【摘要】：互聯(lián)網(wǎng)尤其是移動(dòng)互聯(lián)網(wǎng)的快速發(fā)展使其充斥著大量的帶有情感的短文本,挖掘這些文本包含的情感信息可以獲取大量的商業(yè)信息和社會(huì)信息。本文從兩個(gè)方面使用深度學(xué)習(xí)算法進(jìn)行文本情感分析研究。第一,有效組合多個(gè)深度學(xué)習(xí)算法,緩解單個(gè)算法學(xué)習(xí)偏置問題。對(duì)卷積神經(jīng)網(wǎng)絡(luò)稍作修改,作為基礎(chǔ)算法,使用boosting和bagging兩種算法對(duì)多個(gè)基礎(chǔ)算法進(jìn)行組合研究,并且采用多個(gè)抽樣算法提高基礎(chǔ)算法的多樣性,進(jìn)而提高整個(gè)算法的性能;第二,為文檔生成更優(yōu)的向量表示,減少文檔表示中的噪聲;短文本特征稀疏,直接使用深度學(xué)習(xí)算法得到的文檔的向量表示包含較多噪聲。通過采用多任務(wù)學(xué)習(xí)算法,同時(shí)訓(xùn)練多個(gè)與情感分類相關(guān)的任務(wù),將更多的特征信息回饋到文檔的向量表示中,從而減小噪聲,提高文檔表示的質(zhì)量。使用多個(gè)領(lǐng)域,多種分布的測(cè)試集測(cè)試相關(guān)算法,實(shí)驗(yàn)結(jié)果表明本文研究算法具有比較強(qiáng)的泛化能力,總體效果基本符合預(yù)期。
[Abstract]:With the rapid development of the Internet, especially the mobile Internet, it is flooded with a large number of emotional text books, and the emotional information contained in these texts can be mined to obtain a large amount of business information and social information. In this paper, we use depth learning algorithm to analyze the text emotion from two aspects. First, we can effectively combine multiple depth learning algorithms to alleviate the single algorithm learning bias problem. The convolutional neural network is modified a little, as the basic algorithm, boosting and bagging are used to study the combination of several basic algorithms, and multiple sampling algorithms are used to improve the diversity of the basic algorithm, and then to improve the performance of the whole algorithm. Secondly, a better vector representation is generated for the document to reduce the noise in the document representation, and the vector representation of the document obtained by using the depth learning algorithm directly contains more noise. By using multi-task learning algorithm and training several tasks related to emotion classification, more feature information is returned to the vector representation of documents, thus reducing noise and improving the quality of document representation. The experimental results show that the proposed algorithm has a strong generalization ability and the overall effect is basically in line with the expectation.
【學(xué)位授予單位】：北京郵電大學(xué)
【學(xué)位級(jí)別】：碩士
【學(xué)位授予年份】：2016
【分類號(hào)】：TP391.1;TP181

【參考文獻(xiàn)】

相關(guān)會(huì)議論文 前1條

1 章劍鋒;張奇;吳立德;黃萱菁;;中文評(píng)論挖掘中的主觀性關(guān)系抽取[A];第三屆全國(guó)信息檢索與內(nèi)容安全學(xué)術(shù)會(huì)議論文集[C];2007年

，

本文編號(hào)：2259299