基于微博的意图识别

发布时间：2019-06-01 16:35

【摘要】：微博是一种新兴的社交平台,数以亿计的用户每天在微博上发布海量的微博数据。在这些海量的微博中有些微博具有一定意图,它们通常使用显式或者隐式的表达方式来表达相应的意图。将微博中的这些意图准确地识别出来具有巨大的商业价值,本文针对微博中的意图识别主要进行了以下三方面的研究:(1)微博中的显式意图的识别。具有显式意图的微博通常有“想要”,“希望”等意图触发词。本文我们提出了一种新的基于维基百科的显式意图识别模型。对于每一种意图,我们首先选取一些最能代表这种意图的概念集合,即种子概念。然后将种子概念放在维基百科进行查询,通过概念之间的链接关系,我们可以获得与种子概念相关的概念集合,这些扩展的概念在一定程度上也具有相同意图。之后我们利用获得的概念集合构建相应意图的维基百科链路图,在图中用随机游走算法对每个概念分配意图得分。最后我们将微博映射到相应的意图空间中得到相应的意图得分,我们根据意图得分判断微博是否具有对应的意图。如果微博中的某些词没有被维基百科收录,我们利用显式语义分析(ESA)的方法将其映射到最相关的维基百科概念,然后再进行映射计算相应的意图得分。(2)微博中隐式意图的识别。具有隐式意图的微博通常不具有意图触发词,但是可以通过推理得到微博中的意图。目前大多数的研究工作都是针对显式意图识别的,而本文我们进行了微博中隐式意图的识别。我们利用编码器-解码器模型(Encoder-Decoder)将具有隐式意图的微博“翻译”为相应的显式意图的表达方式,然后再进行显式意图的识别。编码器-解码器模型主要应用于序列到序列问题(seq2seq),例如机器翻译,语音识别,图像描述等,而隐式意图与显式意图的转换也属于seq2seq问题,因此可以使用编码器-解码器模型。传统的基于RNN的编码器-解码器模型的主要思想是将输入的句子编码为一个固定长度的语义向量,然后将该语义向量解码生成相应的输出句子。Bahdanau等人后来提出了注意力模型,改进了基于RNN的编码器-解码器模型,他们将输入句子编码为一个长度不固定的语义向量,这样使得即使句子长度很长,翻译效果同样很好。实验中我们进行了两种模型的对比,结果表明注意力模型要比基于RNN的编码器-解码器模型效果更好。为了训练模型,我们构建了包含隐式意图的微博和相应的显式意图微博的语料库。一旦我们通过注意力模型获得意图的显式表达方式之后,接下来就可以使用本文提出的基于维基百科的显式意图识别模型识别其中的显式意图。(3)微博中意图的识别。我们提出一种基于词向量与卷积神经网络的意图识别模型,这种模型具有通用性,不仅可以用来识别显式意图,也可以用来识别隐式意图。模型的通用性主要得益于两方面,一方面是词的词向量表示具有丰富的语义特征,另一方面是因为卷积神经网络可以提取句子的语义特征。因此,当我们将意图识别问题视为多元分类问题时,即一条微博是否具有某种意图,模型可以对具有意图的微博进行分类,无论其中的意图表达方式是显式的还是隐式的,词向量和卷积神经网络模型都可以提取其中的语义特征然后进行正确的意图识别。
[Abstract]:The Internet is a new social platform, with hundreds of millions of users posting massive microblogging data on microblogs every day. Some of these microblogs have some intent in these massive volumes, and they often use explicit or implicit expressions to express the corresponding intent. The purpose of this paper is to identify the following three aspects: (1) the identification of the explicit intention in the system. Microblogs with explicit intent typically have "want", "hope", and other intent-triggered words. In this paper, a new wikipedia-based explicit intention recognition model is presented. For each intent, we first select a set of concepts that best represent this intent, that is, the concept of a seed. The concept of the seeds is then placed in Wikipedia for query, through the link between concepts, we can obtain a set of concepts related to the concept of the seeds, which have the same intent to some extent. After that, we use the obtained concept set to construct the corresponding intention Wikipedia link diagram, and assign the intention score to each concept by the random walk algorithm in the figure. Finally, we map the micro-blog to the corresponding intention space to get the corresponding intention score, and we judge whether the micro-blog has the corresponding intention according to the intention score. If some of these words are not included in Wikipedia, we map it to the most relevant Wikipedia concept by means of explicit semantic analysis (ESA) and then map the corresponding intent score. (2) The identification of implicit intention in the system. Microblog with implicit intent usually does not have an intention to trigger a word, but it can be inferred by reasoning. At present, most of the research work is identified for explicit intent, and the text of this article is to identify the implicit intent. We use the Encoder-Decoder model (Encoder-Decoder) to use the implicit intent-oriented microblogging "translation" as the corresponding explicit intent expression, and then to identify the explicit intent. The encoder-decoder model is mainly applied to sequence-to-sequence problems (seq2seq), such as machine translation, speech recognition, image description, etc., while the implicit intention and explicit intention conversion also belong to the seq2seq problem, and thus an encoder-decoder model can be used. The main idea of the traditional RNN-based encoder-decoder model is to encode the input sentence into a fixed-length semantic vector, and then decode the semantic vector to generate the corresponding output sentence. The Bahdanau et al. later presented an attention model that improved the RNN-based encoder-decoder model, which codes the input sentence as a length-invariant semantic vector, so that even if the sentence length is long, the translation effect is equally good. The results show that the attention model is better than the RNN-based encoder-decoder model. In order to train the model, we built a corpus of microblogs with implicit intent and corresponding explicit intent microblogs. Once we get the explicit expression of intent through the attention model, the explicit intent in this article can then be identified using the Wikipedia-based explicit intent recognition model proposed in this article. (3) Identification of the intention in the case. We propose an intention recognition model based on the word vector and the convolution neural network. The model has the universality, not only can be used to identify the explicit intention, but also can be used to identify the implicit intention. The generality of the model is mainly from two aspects, on the one hand, the word vector representation of the word has a rich semantic feature, and on the other hand, the convolution neural network can extract the semantic features of the sentence. Therefore, when we view the problem as a multi-classification problem, that is, whether a micro-blog has some intention, the model can classify the micro-blog with intention, whether the intention expression is explicit or implicit, The word vector and the convolution neural network model can extract the semantic features therein and then perform the correct intent recognition.
【学位授予单位】：西华大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.1;TP393.092

【相似文献】