句子分类

学习句子分类，使用深度学习的方法对句子数据集进行分类。

问题

句子分类（Sentence Classification）是指给定一个句子，标注预先设定的若干类别中的一个类别。

句子分类包括情感分析（Sentiment Analysis）、问题分类（Question
Classification）等任务。情感分析又称倾向性分析、意见抽取（Opinion extraction）、意见挖掘（Opinion mining）、情感挖掘（Sentiment mining）、主观分析（Subjectivity analysis），它是对带有情感色彩的主观性文本进行分析、处理、归纳和推理的过程，如从评论文本中分析用户对“数码相机”的“变焦、价格、大小、重量、闪光、易用性”等属性的情感倾向。

应用

了解对电影、商品、Twitter 等的褒贬评价，以此来改善产品和服务、发现竞争对手的优劣势、预测股票走势等。

数据集

Data	c	l	N	|V|	|V_pre|	Test
MR	2	20	10662	18765	16448	CV
SST-1	5	18	11855	17836	16262	2210
SST-2	2	19	9613	16185	14838	1821
Subj	2	23	10000	21323	17913	CV
TREC	6	10	5952	9592	9125	500
CR	2	19	3775	5340	5046	CV
MPQA	2	3	10606	6246	6083	CV

- MR: Movie reviews 电影评论，每条评论包含一个句子。1

• SST-1: Stanford Sentiment Treebank，MR 的扩展但划分了 train/dev/test 集合并提供 5 个细粒度标签（非常积极的，积极的，中性的，负面的，非常消极的）。

• SST-2: 与 SST-1 一样但移除中性评论并用二进制标签。2

• Subj: Subjectivity 主观性数据集，任务是将句子分类为主观或客观的。3

• TREC: TREC question dataset TREC 问题数据集，任务是将一个问题分成 6 类（关于人、位置、数字信息等）。4

• CR: Customer reviews 各种产品的客户评论，任务是预测正面/负面评论。5

• MPQA: MPQA 数据集意见极性检测任务。6

方法

通常会把任务拆分成几个子任务：

1. 分词

   把句子根据意思分成多个词，有时可能还需要去掉停用词、了解词性、转换成词向量等操作。

2. 提取特征

   有时我们不会直接使用分词后的多个词来直接分类，这时需要提取特征来方便分类。

   常用特征：TF-IDF、LDA、LSI

3. 构建分类器

   输入特征或词向量等，通过一些模型，对该句子进行分类。

Naive Bayes

NBSVM: Naive Bayes SVM

MNB: Multinomial Naive Bayes 7

combine-skip

combine-skip + NB 8

Model	MR	SST-1	SST-2	Subj	TREC	CR	MPQA
NBSVM	79.4	-	-	93.2	-	81.8	86.3
MNB	79.0	-	-	93.6	-	80.0	86.3
combine-skip	76.5	-	-	93.6	92.2	80.1	87.1
combine-skip+NB	80.4	-	-	93.6	-	81.3	87.5

RNN

RCNN: Recurrent Convolutional Neural Networks 9

S-LSTM: Long Short-Term Memory Over Recursive Structures 10

LSTM: Long Short-Term Memory

BLSTM: Bidirectional Long Short-Term Memory

Tree-LSTM: Tree-structured Long Short-Term Memory 11

LSTMN: Long Short-Term Memory-Network 12

Multi-Task: Recurrent Neural Network for Text Classification with Multi-Task Learning 13

BLSTM-Att: Bidirectional Long Short-Term Memory, attention-based model

BLSTM-2DPooling: Bidirectional Long Short-Term Memory Networks with Two-Dimensional Max Pooling

BLSTM-2DCNN: Bidirectional Long Short-Term Memory Networks with 2D convolution 14

Model	MR	SST-1	SST-2	Subj	TREC	CR	MPQA
RCNN	-	47.21	-	-	-	-	-
S-LSTM	-	-	81.9	-	-	-	-
LSTM	-	46.4	84.9	-	-	-	-
BLSTM	-	49.1	87.5	-	-	-	-
Tree-LSTM	-	51.0	88.0	-	-	-	-
LSTMN	-	49.3	87.3	-	-	-	-
Multi-Task	-	49.6	87.9	94.1	-	-	-
BLSTM	80.0	49.1	87.6	92.1	93.0	-	-
BLSTM-Att	81.0	49.8	88.2	93.5	93.8	-	-
BLSTM-2DPooling	81.5	50.5	88.3	93.7	94.8	-	-
BLSTM-2DCNN	82.3	52.4	89.5	94.0	96.1	-	-

CNN

DCNN: Dynamic Convolutional Neural Network 15

CNN-non-static: Convolutional Neural Networks, the pretrained vectors are fine-tuned for each task

CNN-multichannel: Convolutional Neural Networks with two sets of word vectors 16

TBCNN: Tree-based Convolutional Neural Network 17

Molding-CNN: Molding Convolutional Neural Networks 18

CNN-Ana: Non-static GloVe+word2vec CNN 19

MVCNN: Multichannel Variable-Size Convolution 20

DSCNN: Dependency Sensitive Convolutional Neural Networks 21

Model	MR	SST-1	SST-2	Subj	TREC	CR	MPQA
DCNN	-	48.5	86.8	-	93.0	-	-
CNN-non-static	81.5	48.0	87.2	93.4	93.6	84.3	89.5
CNN-multichannel	81.1	47.4	88.1	93.2	92.2	85.0	89.4
TBCNN	-	51.4	87.9	-	96.0	-	-
Molding-CNN	-	51.2	88.6	-	-	-	-
CNN-Ana	81.02	45.98	85.45	93.66	91.37	84.65	89.55
MVCNN	-	49.6	89.4	-	-	-	-
DSCNN	81.5	49.7	89.1	93.2	95.4	-	-

Others

RAE: Recursive Autoencoders with pre-trained word vectors from Wikipedia 22

AdaSent: self-adaptive hierarchical sentence model 23

RNTN: Recursive Neural Tensor Network 24

DRNN: Deep Recursive Neural Networks 25

Model	MR	SST-1	SST-2	Subj	TREC	CR	MPQA
RAE	77.7	43.2	82.4	-	-	-	86.4
AdaSent	83.1	-	-	95.5	92.4	86.3	93.3
RNTN	-	45.7	85.4	-	-	-	-
DRNN	-	49.8	86.6	-	-	-	-

参考

---

1. (ACL 2005) Seeing Stars: Exploiting Class Relationships For Sentiment Categorization With Respect To Rating Scales https://www.cs.cornell.edu/people/pabo/movie-review-data/ ↩
2. (EMNLP 2013) Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank https://nlp.stanford.edu/sentiment/ ↩
3. (ACL 2004) A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts http://www.cs.cornell.edu/people/pabo/movie-review-data ↩
4. (ACL 2002) Learning Question Classifiers http://cogcomp.org/Data/QA/QC/ ↩
5. (SIGKDD 2004) Mining and Summarizing Customer Reviews http://www.cs.uic.edu/~liub/FBS/sentiment-analysis.html ↩
6. (Language Resources and Evaluation 2005) Annotating Expressions Of Opinions And Emotions In Language http://mpqa.cs.pitt.edu/ ↩
7. (ACL 2012) Baselines and Bigrams: Simple, Good Sentiment and Topic Classification ↩
8. (NIPS 2015) Skip-Thought Vectors ↩
9. (AAAI 2015) Recurrent Convolutional Neural Networks for Text Classification ↩
10. (ICML 2015) Long Short-Term Memory Over Recursive Structures ↩
11. (ACL 2015) Improved Semantic Representations From Tree-Structured Long Short-Term Memory Networks ↩
12. (EMNLP2016) Long Short-Term Memory-Networks for Machine Reading ↩
13. (IJCAI 2016) Recurrent Neural Network for Text Classification with Multi-Task Learning ↩
14. (COLING 2016) Text Classification Improved by Integrating Bidirectional LSTM with Two-dimensional Max Pooling ↩
15. (ACL 2014) A Convolutional Neural Network for Modelling Sentences ↩
16. (EMNLP 2014) Convolutional Neural Networks for Sentence Classification ↩
17. (EMNLP 2015) Discriminative Neural Sentence Modeling by Tree-Based Convolution ↩
18. (EMNLP 2015) Molding CNNs for text: non-linear, non-consecutive convolutions ↩
19. (IJCNLP 2017) A Sensitivity Analysis of (and Practitioners’ Guide to) Convolutional Neural Networks for Sentence Classification ↩
20. (CoNLL 2015) Multichannel Variable-Size Convolution for Sentence Classification ↩
21. (NAACL 2016) Dependency Sensitive Convolutional Neural Networks for Modeling Sentences and Documents ↩
22. (EMNLP 2011) Semi-Supervised Recursive Autoencoders for Predicting Sentiment Distributions ↩
23. (IJCAI 2015) Self-adaptive hierarchical sentence model ↩
24. (EMNLP 2013) Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank ↩
25. (NIPS 2014) Deep Recursive Neural Networks for Compositionality in Language ↩