机器理解：文档级多方面情感分类的机器阅读理解方法

173 浏览量更新于2024-08-26 收藏 1.6MB PDF 举报

"这篇研究论文探讨了文档级别的多方面情感分类作为机器理解的任务，通过构建伪问题答案对，利用少量与方面相关的关键词和评分来建模。论文提出了一个层次化的迭代注意力模型，通过文档和方面问题之间的频繁交互来构建方面特定的表示。采用层次结构来表示词级和句级信息，并利用一种称为‘a’的机制来引导注意力集中在关键信息上，以实现更准确的情感分类。" 在当前的自然语言处理领域，文档级别的多方面情感分类是一个关键任务，特别是在客户关系管理中，它有助于理解和分析消费者的意见和情绪。传统的单方面或单一情感的分类方法可能无法全面捕捉文档中的复杂情感信息。这篇2017年在 Empirical Methods in Natural Language Processing (EMNLP) 会议上的论文，提出了一种新的方法，将这个问题转化为机器理解的问题。论文作者提出的方法基于伪问题答案对的概念，通过选取少量与特定方面（如产品特性、服务体验等）相关的关键词，结合用户给出的评分，来构建这些问题。这使得模型能够关注到文档中与这些方面相关的关键信息。为了处理这种复杂的交互，他们设计了一个层次化的迭代注意力模型。这个模型分为两个层次：词级和句级，分别捕捉文本中的微小细节和整体语境。通过反复的注意力交互，模型可以逐步聚焦于与各个方面相关的信息，形成对每个方面的特定表示。这种注意力机制允许模型动态地关注文档中的重要部分，从而提高情感分类的准确性。此外，论文中提到的"a"机制，虽然具体细节没有给出，但可以推测这是一种帮助模型聚焦并优先处理关键信息的策略。这可能是通过某种形式的注意力权重分配来实现的，使模型能够在大量文本数据中定位和理解关键情感线索。这篇研究论文为处理多维度、多层次的情感分析提供了一种创新的机器学习方法，通过模拟人类理解过程，提高了情感分类的深度和精度。这种方法不仅对于NLP领域的理论研究具有价值，还可能对实际应用如客户服务、市场分析等领域产生深远影响。

, q

, . . . , q

}, we use N

aspect-related key-

words,

, q

. . . q

, to represent it. Simi-

larly, we use q

, q

as the one-hot encoding and

word embedding for q

respectively.

There are several sophisticated methods for

choosing aspect keywords (e.g., topic model).

Here, we consider a simple way where ﬁve seeds

were ﬁrst manually selected for each aspect and

then more words were obtained based on their co-

sine similarities with seeds

As shown in Figure 2 (left), our framework fol-

lows the idea of multi-task learning, which learns

different aspects simultaneously. In this case, all

these tasks share the representations of words and

architecture of semantic model for the ﬁnal clas-

siﬁers. Different from straightforward neural net-

work based multi-task learning (Collobert et al.,

2011), for each document d and an aspect q

, our

model uses both the content of d and all the related

keywords

, q

. . . q

as input. Since the

keywords can cover most of the semantic mean-

ings of the aspect, and we do not know which

document mentions which semantic meaning, we

build an attention model to automatically decide

it (introduced in Section 2.3). Assuming that the

keywords have been decided, we use a hierarchi-

cal attention model to select useful information

from the review documents. As shown in Figure 2

(right), the hierarchical attention of keywords is

applied to both sentence level (to select meaning-

ful words) and document level (to select mean-

ingful sentence). Thus, our model builds aspect-

speciﬁc representations in a bottom-up manner.

Speciﬁcally, we obtain sentence representa-

tions



, s

, . . . s



using the input encoder (Sec-

tion 2.2) and iterative attention module (Sec-

tion 2.3) at the word level. Then we take sen-

tence representations and k -th aspect as input and

apply the sentence-level input encoder and atten-

tion model to generate the document representa-

tion d

for ﬁnal classiﬁcation. As shown in Fig-

ure 2 (right), the attention model is applied twice

at different levels of the representation.

2.2 Input Encoder

The input module builds memory vectors for the

iterative attention module and is performed both at

word and sentence levels. For a document, it con-

For example, the words “value,” “price,” “worth,” “cost,”

and “$” are selected as seeds for aspect Price. The informa-

tion for seeds can be found in our released resource.

verts word sequence into word level memory M

and sentence sequence into sentence level mem-

ory M

respectively. For an aspect question q

, it

takes a set of aspect-speciﬁc words {q

}

1≤i≤N

as input and derives word level memory M

and

sentence level memory M

To construct M

, we obtain word embeddings

, w

, . . . w

from an embedding matrix

applied to all words shown in the corpus.

Then, LSTM (Hochreiter and Schmidhuber, 1997)

model is used as the encoder to produce hidden

vectors of words based on the word embeddings.

At each step, LSTM takes input w

and derives

a new hidden vector by h

= LSTM(w

, h

t−1

To preserve the subsequent context information

for words, another LSTM is ran over word se-

quence in a reverse order simultaneously. Then the

forward hidden vector

−→

and backward hidden

vector

←−

are concatenated as phrase embedding

. We stack these phrase embeddings together

as word level memory M

. Similarly, we feed

sentence representations into another Bi-LSTM to

derive the sentence level memory M

. Note that,

the sentence representations are obtained using the

iterative attention module which is described as

Eq. (5) in Section 2.3.

Since we have question keywords as input, to

allow the interactions between questions and doc-

uments, we also build question memory in follow-

ing way. We obtain Q





1≤i≤N

by look-

ing up an embedding matrix

applied to all

question keywords. Then a non-linear mapping

is applied to obtain the question memory at word

level:

= tanh(Q

), (1)

where W

is the parameter matrix to adapt q

word level. Similarly, we use another mapping to

obtain the sentence level memory:

= tanh(Q

), (2)

where W

is the parameter matrix to adapt q

sentence level.

2.3 Iterative Attention Module

The iterative attention module (IAM) attends and

reads memories of questions and documents al-

ternatively with a multi-hop mechanism, deriving

and E

are initialized by the same pre-trained em-

beddings but are different embedding matrices with different

updates.

2046

剩余10页未读，继续阅读

weixin_38749268

粉丝: 5
资源: 943

机器理解：文档级多方面情感分类的机器阅读理解方法

基于深度学习的情感分类研究.pdf

tf-idf词袋模型、jieba 文本情感分类

请写出机器学习对文档数据分类的完整的Scala命令

ara::com文档理解

请写出机器学习对文档数据分类的Scala命令

tfidf怎么运用在情感分类

基于深度学习的机器视觉:垃圾分类python仿真(完整源码+数据+文档).rar

将文档内容分类 c++

多模态 智能文档解析

tensorflow文档

最新资源

多模态智能文档解析