情感分析是自然語(yǔ)言處理中一個(gè)非�；钴S的研究領(lǐng)域。近來(lái),互聯(lián)網(wǎng)上出現(xiàn)了許多包含用戶評(píng)論的文本資源:互聯(lián)網(wǎng)用戶的想法、論壇、社交網(wǎng)絡(luò)、消費(fèi)者調(diào)查等�？紤]到數(shù)據(jù)的豐富性,自動(dòng)綜合多個(gè)觀點(diǎn)對(duì)于獲得對(duì)即定主題情感的概述變得至關(guān)重要。該研究對(duì)于希望了解客戶對(duì)其產(chǎn)品的反饋意見(jiàn)的公司,以及希望查詢關(guān)于產(chǎn)品或旅行的評(píng)論的人來(lái)說(shuō)都很有吸引力。在過(guò)去的十年里,推特已經(jīng)變得很流行,并且成為許多人日常生活的一部分。在該論文中,筆者研究了Twitter消息中的諷刺檢測(cè)。盡管大多數(shù)關(guān)于諷刺檢測(cè)的研究都強(qiáng)調(diào)詞匯、句法或語(yǔ)用特征的使用。這些特征通常通過(guò)比喻手段來(lái)表達(dá),如單詞、表情符號(hào)和感嘆號(hào)。在本文中,作者將注意力機(jī)制與深度神經(jīng)模型相結(jié)合,并將其與目前最先進(jìn)的特征工程方法進(jìn)行了比較,探索深度學(xué)習(xí)在諷刺檢測(cè)任務(wù)中的應(yīng)用。作者還建立了一個(gè)關(guān)于諷刺存在的tweet手動(dòng)注釋數(shù)據(jù)集。因?yàn)檫f歸神經(jīng)網(wǎng)絡(luò)（RNN）模型通常不能囊括其最終的所有重要信息隱藏狀態(tài),作者重點(diǎn)關(guān)注機(jī)制的影響,首先通過(guò)與長(zhǎng)期短期記憶（LSTM）結(jié)合探究,然后與雙向長(zhǎng)期短期記憶（BLSTM）結(jié)合探究找出句子中的每個(gè)單詞的相對(duì)作用。結(jié)果表明,基于注意機(jī)制下的LSTM... 

【文章頁(yè)數(shù)】：59 頁(yè)

【學(xué)位級(jí)別】：碩士

【文章目錄】：
摘要
Abstract
1 Introduction
    1.1 Twitter Overview
    1.2 Definition of Sarcasm
    1.3 Why is sarcasm detection interesting?
    1.4 Objectives and Contributions
    1.5 Thesis Structure
2 Related Work
    2.1 Previous Works on Sentiment Analysis
    2.2 Previous Works on Sarcasm Detection
3 Background
    3.1 General Neural Networks
    3.2 Convolutional Neural Network
        3.2.1 Convolution
        3.2.2 Max Pooling
    3.3 Recurrent Neural Networks
        3.3.1 Long Short-Term Memory(LSTM)
    3.4 Training
        3.4.1 Cost Function
        3.4.2 Gradient Descent
        3.4.3 Backpropagation
    3.5 Overfitting
    3.6 Word Embedding
    3.7 Attention Mechanism
4 Proposed Method
    4.1 Data Collection
    4.2 Data Preprocessing
    4.3 Proposed Model Architecture
        4.3.1 Input Layer
        4.3.2 Embedding Layer
        4.3.3 LSTM Layer
        4.3.4 Attention Layer
        4.3.5 Bidirectional Long Short-Term Memory(BLSTM)Layer
5 Data and Experiment Setup
    5.1 Datasets
        5.1.1 Collected Dataset
        5.1.2 Gosh and Veale Dataset
    5.2 Parameter Setting
        5.2.1 Hardware and Software Details
        5.2.2 Hyperparameter
6 Results and Analysis
    6.1 Scoring Methods
        6.1.1 Accuracy
        6.1.2 F1-Score
    6.2 Result on Ghosh and Veale Dataset
    6.3 The Result of Collected Dataset
    6.4 Results Comparison on Different Datasets
    6.5 Evaluation
Conclusion and Future Work
References
Research Projects and Publications in Master Study
Acknowledgements



本文編號(hào)：3731518