深度学习在细粒度情感分析中的比较研究：从提取到分类

版权申诉

5星 · 超过95%的资源 51 浏览量更新于2024-07-21 收藏 4.92MB PDF 举报

随着互联网上用户生成内容的爆炸性增长，情感分析已经成为挖掘人类情绪状态的重要工具。这篇由Hai Ha Do和PWC Prasad撰写的深度学习在基于方面的情感分析中的比较综述文章发表于《专家系统与应用》杂志（Expert Systems with Applications）第118期，2019年，272-299页。该研究关注的是提高情感分析的精细度，特别是在两个关键任务上的改进：方面提取和产品评论的情感分类，以及依赖目标的情感分类，如目标导向的推文。深度学习方法因其能够捕捉文本中的语法和语义特征而被看作是实现这些目标的有前景途径。传统的技术往往依赖于手工设计的特征工程，而深度学习通过自动学习的方式，避免了对高级特征的繁琐设计，能够更好地处理复杂的情感表达。文章可能探讨了以下知识点： 1. **深度学习模型**：文中可能会介绍各种深度学习模型，如卷积神经网络（CNN）、循环神经网络（RNN）、长短时记忆网络（LSTM）、双向LSTM、注意力机制等，它们如何在句子结构和上下文理解中发挥作用。 2. **情感表示**：如何通过深度学习捕捉情感词汇、情感强度、否定词和情感转移等，以更准确地判断情感极性。 3. **方面检测**：阐述了如何利用深度学习进行多层特征提取，以便从文本中识别和定位关键的评价方面。 4. **迁移学习和预训练模型**：深度学习模型如BERT、ELMo等在情感分析任务中的应用，以及如何利用预训练模型在少量标注数据下提升性能。 5. **评估和基准**：文中可能包括了对不同深度学习模型在各种公开数据集（如SemEval、Yelp、Amazon等）上的实验结果比较，以及对于模型性能（如精确度、召回率、F1分数）的分析。 6. **挑战和未来方向**：深度学习在情感分析中的局限性，例如数据依赖性、过拟合风险，以及针对这些问题可能的解决策略和未来研究的展望。这篇文章深入剖析了深度学习在提高情感分析的方面级别精确度方面的潜力和进展，并为研究人员和开发者提供了宝贵的参考，帮助他们选择最适合的深度学习方法来处理日益增长的在线文本数据。

H.H. Do et al. / Expert Systems With Applications 118 (2019) 272–299 277

Fig. 3. Comparison of classic machine learning and deep learning processes for NLP. Deep learning architecture is characterized by dense embeddings and hidden layers -

adapted from Thanaki (2017) .

3.2. Initialization of input vectors

Before reaching the ﬁrst hidden layer of the DNN, the input

layer is encoded with a distributed representation, or word em-

beddings, which represent each word as a low-dimensional, real-

valued and continuous vector to encode its semantic and syntactic

properties ( Tang, Wei, et al., 2016 ).

3.2.1. Word embeddings vectors

One of the ﬁrst word embedding models is by Bengio, Schwenk,

Senécal, Morin, & Gauvain (2006) who proposed a neural proba-

bilistic language model with a shared lookup table. Given a word

and its previous words, the model can look up its continuous vec-

tor, feed the vector to a feed-forward neural network and predict

the probability function for its next word. Assuming a sequence of

T words w

, w

, , w

with n previous words fed into the model,

the model can predict the probability p of the words u

based on

ﬁnding the model parameter θ that maximises the objective func-

tion J :

J =



t=1

log g( w

, w

t−1

, ··· , w

t−n +1

; θ ) + R

(

)

(11)

where R ( θ) is the regularisation term, and g ( w

, w

t − 1

, , w

t − n + 1

;

θ) can be estimated by the softmax function as p ( w

| w

t − 1

, ,

t − n + 1

Mikolov et al. (2013) developed the word2vec with two dif-

ferent neural network models for creating word embeddings for

training on large corpora: a bag-of-words based model (CBOW)

that obtains word context from sequential word context; and the

skip-gram model that predicts the word embeddings from neigh-

bouring words ( Fig. 4 ).

The CBOW for the target as the word w

at time step t , the

model receives a window of n words around w

, and the loss func-

tion J can be written as:

J =



t=1

log p ( w

| w

t−n

, ··· , w

t−1

, w

t+1

, ··· , w

t+ n

) (12)

In contrast, the skip-gram model uses the centre word w

predict the neighbouring words w

t + j

. In this case, the objective

function is:

J =



t=1



−n ≤ j≤n,  =0

log p ( w

t+ j

| w

) (13)

In addition to word2vec framework, a number of software have

been developed for training word embeddings such as GloVe de-

veloped by Stanford University or fastText developed by Facebook.

剩余27页未读，继续阅读

Fun_He

粉丝: 19
资源: 104

深度学习在细粒度情感分析中的比较研究：从提取到分类

Deep learning for sentiment analysis A survey 深度学习情感分析综述

Aspect-Based-Sentiment-Analysis:一个为SemEval 2016数据集实现基于方面的情感分析分类系统的python程序

deep-learning-for-sentiment-analysis:多模态情感分析的深度学习

马云飞A_Novel_Hybrid_Deep_Learning_Model_for_Sentiment_Classification.pdf

Sentiment_analysis_deep_learning

[machine_learning_mastery系列]deep_learning_with_python.pdf(with code)

Semi-Supervised-CNN-for-Sentiment-Analysis.zip_https//semi.CNN_l

PyPI 官网下载 | aspect-based-sentiment-analysis-2.0.0.tar.gz

Sentiment_Analysis_or_Opinion_Mining_A_Review.pdf

Sentiment_Analysis_Deep_Learning：使用深度学习（CNN）进行情感分析

最新资源