識。

英文摘要

The main goal of this thesis is to predict dark humor intensity within internet memes (IMs) and fill in the missing pieces in the puzzle of machine understanding towards dark humor by offering cognitive interpretations for computational results. Dark humor meme, or “hellish gag” meme, is a type of IMs that teases, ridicules, or takes others’ misfortune as jokes. These jokes are often malicious, but sometimes make readers laugh. When this form of dark humor is taken to extreme, it may become purely offensive, insulting or in a form of hate speech. In this thesis, this continuum of dark humor, ranging from humor to hatred, is analyzed and studied within the context of IMs to predict the dark humor level.

With the growing volume of multimodal social media, detecting offensive content on online social media is still an ongoing struggle. IMs, as a relatively complex, multi-layered combination of image and text, have doubled the challenge. Therefore, one of the aims of this study is to build an automatic filter to improper memes on social media by measuring the dark humor level of IMs. If the meme scores very high on our dark humor scale, it is highly possible that the meme is considered as an offensive meme by our machine system. Additionally, as humor is always situated in a broader context that sometimes requires a lot of external knowledge to fully understand it, insights from different perspectives could provide a deeper understanding of memes. In this thesis, we prove that insights from cognitive linguistics help identify several kinds of processing difficulties in understanding memes, which provides practical implications for improving the performance of machines in the future.

To compute dark humor in IMs, this thesis trains a neural network that can not only separate dark humor memes from safe humor memes but also predict dark humor score of memes. We evaluate our model on predicting dark humor from internet memes, using the captions, template names, and image features. We achieve a 64% prediction on predicting dark humor level and show that among all the modalities investigated in this study, the captions and template names act as the strongest predictor for predicting dark humor in memes. A cognitive explanation of computational results is also presented to determine how machines can improve the understanding of the level of dark humor by gaining insights from how humans process dark humor in memes. To conclude, our study not only sheds light on the continuous and subjective nature of dark humor but also provides practical implications on online hate speech detection.

論文貢獻

本論文運用神經網路模型預測網路迷因的黑色幽默程度。多模態模型透過梗圖裡的字幕、圖片、及模板名稱等特徵，辨識出已失去幽默效果的冒犯性梗圖，結果顯示模型預測效果達到64%。另外，本文結合認知上的觀點來解釋神經網路模型的結果，並且進一步分析文字及圖像在黑色幽默理解中的互動關係。我們發現，網路迷因是一種特殊的構式 (construction)，需要多模態的理解。而人類理解黑色幽默梗圖的歷程對於增進機器理解黑色幽默有極大幫助，有助於黑色幽默的認知處理。