每天给你送来NLP技术干货！
来自：看个通俗理解吧
有趣大会 · EMNLP2022 (Findings & CL篇)
Interesting · EMNLP2022
(Findings&CL Papers)
一句话了解“有趣大会”
(Explain"Interesting" in One Sentence)
奇妙的角度，有趣的任务
(Fantastic Angles, Interesting Tasks)
趣言趣语提前看：（Preview of Interesting Words）
读懂故事，找到SoulMate（Understand Story, Find SoulMate）
喜剧界AI“王建国” （AI Stand-Up Comedian "WANG Jianguo"）
TA曾说的有所谓，我怕一天一天被忘记（What they once said has meaning, Afraid to be forgotten day by day）
意在笔先，下笔成章（Think Before Write, Smooth and Well）
论文评审加速器（Paper Review Accelerator）
【有趣大会 · 目录】
Findings&CL Papers (←)
Main Conference
Findings
1. 读懂故事，找到SoulMate（Understand Story, Find SoulMate）
MBTI Personality Prediction for Fictional Characters Using Movie Scripts https://arxiv.org/abs/2210.10994
论文配图 (Figure from the paper)：
“墨菲斯（Morpheus），系列电影《黑客帝国》中的角色，由劳伦斯·菲什伯恩饰演。Morpheus, a character from the film series, The Matrix, played by Laurence Fishburne.
”
NLP也许能读懂一个故事，那么它是否也能看穿主人公的人格，为你在异次元找出人格匹配的灵魂伴侣呢？这篇论文的结果告诉我们，对现有的、用于文本分类的深度学习模型来说，并不容易——比随机猜测没有好很多；论文针对此任务提出了一个“多角度（Multi-View）”模型，效果才有了些许提升。这无疑为搭建相关的NLP应用带来了新的启示：如何识别叙事中人物的人格是一个不小的挑战。
NLP may be able to understand a story, but can it also figure out the personality of the characters and find a personality-matching soul mate for you in the story? The results of this paper suggest that it is not easy for existing deep learning models for text classification - not much better than random guessing; the paper proposes a 'Multi-View' model for this task, and the results are only marginally better. This certainly sheds new light on building relevant NLP applications: it is a big challenge to identify the personalities of the characters in a story.
图片来源于网络（Image from the internet）
2. 喜剧界AI“王建国”（AI Stand-Up Comedian "WANG Jianguo"）
A Unified Framework for Pun Generation with Humor Principles https://arxiv.org/abs/2210.13055
论文配图 (Figure from the paper)：
“Pun”可以简单理解为“梗”，或者理解为“谐音、双关梗”。是的，你是不是也想到了他（谐音梗之王——王建国）？这个工作参考了幽默相关的理论，融合了3个角度的指标来构建了一个AI造梗机器。这3个角度分别为：歧义性（ambiguity）、独特性（distinctiveness）和惊喜性（surprise）。论文的工作在英文语料上收获了还不错的效果。但是这个方法是否也适用于汉语，又能否打败王建国还需要实践才能知道结果！
"Pun" is "the usually humorous use of a word in such a way as to suggest two or more of its meanings or the meaning of another word similar in sound". Yes, did you also think of him (the king of pun - WANG Jianguo)? This work draws on theories related to humour and incorporates three perspectives to build an AI Pun Creator. The three perspectives are ambiguity, distinctiveness and surprise. The work gained good results on the English corpus. However, whether this method is also applicable to Chinese and whether it can beat WANG Jianguo needs to be practised to know the results!
3. TA曾说的有所谓，我怕一天一天被忘记（What they once said has meaning, Afraid to be forgotten day by day）
He Said, She Said: Style Transfer for Shifting the Perspective of Dialogues https://arxiv.org/abs/2210.15462
论文配图 (Figure from the paper)：
这里提到的文本风格转换和之前我们了解到的转换可能有所不同。这篇论文解决的问题是视角转换：把一段对话内容从非正式的第一人称视角转变为正式的第三人称视角的文本。这样的复述转变是有很大挑战的，复述的文本既要保留原始的含义，有正确的人称指代，考虑当时的情绪，又要保证是正式的语言风格。也许把生活中的开会、闲聊、吵架(×)等等转变成正式日记记在小本本上是一个大有可为的应用方向。
The textual style shift mentioned here may be different from previous work. The problem addressed in this paper is perspective shifting: transforming the content of a conversation from an informal first-person perspective to a formal third-person perspective text. Such a transformation is challenging: the transformed text has to retain its original meaning, have the correct coreference resolution, take into account the mood of the moment, and also ensure a formal style of language. Perhaps turning discussions in meetings, casual conversations, arguments between lovers (joking), etc. into formal diaries is a promising direction of application.
4. 意在笔先，下笔成章（Think Before Write, Smooth and Well）
Sequentially Controlled Text Generation https://openreview.net/pdf?id=cMGM38rfl-
论文配图 (Figure from the paper)：
相比当前的大模型想到哪里就写到哪里的创作方式，这篇论文主张在写作之前先构思好大体框架再下笔，并且在AI创作过程中，人为可以互动参与。以这样的方式去引导大模型创作，可以生成结构更合理、逻辑更顺畅、主题更鲜明的文章。那……如果我心里有一个还不错的修仙小说大纲，要不让它帮我完成这个“巨作”？
In contrast to the large AI model that outputs the next character until the end of an article, this paper advocates that the general structure of an article is conceived before "writing". Moreover, humans can interactively participate in the AI writing process. Guiding the "writing" of large models in this way can produce more structured, logical and topical articles. So... if I have a decent outline for an Im-Mortal novel in mind, should I let it help me finish the novel?
CL (Computational Linguistics)
5. 论文评审加速器（Paper Review Accelerator）
Revise and Resubmit: An Intertextual Model of Text-based Collaborationin Peer Review https://direct.mit.edu/coli/article/doi/10.1162/coli_a_00455/112555/Revise-and-Resubmit-An-Intertextual-Model-of-Text
论文配图 (Figure from the paper)：
长达几个月甚至1、2年的论文审核时间，对于投稿者来说，可能是担心和焦虑的。论文一天不出结果，心里的石头就会一直放不下。这个工作从审稿者的角度出发，从文学领域的“互文性（Intertextuality）”获得灵感，提出了一套完整的基于NLP的多人审稿协作机制，用以提高审稿的工作效率。“互文性”是一个很大的概念，文中写道：“将文本视为一个由作者创造并由读者在其他文本的背景下进行解释的不断发展的实体，是互文性理论的核心”。它在某种程度上体现了文本含义的传承和演变。作者们希望这个工作能够对现有的审稿流程和工具的改善有所启发。NLP支撑的互动工具带来的高效率审稿对于每一方来说都将带来极大的便利！
Months or even 1 or 2 years of paper review time can be worrying and anxious for the authors. As long as they don't receive the results, they will keep dwelling on the matter. This work proposes a complete NLP-based multi-reviewer collaboration mechanism to improve the efficiency of the review process from the reviewer's perspective, inspired by Intertextuality in literature——"Treating text as an evolving entity created by the author and interpreted by the reader in the context of other texts is the core of the intertextuality theory". It reflects, in a way, the transmission and evolution of the meaning of the text. The authors hope that this work can be an inspiration for improvements to existing review processes and tools. The efficiency of the NLP-backed interactive tool for reviewing drafts will be a great convenience for everyone involved!
结束语 (Next)
下一篇会进入有趣大会的主会篇章~
We will move on to the Interesting · EMNLP2022 (Main Conference) session~
回顾（Previous posts）:
有趣大会 Interesting · ACL2022
有趣大会 Interesting · NeurIPS2022
声明：
内容仅代表个人观点
个人难免会出现遗漏，欢迎小伙伴们私信提出意见或者补充
Statement：
The content represents personal views only
We may have missed some papers, thus greatly appreciate your valuable comments
📝论文解读投稿，让你的文章被更多不同背景、不同方向的人看到，不被石沉大海，或许还能增加不少引用的呦~ 投稿加下面微信备注“投稿”即可。
最近文章
COLING'22 | SelfMix：针对带噪数据集的半监督学习方法
ACMMM 2022 | 首个针对跨语言跨模态检索的噪声鲁棒研究工作
ACM MM 2022 Oral | PRVR: 新的文本到视频跨模态检索子任务
点击这里进群—>加入NLP交流群
778
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
