基于深度表示學習的跨領域情感分析
發(fā)布時間：2018-11-17 21:39
【摘要】：【目的】通過在標注資源豐富的源領域中學習,并將目標領域的文檔投影到與源領域相同的特征空間中去,從而解決目標領域因數(shù)據(jù)量較小難以獲得好的分類模型的問題�！痉椒ā窟x擇亞馬遜在線購物網(wǎng)站在書籍、DVD和音樂類目下的中文、英文和日文評論作為實驗數(shù)據(jù),在卷積神經(jīng)網(wǎng)絡和結(jié)構對應學習的基礎上提出跨領域深度表示模型(CDDRM),以實現(xiàn)不同領域環(huán)境下的知識遷移,并將其應用到跨領域情感分析任務之中�！窘Y(jié)果】實驗結(jié)果表明,CDDRM在跨領域環(huán)境下最優(yōu)的F值達到0.7368,證明了該模型的有效性�！揪窒蕖緾DDRM針對長文本的跨領域情感分類F值仍然有待提升�！窘Y(jié)論】知識遷移能夠解決監(jiān)督學習在小數(shù)據(jù)集上難以獲得好的分類效果的問題,與傳統(tǒng)監(jiān)督學習的基本假設相比,它并不要求訓練集和測試集服從相同或相似的數(shù)據(jù)分布。
[Abstract]:[objective] by learning in resource-rich source fields and projecting documents from target domains into the same feature space as source domains, In order to solve the problem that the target area is difficult to obtain good classification model because of the small amount of data. [methods] the Chinese, English and Japanese reviews of Amazon online shopping website under books, DVD and music categories are selected as experimental data. On the basis of convolution neural network and structure correspondence learning, a cross-domain depth representation model (CDDRM),) is proposed to realize knowledge transfer in different domain environments. The experimental results show that the optimal F value of CDDRM in cross-domain environment is 0.7368. It is proved that this model is effective. [limitations] the F value of CDDRM's cross-domain affective classification for long text still needs to be improved. [conclusion] knowledge transfer can solve the problem that supervised learning is difficult to obtain good classification effect on small data sets. Compared with the traditional supervised learning hypothesis, it does not require the training set and the test set to be distributed from the same or similar data.
【作者單位】： 中南財經(jīng)政法大學信息與安全工程學院;武漢大學信息管理學院;
【基金】：國家自然科學基金面上項目“大數(shù)據(jù)環(huán)境下基于領域知識獲取與對齊的觀點檢索研究”(項目編號:71373286);國家自然科學基金青年項目“突發(fā)公共衛(wèi)生事件社交媒體信息主題演化與影響力建�！�(項目編號:71603189) 武漢大學人文社會科學青年學者學術發(fā)展計劃學術團隊項目“人機交互與協(xié)作創(chuàng)新”(項目編號:Whu2016020)的研究成果之一
【分類號】：TP391.1
本文編號：2339113
[Abstract]:[objective] by learning in resource-rich source fields and projecting documents from target domains into the same feature space as source domains, In order to solve the problem that the target area is difficult to obtain good classification model because of the small amount of data. [methods] the Chinese, English and Japanese reviews of Amazon online shopping website under books, DVD and music categories are selected as experimental data. On the basis of convolution neural network and structure correspondence learning, a cross-domain depth representation model (CDDRM),) is proposed to realize knowledge transfer in different domain environments. The experimental results show that the optimal F value of CDDRM in cross-domain environment is 0.7368. It is proved that this model is effective. [limitations] the F value of CDDRM's cross-domain affective classification for long text still needs to be improved. [conclusion] knowledge transfer can solve the problem that supervised learning is difficult to obtain good classification effect on small data sets. Compared with the traditional supervised learning hypothesis, it does not require the training set and the test set to be distributed from the same or similar data.
【作者單位】： 中南財經(jīng)政法大學信息與安全工程學院;武漢大學信息管理學院;
【基金】：國家自然科學基金面上項目“大數(shù)據(jù)環(huán)境下基于領域知識獲取與對齊的觀點檢索研究”(項目編號:71373286);國家自然科學基金青年項目“突發(fā)公共衛(wèi)生事件社交媒體信息主題演化與影響力建�！�(項目編號:71603189) 武漢大學人文社會科學青年學者學術發(fā)展計劃學術團隊項目“人機交互與協(xié)作創(chuàng)新”(項目編號:Whu2016020)的研究成果之一
【分類號】：TP391.1
【相似文獻】
相關期刊論文 前1條
1 陳紅;;e-Learning培訓的效能及其適用條件[J];教育教學論壇;2010年14期
，本文編號：2339113
本文鏈接：http://sikaile.net/kejilunwen/ruanjiangongchenglunwen/2339113.html
最近更新
教材專著
