【摘要】：傳統(tǒng)的年齡回歸方法不能學(xué)習(xí)深層次信息,因此利用能充分挖掘上下文關(guān)系信息的深度學(xué)習(xí)方法來識別用戶的年齡。具體而言,提出了一種基于LSTM的年齡回歸方法,其能夠?qū)W習(xí)長期依賴關(guān)系即建立輸入值之間的長相關(guān)聯(lián)系。采用了兩種不同的特征,即文本特征和社交特征。為了有效地區(qū)分這兩種特征,充分利用這兩種特征之間的信息,進一步提出了基于雙通道LSTM的年齡回歸方法,具體實現(xiàn)是在神經(jīng)網(wǎng)絡(luò)中加入Merge層,將LSTM分別產(chǎn)生的文本特征表示和社交特征表示結(jié)合進行集成學(xué)習(xí)以充分學(xué)習(xí)文本特征和社交特征間的聯(lián)系。實驗結(jié)果表明,基于雙通道LSTM的年齡回歸方法能夠有效地區(qū)分文本特征和社交特征,并且較單個LSTM方法能夠取得更好的年齡回歸性能。
[Abstract]:The traditional age regression method can not learn the deep level information, so the depth learning method which can fully mine the contextual information can be used to identify the age of the user. Specifically, an age regression method based on LSTM is proposed, which can learn long-term dependency, that is, establish long correlation between input values. Two different features, text feature and social feature, are adopted. In order to effectively distinguish the two features and make full use of the information between the two features, a new age regression method based on two-channel LSTM is proposed, which is realized by adding merge layer into neural network. The text feature representation and the social feature representation generated by LSTM are integrated to learn the relationship between the text feature and the social feature. The experimental results show that the age regression method based on two-channel LSTM can effectively distinguish text features from social features, and can achieve better age regression performance than the single LSTM method.
【作者單位】： 蘇州大學(xué)自然語言處理實驗室;
【基金】：國家自然科學(xué)基金重點資助項目(61331011);國家自然科學(xué)基金資助項目(61375073,61273320)
【分類號】：O212.1;TP391.1
，

本文編號：2119604