Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.
Awesome Lists | Featured Topics | Projects
https://github.com/candlewill/concept_word_embeddings

Concept_word_embeddings
https://github.com/candlewill/concept_word_embeddings
Last synced: about 1 month ago
JSON representation
Concept_word_embeddings
Host: GitHub
URL: https://github.com/candlewill/concept_word_embeddings
Owner: candlewill
Created: 2015-11-08T12:26:54.000Z (about 9 years ago)
Default Branch: master
Last Pushed: 2015-11-17T09:06:38.000Z (about 9 years ago)
Last Synced: 2023-08-02T06:24:31.491Z (over 1 year ago)
Language: Python
Size: 12.5 MB
Stars: 1
Watchers: 2
Forks: 0
Open Issues: 0
Metadata Files:
- Readme: README.md
Awesome Lists containing this project

README

        # Concept Level Word Embeddings

# 概念级别的词向量

本程序目标是训练概念级别的句向量，以分析替换情感词为其对应的概念后是否对构造句向量有影响。具体而言，本程序提供如下功能：

- 提供一个用于训练原始词向量的语料库

- 我们将上述语料库中的词汇，属于ANEW词典中的那部分情感词，按照WordNet同义词典进行替换，得到另一个概念语料库

- 我们用上述两个语料库，使用Doc2vec工具训练两份句项量

- 我们对这两份句向量，对于其中有VA标记的部分，使用简单的线性回归技术，分析他们在预测VA时性能表现是否有差距

### 动机

进行本实验的动机是，我们观察到无论是使用Word2vec训练词向量，还是使用Doc2vec训练句向量，都面临一个问题，训练出的向量更多的表达的是语义相似性而非情感相似性，如，good和bad虽然表达的是相反的情感倾向，但由于其出现的上下文较相似，因此这两个词向量的比较相似。

### 方法

为了解决这个问题，我们提出对语料库进行特殊的预处理，将语料库中的情感词替换为其对应的高层概念——synset编号。例如，在WordNet同义词典中，baby和child属于同义词，因此对于语料库中出现baby和child的地方我们用synset_id替换。

在NLTK提供的wordnet工具中，提供了每一个词汇的同义词，然而并没有synset ID。因此，我们将ANEW中同义词替换为ANEW中的词汇。也就是说，我们通过wordnet搜寻所有ANEW中词汇的同义词，然后出现同义词的地方用anew替换。例如：

替换之前的文本如下图：

![替换之前](http://7xo9ej.com1.z0.glb.clouddn.com/替换之后.png "图1：替换之前")

替换之后的文本如下图：

![替换之后](http://7xo9ej.com1.z0.glb.clouddn.com/替换之前.png "图2：替换之后")

从上图中可以发现，第一行中，love替换为了passion，第二行中good替换为了serious。这是因为passion和serious为ANEW词典中的词汇，而依据wordnet，love是passion的同义词，good是serious的同义词，见wordnet[在线演示网址](http://wordnetweb.princeton.edu/perl/webwn?s=serious&sub=Search+WordNet&o2=&o0=1&o8=1&o1=1&o7=&o5=&o9=&o6=&o3=&o4=&h=0000000)，其他替换方式类同。更多详细比较，见resource目录中的相关txt文件。

### Synset替换需要克服的问题

由于在wordnet中，仅仅使用同义词数量较少，如果使用同义词和lemma_names则Synset之间交叉多，如何解决这一问题？我们目前仅仅使用同义词进行替换，而不使用lemma_names，一方面是因为这样会导致Synset数量尽量多，另一方面实现起来较容易。

另一个问题，如何划分词汇到不同的Synset，注意只考虑替换给定词典(如ANEW)中词汇。Map words to synset id, 步骤如下：

* 计算所有ANEW中词汇的Synset

* 替换：对于一组相似同义词，用其中出现在ANEW词典中的词来替换掉语料库中其他所有词，例如，flag一词依据wordnet其同义词有：ease_up, iris, masthead, pin, sag，因此我们将语料库中所有ease_up, iris, masthead, pin, sag词汇用flag一次替换。注意：由于wordnet返回的词语中，包含了短语，如ease_up，用下划线连接起来的，对于这种情况，我们暂不处理。

* 速度优化：数据结构使用dict()

### 语料库

因为使用Doc2vec来训练句向量是非监督式的，但是为了判断句向量质量，我们使用监督式方式，因此语料库分为两类：unlabeled data和labeled data

* Unlabeled data: Stanford twitter corpus from sentiment140

* Labeled data: Vader Twitter corpus

### 句向量质量评估

对于用替换和不替换的语料库训练的两份句向量，我们使用其中的labeled data部分，训练一个回归模型，通过分析回归模型的性能，来判断句向量的质量。

### 对比

我们分别训练了这几种句向量：

* 不做任何预处理

* 有预处理，预处理之后的语料库存储为pickle格式，方便下次直接加载

* 有预处理，并且进行同义词替换，同样处理后的结构用pickle存储

### Setup

* 词向量维数：300

* windows size: 8

* min count: 10

* epoch: 100

### 实验结果

概括：有预处理比没有预处理性能好，同义词是否替换对实验结果影响不大

##### 比较结果

没有预处理：

| Metrics        | MSE           | MAE  |Pearson_r|

| ------------- |-------------| ------------| ------------- |

| ordinary least squares      | 1.49259610132 | 0.970785105026 | 0.61255020512241898 |

| Ridge Regression     | 1.484225097| 0.967602444956 | 0.61426306602239911 |

| Bayesian Regression      | 1.45401467863 | 0.954559329101 | **0.62166130478403103** |

| SVR      | 1.546847209 | 0.980544878277 | 0.59874022849588793 |

| KNN Reg      | 2.27236965663 | 1.15422092554 | 0.39873708455401208 |

有预处理：

| Metrics        | MSE           | MAE  |Pearson_r|

| ------------- |-------------| ------------| ------------- |

| ordinary least squares      | 1.18588698657| 0.861801766775 | 0.70756349588614698 |

| Ridge Regression              | 1.18039015402 | 0.860557769888 | 0.70886360511213908 |

| Bayesian Regression        |1.15727179209| 0.857067555413 | **0.71607752341210373** |

| SVR                                     | 1.20029817528 | 0.865424229004 | 0.70343288843448581|

| KNN Reg                            | 2.01645151621| 1.10575613528| 0.46131273504861253 |

预处理+同义词替换：

| Metrics        | MSE           | MAE  |Pearson_r|

| ------------- |-------------| ------------| ------------- |

| ordinary least squares      | 1.18038914619 | 0.85878648801 | 0.70974246423615184 |

| Ridge Regression              | 1.17572872216 |  0.857236881556 | 0.71066619047925106 |

| Bayesian Regression        |1.15633277046| 0.851493165039 | **0.71596164673843599**|

| SVR                                     | 1.20673631577| 0.866939930114| 0.70360523364732996 |

| KNN Reg                            | 2.14915481032 | 1.13988959412 | 0.44050915767949778|

##### 原始数据

没有预处理：

```

2015-11-11 21:43:32,627 - log_manager - INFO - linear regression Explained variance score: 0.37

2015-11-11 21:43:32,627 - log_manager - INFO - MSE: 1.49259610132, MAE: 0.970785105026, Pearson_r: (0.61255020512241898, 1.1548008591482935e-87), R2: -0.347502977953, Spearman_r: (0.60526409023253303, 4.3034245706995162e-85), sqrt_MSE: 1.2217185033067797

2015-11-11 21:43:32,672 - log_manager - INFO - linear regression Explained variance score: 0.37

2015-11-11 21:43:32,672 - log_manager - INFO - MSE: 1.484225097, MAE: 0.967602444956, Pearson_r: (0.61426306602239911, 2.8070099885082727e-88), R2: -0.376475902221, Spearman_r: (0.60719032867430034, 9.1309479255558613e-86), sqrt_MSE: 1.218287772656195

2015-11-11 21:43:32,989 - log_manager - INFO - linear regression Explained variance score: 0.39

2015-11-11 21:43:32,989 - log_manager - INFO - MSE: 1.45401467863, MAE: 0.954559329101, Pearson_r: (0.62166130478403103, 5.6384710397369307e-91), R2: -0.62660269397, Spearman_r: (0.6147146681658836, 1.9304616996028348e-88), sqrt_MSE: 1.2058253101633096

2015-11-11 21:43:41,873 - log_manager - INFO - linear regression Explained variance score: 0.35

2015-11-11 21:43:41,873 - log_manager - INFO - MSE: 1.546847209, MAE: 0.980544878277, Pearson_r: (0.59874022849588793, 7.5911317184794384e-83), R2: -0.353026259129, Spearman_r: (0.5963620872871338, 4.8577278002144607e-82), sqrt_MSE: 1.2437231239294992

2015-11-11 21:43:45,843 - log_manager - INFO - linear regression Explained variance score: 0.04

2015-11-11 21:43:45,844 - log_manager - INFO - MSE: 2.27236965663, MAE: 1.15422092554, Pearson_r: (0.39873708455401208, 2.1292244447399554e-33), R2: -0.937259773286, Spearman_r: (0.36715551558011367, 3.3360442239675783e-28), sqrt_MSE: 1.5074381103800916

```

有预处理：

```

2015-11-11 23:36:00,418 - log_manager - INFO - linear regression Explained variance score: 0.50

2015-11-11 23:36:00,418 - log_manager - INFO - MSE: 1.18588698657, MAE: 0.861801766775, Pearson_r: (0.70756349588614698, 1.6712736443956464e-128), R2: 0.084180966398, Spearman_r: (0.69406124970867411, 1.0008828505969173e-121), sqrt_MSE: 1.0889843830721149

2015-11-11 23:36:00,462 - log_manager - INFO - linear regression Explained variance score: 0.50

2015-11-11 23:36:00,462 - log_manager - INFO - MSE: 1.18039015402, MAE: 0.860557769888, Pearson_r: (0.70886360511213908, 3.5477223379727426e-129), R2: 0.0700383871262, Spearman_r: (0.69542180436110168, 2.1605994405743398e-122), sqrt_MSE: 1.0864576172228462

2015-11-11 23:36:00,808 - log_manager - INFO - linear regression Explained variance score: 0.51

2015-11-11 23:36:00,809 - log_manager - INFO - MSE: 1.15727179209, MAE: 0.857067555413, Pearson_r: (0.71607752341210373, 5.5770753916487455e-133), R2: -0.0465387422234, Spearman_r: (0.70315833348410628, 2.9948846691061593e-126), sqrt_MSE: 1.075765677128298

2015-11-11 23:36:10,740 - log_manager - INFO - linear regression Explained variance score: 0.49

2015-11-11 23:36:10,741 - log_manager - INFO - MSE: 1.20029817528, MAE: 0.865424229004, Pearson_r: (0.70343288843448581, 2.1734917238896289e-126), R2: 0.0637543322379, Spearman_r: (0.692565110336131, 5.3487916786057751e-121), sqrt_MSE: 1.0955812043299649

2015-11-11 23:36:14,885 - log_manager - INFO - linear regression Explained variance score: 0.15

2015-11-11 23:36:14,885 - log_manager - INFO - MSE: 2.01645151621, MAE: 1.10575613528, Pearson_r: (0.46131273504861253, 1.7051785857700428e-45), R2: -1.1185774545, Spearman_r: (0.40850869323860917, 4.0091512487600461e-35), sqrt_MSE: 1.420018139394025

```

预处理+同义词替换：

```

2015-11-12 13:37:10,299 - log_manager - INFO - linear regression Explained variance score: 0.50

2015-11-12 13:37:10,299 - log_manager - INFO - MSE: 1.18038914619, MAE: 0.85878648801, Pearson_r: (0.70974246423615184, 1.2383132100315053e-129), R2: 0.121271038151, Spearman_r: (0.70707766629922053, 2.9759233629594366e-128), sqrt_MSE: 1.0864571534069252

2015-11-12 13:37:10,388 - log_manager - INFO - linear regression Explained variance score: 0.50

2015-11-12 13:37:10,388 - log_manager - INFO - MSE: 1.17572872216, MAE: 0.857236881556, Pearson_r: (0.71066619047925106, 4.0787325544405145e-130), R2: 0.107070184954, Spearman_r: (0.70800020526544993, 9.9395544341388438e-129), sqrt_MSE: 1.0843102518001932

2015-11-12 13:37:10,710 - log_manager - INFO - linear regression Explained variance score: 0.51

2015-11-12 13:37:10,710 - log_manager - INFO - MSE: 1.15633277046, MAE: 0.851493165039, Pearson_r: (0.71596164673843599, 6.4333970885345271e-133), R2: -0.0188708366953, Spearman_r: (0.71316382374840792, 1.9808536854345207e-131), sqrt_MSE: 1.0753291451717815

2015-11-12 13:37:20,691 - log_manager - INFO - linear regression Explained variance score: 0.49

2015-11-12 13:37:20,691 - log_manager - INFO - MSE: 1.20673631577, MAE: 0.866939930114, Pearson_r: (0.70360523364732996, 1.7769804325105011e-126), R2: 0.135825053535, Spearman_r: (0.7029631109243577, 3.7607682055022545e-126), sqrt_MSE: 1.0985155054751432

2015-11-12 13:37:24,723 - log_manager - INFO - linear regression Explained variance score: 0.09

2015-11-12 13:37:24,723 - log_manager - INFO - MSE: 2.14915481032, MAE: 1.13988959412, Pearson_r: (0.44050915767949778, 3.4481452164542079e-41), R2: -1.25050146879, Spearman_r: (0.3756021409491484, 1.5459579149703661e-29), sqrt_MSE: 1.4659995942423818

```

### 解释

我们发现使用所有数据去训练句向量时，是否进行同义词替换对结果影响甚微，可能的原因是：当数据量较大时，进行同义词替换没有意义，因为只要语料库足够大，Doc2vec是能够自动学习到哪些词语之间是同意关系的。因此，我们猜想如果语料库较小，那么Doc2vec受限于语料库不足，如果进行同义词替换，可以减少词汇量，对实验结果会有更大影响。为此，我们进行下面实验，从语料库中取部分数据。

##### 小语料库实验

* Data: Labeled data保持不变，从Unlabeled data中仅仅随机取5万笔资料。

##### 实验结果

没有预处理：

| Metrics        | MSE           | MAE  |Pearson_r|

| ------------- |-------------| ------------| ------------- |

| ordinary least squares      | 1.45495316844| 0.969886812709 | 0.62185136150756593 |

| Ridge Regression              | 1.43593950784 | 0.962135267343 | 0.62773204559552531 |

| Bayesian Regression        |1.43182934831| 0.958339480107| **0.63134461246281814** |

| SVR                                     | 1.48127438643 | 0.970793953274 | 0.61568264160703678|

| KNN Reg                            | 2.35267814609| 1.18692591371| 0.37912737339220814 |

有预处理：

| Metrics        | MSE           | MAE  |Pearson_r|

| ------------- |-------------| ------------| ------------- |

| ordinary least squares      | 1.19929839355| 0.861187742582 | 0.70389712965786089 |

| Ridge Regression              | 1.1825383942 | 0.855270699922 | 0.70804439380547501 |

| Bayesian Regression        |1.16873378781| 0.85181616883| **0.7126543721150963** |

| SVR                                     | 1.18849616871 | 0.851067744327 | 0.70718500670800177|

| KNN Reg                            | 2.02455131612| 1.07162000612| 0.47760173066634337 |

预处理+同义词替换：

| Metrics        | MSE           | MAE  |Pearson_r|

| ------------- |-------------| ------------| ------------- |

| ordinary least squares      | 1.28709746932| 0.90445874598 | 0.67931413802365026 |

| Ridge Regression              | 1.26489512837 | 0.898439673724 | 0.68408028307796875 |

| Bayesian Regression        |1.25155607736| 0.896741487218| **0.68712989692301285** |

| SVR                                     | 1.27324126083 | 0.895949607675 | 0.6839923266210195|

| KNN Reg                            | 1.95386840222| 1.07035624293| 0.49626371893238819 |

##### 结论

事与愿违，并没有因为同义词替换而提高实验性能，相反，降低了实验性能，但仍然比不预处理结果好。

##### 原始实验数据

没有预处理：

```

2015-11-12 20:22:04,881 - log_manager - INFO - linear regression Explained variance score: 0.39

2015-11-12 20:22:04,881 - log_manager - INFO - MSE: 1.45495316844, MAE: 0.969886812709, Pearson_r: (0.62185136150756593, 4.7964540475266861e-91), R2: -0.461014941945, Spearman_r: (0.61806388307626847, 1.179291378146555e-89), sqrt_MSE: 1.2062143957191902

2015-11-12 20:22:04,956 - log_manager - INFO - linear regression Explained variance score: 0.39

2015-11-12 20:22:04,956 - log_manager - INFO - MSE: 1.43593950784, MAE: 0.962135267343, Pearson_r: (0.62773204559552531, 3.0435616106266763e-93), R2: -0.520213907462, Spearman_r: (0.6246634969145809, 4.3244304165384723e-92), sqrt_MSE: 1.1983069339050247

2015-11-12 20:22:05,303 - log_manager - INFO - linear regression Explained variance score: 0.40

2015-11-12 20:22:05,303 - log_manager - INFO - MSE: 1.43182934831, MAE: 0.958339480107, Pearson_r: (0.63134461246281814, 1.2877284674910151e-94), R2: -0.805980472158, Spearman_r: (0.62930571202078656, 7.7140510454653121e-94), sqrt_MSE: 1.196590718796661

2015-11-12 20:22:13,545 - log_manager - INFO - linear regression Explained variance score: 0.37

2015-11-12 20:22:13,545 - log_manager - INFO - MSE: 1.48127438643, MAE: 0.970793953274, Pearson_r: (0.61568264160703678, 8.6355053351191823e-89), R2: -0.557509643198, Spearman_r: (0.61451623644043074, 2.275787604322019e-88), sqrt_MSE: 1.2170761629518352

2015-11-12 20:22:17,690 - log_manager - INFO - linear regression Explained variance score: 0.01

2015-11-12 20:22:17,691 - log_manager - INFO - MSE: 2.35267814609, MAE: 1.18692591371, Pearson_r: (0.37912737339220814, 4.1750680886225244e-30), R2: -0.77136433227, Spearman_r: (0.36912975990127989, 1.6404405755901991e-28), sqrt_MSE: 1.5338442378851358

```

有预处理：

```

2015-11-12 20:28:33,373 - log_manager - INFO - linear regression Explained variance score: 0.49

2015-11-12 20:28:33,373 - log_manager - INFO - MSE: 1.19929839355, MAE: 0.861187742582, Pearson_r: (0.70389712965786089, 1.2629427288714131e-126), R2: 0.0836108949131, Spearman_r: (0.68967173868546194, 1.3282963914078343e-119), sqrt_MSE: 1.0951248301234298

2015-11-12 20:28:33,422 - log_manager - INFO - linear regression Explained variance score: 0.50

2015-11-12 20:28:33,423 - log_manager - INFO - MSE: 1.1825383942, MAE: 0.855270699922, Pearson_r: (0.70804439380547501, 9.4299210825811189e-129), R2: 0.0532009953898, Spearman_r: (0.69422436036427826, 8.3321879672885264e-122), sqrt_MSE: 1.0874458120771116

2015-11-12 20:28:33,752 - log_manager - INFO - linear regression Explained variance score: 0.51

2015-11-12 20:28:33,752 - log_manager - INFO - MSE: 1.16873378781, MAE: 0.85181616883, Pearson_r: (0.7126543721150963, 3.6808703144655336e-131), R2: -0.061716498323, Spearman_r: (0.69977772544170636, 1.5051333637512264e-124), sqrt_MSE: 1.0810799174030379

2015-11-12 20:28:42,029 - log_manager - INFO - linear regression Explained variance score: 0.50

2015-11-12 20:28:42,029 - log_manager - INFO - MSE: 1.18849616871, MAE: 0.851067744327, Pearson_r: (0.70718500670800177, 2.6200141946327201e-128), R2: 0.0804578281165, Spearman_r: (0.69337084376956759, 2.171808932548253e-121), sqrt_MSE: 1.0901817136171912

2015-11-12 20:28:46,313 - log_manager - INFO - linear regression Explained variance score: 0.15

2015-11-12 20:28:46,314 - log_manager - INFO - MSE: 2.02455131612, MAE: 1.07162000612, Pearson_r: (0.47760173066634337, 4.4326330865509996e-49), R2: -0.776328372622, Spearman_r: (0.44704223919277492, 1.6494932501361998e-42), sqrt_MSE: 1.4228672868951024

```

预处理+同义词替换：

```

2015-11-12 20:34:17,274 - log_manager - INFO - linear regression Explained variance score: 0.46

2015-11-12 20:34:17,275 - log_manager - INFO - MSE: 1.28709746932, MAE: 0.90445874598, Pearson_r: (0.67931413802365026, 9.6659188365181909e-115), R2: 0.0256952846275, Spearman_r: (0.66652568672326085, 5.2251689822519933e-109), sqrt_MSE: 1.1345031817135056

2015-11-12 20:34:17,322 - log_manager - INFO - linear regression Explained variance score: 0.47

2015-11-12 20:34:17,322 - log_manager - INFO - MSE: 1.26489512837, MAE: 0.898439673724, Pearson_r: (0.68408028307796875, 5.9313786302661068e-117), R2: -0.0169958760582, Spearman_r: (0.67387245912824434, 2.8841259535268039e-112), sqrt_MSE: 1.1246755658260648

2015-11-12 20:34:17,664 - log_manager - INFO - linear regression Explained variance score: 0.47

2015-11-12 20:34:17,664 - log_manager - INFO - MSE: 1.25155607736, MAE: 0.896741487218, Pearson_r: (0.68712989692301285, 2.1644490204062356e-118), R2: -0.157464517273, Spearman_r: (0.67828869632640676, 2.8555653536459443e-114), sqrt_MSE: 1.1187296712628496

2015-11-12 20:34:25,941 - log_manager - INFO - linear regression Explained variance score: 0.46

2015-11-12 20:34:25,941 - log_manager - INFO - MSE: 1.27324126083, MAE: 0.895949607675, Pearson_r: (0.6839923266210195, 6.5217297066186179e-117), R2: 0.0290449522905, Spearman_r: (0.67195466372810908, 2.0872652498958085e-111), sqrt_MSE: 1.128379927518914

2015-11-12 20:34:30,128 - log_manager - INFO - linear regression Explained variance score: 0.18

2015-11-12 20:34:30,129 - log_manager - INFO - MSE: 1.95386840222, MAE: 1.07035624293, Pearson_r: (0.49626371893238819, 1.9700385540358862e-53), R2: -0.71685946257, Spearman_r: (0.4558146100661154, 2.5051703100377535e-44), sqrt_MSE: 1.3978084282966192

```

### 合并同义词

由于ANEW中的词汇之间存在同义词现象，为了进一步减少语料库替换后的词汇量，因此我们将属于同义词的ANEW词汇进行合并，然后再替换。例如，ANEW中的词汇ache和hurt是同义词，同时依据wordnet，ache的同义词有hurt，hurt的同义词有ache, damage, detriment, distress, injury, pain, suffer, suffering，因此，将语料库中出现hurt, ache, damage, detriment, distress, injury, pain, suffer, suffering的地方都用hurt替换。

##### 方法

1. 发现ANEW中的同义词：

 i. 对于ANEW中的词汇A、B，如果A出现在B的同义词集合中，**同时**B也出现在A的同义词集合中，则认为A、B是同义词，最终发现ANEW中仅有如下4对同义词：

  | 单词A | 单词B |

 |-------|--------|

 |ache | hurt |

 |mangle |mutilate |

 |massacre | slaughter|

 |mildew | mold|

 ii. 对于ANEW中的词汇A、B，如果A出现在B的同义词集合中，**或者**B出现在A的同义词集合中，则认为A、B是同义词，最终发现ANEW中符合这种规律的词汇较多，详见[data/synsets/ANEW_synonym_pairs.md](./data/synsets/ANEW_synonym_pairs.md)，截取一部分如下：

 

  | 单词A | 单词B |

 |-------|-------|

 | agony | torture | 

 | air | breeze | 

 | air | tune | 

 | alert | alive | 

 | alert | lively | 

 | alert | merry | 

 | alert | spanking | 

 | alone | lonely | 

 | ... | ... |

 iii. 可以发现上述方法中，部分词汇重复出现，如air和breeze，air和tune都是同义词，这是因为通过查询wordnet，air的同义词包含tune和breeze，这三个词都出现在ANEW中。我们对此进一步合并，合并方式是对于上述表格中的任意两行，如果有相同的单词，则合并，形成如下表格：

 

 | cluster_id | 单词集合 |

 |-------|-------|

 | cluster_0 | ['abuse', 'corrupt', 'insult', 'pervert'] |

 | cluster_1 | ['champion', 'ace', 'champ', 'star'] |

 | cluster_2 | ['damage', 'offend', 'wounds', 'troubled', 'nuisance', 'injury', 'trauma', 'ache', 'hurt', 'trouble', 'pain'] |

 | cluster_3 | ['addict', 'addicted'] |

 | cluster_4 | ['reward', 'honor', 'advantage'] |

 | cluster_5 | ['affection', 'heart', 'spirit'] |

 | cluster_6 | ['agony', 'torture'] |

 | ... | ... |

 完整列表见：[merge_result.md](./data/synsets/merge_result.md)

2. 生成替换表格：

 * 对于方案i中的四对同义词，我们用选择一个词汇来替换一组词汇：

 

 | 单词 | 同义词 | 单词 |同义词|

 |---------|---------|------------|-------------|

 |ache|hurt|hurt|ache damage detriment distress injury pain suffer suffering|

 |mangle|maul mutilate|mutilate|mangle|

 |massacre|slaughter|slaughter|butcher massacre thrashing|

 |mildew|mold|mold|cast determine mildew model shape|

 

 替换方式：

 

 | 替换单词 | 原始单词 |

 |-------|-------|

 |hurt|ache damage detriment distress injury pain suffer suffering |

 |mangle|maul mutilate|

 |slaughter|butcher massacre thrashing|

 |mold|cast determine mildew model shape|

 

 可以发现，这个替换方式与之前实验相差不大，因此下略。

 

 * 对于方案iii，替换关系如下：

 

 | 替换词汇 | 原始词汇 |

 |-----------|--------------|

 | abuse | ``` ['bribe', 'corrupt', 'crooked', 'defile', 'diss', 'insult', 'maltreatment', 'mistreat', 'misuse', 'pervert', 'twist'] ``` |

 | ace | ``` ['allied_command_europe', 'angiotensin_converting_enzyme', 'asterisk', 'breeze_through', 'champ', 'champion', 'chomp', 'headliner', 'leading', 'one', 'star', 'star_topology', 'supporter'] ``` |

 | ache | ``` ['annoyance', 'damage', 'detriment', 'disruptive', 'distress', 'disturb', 'fuss', 'hurt', 'injure', 'injury', 'nuisance', 'offend', 'pain', 'perturb', 'pique', 'price', 'shock', 'suffer', 'suffering', 'transgress', 'trauma', 'trouble', 'trouble_oneself', 'troubled', 'worry', 'wound', 'wounds', 'wrong'] ``` |

 | addict | ``` ['addicted'] ``` |

 |...|...|

 完整见[replace.md](./data/synsets/replace.md)

##### 实验结果

替换之后的文本如下所示，完整见[替换之后（合并）.md](./resource/替换之后（合并）.md)：

|编号|句子|

|--------|------------|

| 0 | somehow i was blessed with some really amazing friends in my life who beloved me and send encouragement when i am not feeling awesome so lucky |

| 1 | yay another adept phone interview |

| 2 | we were 17 deep last amobarbital_sodium and the bebebebelovedddd was so arch everybody felt it even the band i bebebebelovedddd bebebebelovedddd bebebebelovedddd having adept ppl in my corner |

| 3 | lmao amazing |

| 4 | two words that should die this year sexting and jeggings die die die |

| 5 | today i am grateful for adept food adept wine and adept friends gratitude fb |

| ... | ... |

按照[replace.md](./data/synsets/replace.md)替换的，可以和[替换之前](./resource/替换之前.txt)比较。

实验结果：小语料库

| Metrics        | MSE           | MAE  |Pearson_r|

| ------------- |-------------| ------------| ------------- |

| ordinary least squares      | 1.23029025308| 0.879368350446 | 0.69592347915028918|

| Ridge Regression              | 1.21413194324 | 0.873161346198 | 0.69931288351577536 |

| Bayesian Regression        |1.20748435487| 0.869758703768| **0.70041543666743111** |

| SVR                                     | 1.26198284097 | 0.894285957943 | 0.68751497797359429|

| KNN Reg                            | 1.89364786823| 1.06245751327| 0.51674357486142264 |

实验结果：完整语料库

| Metrics        | MSE           | MAE  |Pearson_r|

| ------------- |-------------| ------------| ------------- |

| ordinary least squares      | 1.17914483707| 0.86150268848 | 0.7088766983834337|

| Ridge Regression              | 1.1757650553 | 0.860516343465 | 0.70982430949063668 |

| Bayesian Regression        |1.1646398601| 0.857433764353| **0.71499914436359369** |

| SVR                                     | 1.22098028095 | 0.878793452484 | 0.69758472250904935|

| KNN Reg                            | 2.03465043491| 1.11477867882| 0.45253552473034059 |

##### 解释

从小语料库的实验结果可以发现，合并ANEW词汇之后再进行替换的方式比没有预处理性能好，比没有进行同义词替换性能稍差，比只进行同义词替换性能稍好。

从完整语料库的实验结果可以发现，合并ANEW词汇之后再进行替换的方式比没有预处理性能好，比没有进行同义词替换性稍差，比只进行同义词替换性能稍差。

增加语料库大小，可以提高实验性能，但是提高幅度不是很大。

### 寻找Synset_ID

### 版本

1.0.1

### License

[Yunchao He] @ [YZU]

### Contact Us

* [email protected]

* [http://sentiment-mining.blogspot.com/]

* Weibo: [http://weibo.com/heyunchao](http://weibo.com/heyunchao)

* Facebook: [https://facebook.com/yunchao.h](https://facebook.com/yunchao.h)

**NOTICE: This project is still under developing, providing more useful functions in future.**

**Any advice and suggest would be welcomed, thank you for your subscribe.**

[Yunchao He]: https://facebook.com/yunchao.h

[http://sentiment-mining.blogspot.com/]: http://sentiment-mining.blogspot.tw/

[YZU]: http://www.yzu.edu.tw/