本项目是硕士毕业论文中部分补充实验
-
CVAT 2.0 数据统计
-
CVAT 使用word2vec词向量,CNN预测结果
-
CVAT 使用随机词向量,CNN预测结果
-
CVAT 使用不同维度词向量预测结果
- CVAT 2.0 数据统计
语料库大小 | 文字总数 | 平均句子长度 | 词汇量 | 标记维度 | 句子最大长度 | 句子最小长度 |
---|---|---|---|---|---|---|
2,009 | 111,558 | 55.52 | 14,708 | V+A | 247 | 4 |
VA 分布图
- 使用CNN、word2vec预测va值
Hyper-parameters:
N_fm = 100, kernel_size = 5, maxlen = 100, hidden_dims = 100, dropout_p = 0.5, pool_length = pool_length=math.floor((maxlen-kernel_size+1)/2), batch_size = 8, max_nb_epoch=50
Experiement result:
Metrics | MSE | MAE | pearson_r | R2 | Spearman_r | rMSE |
---|---|---|---|---|---|---|
Valence | 1.189 | 0.874 | 0.639 | 0.374 | 0.662 | 1.09 |
Arousal | 0.773 | 0.716 | 0.387 | 0.117 | 0.359 | 0.879 |
Valence
部分预测Valence预测图
valence迭代过程
提前停止防止过拟合,early_stopping (论文中解释下)
valence直方图,绝对误差和频率
Arousal
部分预测值和实际值之间的关系
Arousal迭代过程
- 使用随机词向量预测CVAT2.0 VA值
Metrics | MSE | MAE | pearson_r | R2 | Spearman_r | rMSE |
---|---|---|---|---|---|---|
Valence | 1.227 | 0.896 | 0.611 | 0.355 | 0.636 | 1.108 |
Arousal | 1.042 | 0.827 | 0.292 | -0.190 | 0.250 | 1.021 |
Valence
Arousal
- 随机产生词向量,不同维度对实验(Valence)的影响
10维
prediction result:
MSE, MAE, Pearson_r, R2, Spearman_r, MSE_sqrt
1.25836214085 0.910268113036 (0.60011948439809737, 1.0890142012958216e-40) 0.338585582905 SpearmanrResult(correlation=0.63293688023783556, pvalue=2.1862829634462172e-46) 1.1217674183395847
50维
prediction result:
MSE, MAE, Pearson_r, R2, Spearman_r, MSE_sqrt
1.3131630047 0.931548146482 (0.59641990871786488, 4.3503589427249966e-40) 0.309781409411 SpearmanrResult(correlation=0.63282173346108705, pvalue=2.2956331299069574e-46) 1.1459332461809077
100维
prediction result:
MSE, MAE, Pearson_r, R2, Spearman_r, MSE_sqrt
1.30339259081 0.917410494447 (0.59833307930982149, 2.1303463971132947e-40) 0.314916888616 SpearmanrResult(correlation=0.61738254682301019, pvalue=1.3300052107696708e-43) 1.1416622052139032
200维
prediction result:
MSE, MAE, Pearson_r, R2, Spearman_r, MSE_sqrt
1.23475371152 0.882428248917 (0.62016164255507433, 4.3436508709342757e-44) 0.350994535 SpearmanrResult(correlation=0.64549360851831772, pvalue=9.4037467810345631e-49) 1.111194722592499
- RMV方法实验结果
Valence
prediction result:
MSE, MAE, Pearson_r, R2, Spearman_r, MSE_sqrt
1.2835520318 0.94156873862 (0.57191450528400101, 2.7101874320614917e-36) 0.325345390358 SpearmanrResult(correlation=0.62216714491239822, pvalue=1.9235418374036149e-44) 1.1329395534605917
Arousal
prediction result:
MSE, MAE, Pearson_r, R2, Spearman_r, MSE_sqrt
0.853916672335 0.752925855315 (0.275446058880638, 1.971860439753554e-08) 0.0241592254039 SpearmanrResult(correlation=0.25269012595517543, pvalue=2.8334554626819371e-07) 0.9240761182578903
- RMV-rand实验结果
Valence
prediction result:
MSE, MAE, Pearson_r, R2, Spearman_r, MSE_sqrt
1.2889382787 0.938735591411 (0.57212649720456332, 2.5207511400979108e-36) 0.322514296476 SpearmanrResult(correlation=0.62359280247652293, pvalue=1.074145807226024e-44) 1.1353141762099406
Arousal
prediction result:
MSE, MAE, Pearson_r, R2, Spearman_r, MSE_sqrt
0.862190900749 0.756158571479 (0.27546894983520076, 1.9663319818446003e-08) 0.0147035844416 SpearmanrResult(correlation=0.2557304787319743, pvalue=2.0133710378673852e-07) 0.9285423526954456
- 使用GloVe词向量
Valence
prediction result:
MSE, MAE, Pearson_r, R2, Spearman_r, MSE_sqrt
1.1164321125 0.867102129429 (0.68690511510825092, 2.0695044715449238e-57) 0.413186179919 SpearmanrResult(correlation=0.70279031194106056, pvalue=3.9585791260288134e-61) 1.056613511412774
Arousal
prediction result:
MSE, MAE, Pearson_r, R2, Spearman_r, MSE_sqrt
1.03420569678 0.825181955229 (0.32846545402454846, 1.4433020927882028e-11) -0.181871862839 SpearmanrResult(correlation=0.29679482897748571, pvalue=1.2832242048569735e-09) 1.0169590438083371
Valence
prediction result:
MSE, MAE, Pearson_r, R2, Spearman_r, MSE_sqrt
1.23910091076 0.924262162726 (0.59081545611658148, 3.4276731557030528e-39) 0.348709580488 SpearmanrResult(correlation=0.64842236586078461, pvalue=2.5424448943081574e-49) 1.113149096373635
Arousal
prediction result:
MSE, MAE, Pearson_r, R2, Spearman_r, MSE_sqrt
0.807147009032 0.733752654548 (0.3155619729987712, 9.5800508818236855e-11) 0.0776067641902 SpearmanrResult(correlation=0.28526844931242679, pvalue=5.771686920239029e-09) 0.8984136068826754
- Additional Experiment
The above experiment use sigma=1 to exclude the outlier. sigma = 1 is the best way? Maybe we can use sigma = 1.5 or 2.0 as the outlier selection mechanism.
- sigma=1.5
Valence-word2vec
prediction result:
MSE, MAE, Pearson_r, R2, Spearman_r, MSE_sqrt
1.17501734133 0.857153185261 (0.64596081596608124, 7.6401374203288254e-49) 0.389370319799 SpearmanrResult(correlation=0.678238647820257, pvalue=1.7600146928857212e-55) 1.0839821683646538
Arousal-word2vec
prediction result:
MSE, MAE, Pearson_r, R2, Spearman_r, MSE_sqrt
0.921457808986 0.767747403198 (0.34734165685328955, 7.6746875625596964e-13) -0.115481114959 SpearmanrResult(correlation=0.31769903648833775, pvalue=7.0456692082057659e-11) 0.9599259393233046
- sigma=2.0
Valence-word2ve
prediction result:
MSE, MAE, Pearson_r, R2, Spearman_r, MSE_sqrt
1.23509052786 0.88814520616 (0.62644756090465481, 3.3145568313467529e-45) 0.358578117496 SpearmanrResult(correlation=0.65623542877526542, pvalue=7.2243601114955732e-51) 1.1113462682094133
Arosual-word2ve
prediction result:
MSE, MAE, Pearson_r, R2, Spearman_r, MSE_sqrt
0.816087390281 0.732002397982 (0.33742783810355886, 3.6739148370540297e-12) 0.0124040738807 SpearmanrResult(correlation=0.31302769736912855, pvalue=1.3748173925512652e-10) 0.9033755532896222
sigma | MSE | MAE | r |
---|---|---|---|
Valence(sigma=1.0) | 1.189 | 0.874 | 0.639 |
Valence(sigma=1.5) | 1.175 | 0.857 | 0.646 |
Valence(sigma=2.0) | 1.235 | 0.888 | 0.626 |
Arousal(sigma=1.0) | 0.773 | 0.716 | 0.387 |
Arousal(sigma=1.5) | 0.921 | 0.768 | 0.347 |
Arousal(sigma=2.0) | 0.816 | 0.732 | 0.337 |
RMV method:
sigma = 1.5
prediction result:
MSE, MAE, Pearson_r, R2, Spearman_r, MSE_sqrt
1.25285167125 0.925782175414 (0.59217695709096529, 2.0836660810940871e-39) 0.348921595928 SpearmanrResult(correlation=0.64716559558699882, pvalue=4.4645167616700477e-49) 1.119308568379042
prediction result:
MSE, MAE, Pearson_r, R2, Spearman_r, MSE_sqrt
0.811448281476 0.725746071187 (0.27298738031183734, 2.6618926487311099e-08) 0.0176921559236 SpearmanrResult(correlation=0.2505769297144147, pvalue=3.5837527768889621e-07) 0.9008042414840945
sigma = 2.0
prediction result:
MSE, MAE, Pearson_r, R2, Spearman_r, MSE_sqrt
1.25824152441 0.925569502635 (0.58972330837386944, 5.1011083732185485e-39) 0.346555066994 SpearmanrResult(correlation=0.64854313773014549, pvalue=2.4082009431610321e-49) 1.1217136552679012
prediction result:
MSE, MAE, Pearson_r, R2, Spearman_r, MSE_sqrt
0.80917985945 0.723266230294 (0.27168756180052012, 3.115749106640244e-08) 0.0207632880898 SpearmanrResult(correlation=0.24893271210102805, pvalue=4.2962166603105685e-07) 0.8995442509681696
sigma | MSE | MAE | r |
---|---|---|---|
Valence(sigma=1.0) | 1.284 | 0.942 | 0.572 |
Valence(sigma=1.5) | 1.253 | 0.926 | 0.592 |
Valence(sigma=2.0) | 1.258 | 0.926 | 0.589 |
Arousal(sigma=1.0) | 0.854 | 0.753 | 0.275 |
Arousal(sigma=1.5) | 0.811 | 0.726 | 0.273 |
Arousal(sigma=2.0) | 0.809 | 0.723 | 0.272 |
- Worst result
Let all the predicted valence and arousal be 5, and the performance is:
sigma | MSE | MAE | r |
---|---|---|---|
Valence(sigma=1.0) | 1.945 | 1.192 | nan |
Valence(sigma=1.5) | 1.971 | 0.204 | nan |
Valence(sigma=2.0) | 1.968 | 1.204 | nan |
Arousal(sigma=1.0) | 0.898 | 0.776 | nan |
Arousal(sigma=1.5) | 0.830 | 0.731 | nan |
Arousal(sigma=2.0) | 0.831 | 0.733 | nan |
- Tradictional Method
sigma = 1.0 using CVAW lexicon
Dims | Mean Method | Regression | MSE | MAE | r |
---|---|---|---|---|---|
Valence | Geometric | False | 2.162 | 1.149 | 0.527 |
Valence | Arithmetic | False | 2.227 | 1.169 | 0.520 |
Valence | Geometric | True | 1.434 | 0.988 | 0.506 |
Valence | Arithmetic | True | 1.457 | 0.997 | 0.494 |
Arousal | Geometric | False | 1.517 | 0.982 | 0.134 |
Arousal | Arithmetic | False | 1.932 | 1.109 | 0.147 |
Arousal | Geometric | True | 0.848 | 0.753 | 0.184 |
Arousal | Arithmetic | True | 0.839 | 0.751 | 0.216 |
sigma = 1.5 using CVAW lexicon
Dims | Mean Method | Regression | MSE | MAE | r |
---|---|---|---|---|---|
Valence | Geometric | False | 2.162 | 1.152 | 0.528 |
Valence | Arithmetic | False | 2.230 | 1.175 | 0.521 |
Valence | Geometric | True | 1.460 | 0.984 | 0.503 |
Valence | Arithmetic | True | 1.486 | 0.993 | 0.489 |
Arousal | Geometric | False | 1.647 | 1.011 | 0.140 |
Arousal | Arithmetic | False | 1.656 | 1.016 | 0.147 |
Arousal | Geometric | True | 0.794 | 0.724 | 0.213 |
Arousal | Arithmetic | True | 0.793 | 0.724 | 0.217 |
sigma = 2.0 using CVAW lexicon
Dims | Mean Method | Regression | MSE | MAE | r |
---|---|---|---|---|---|
Valence | Geometric | False | 2.143 | 1.150 | 0.531 |
Valence | Arithmetic | False | 2.210 | 1.172 | 0.524 |
Valence | Geometric | True | 1.456 | 0.979 | 0.505 |
Valence | Arithmetic | True | 1.480 | 0.990 | 0.492 |
Arousal | Geometric | False | 1.646 | 1.012 | 0.140 |
Arousal | Arithmetic | False | 1.656 | 1.016 | 0.147 |
Arousal | Geometric | True | 0.795 | 0.726 | 0.212 |
Arousal | Arithmetic | True | 0.793 | 0.725 | 0.216 |
Extended Lexicon
sigma = 1.0 using extended CVAW lexicon (Neural_cand)
Dims | Mean Method | Regression | MSE | MAE | r |
---|---|---|---|---|---|
Valence | Geometric | False | 1.630 | 1.024 | 0.507 |
Valence | Arithmetic | False | 1.682 | 1.039 | 0.504 |
Valence | Geometric | True | 1.466 | 0.987 | 0.486 |
Valence | Arithmetic | True | 1.490 | 0.998 | 0.474 |
Arousal | Geometric | False | 1.517 | 0.982 | 0.134 |
Arousal | Arithmetic | False | 1.543 | 0.990 | 0.141 |
Arousal | Geometric | True | 0.848 | 0.753 | 0.184 |
Arousal | Arithmetic | True | 0.845 | 0.751 | 0.193 |
sigma = 1.5 using CVAW lexicon
Dims | Mean Method | Regression | MSE | MAE | r |
---|---|---|---|---|---|
Valence | Geometric | False | 1.625 | 1.027 | 0.511 |
Valence | Arithmetic | False | 1.678 | 1.044 | 0.508 |
Valence | Geometric | True | 1.481 | 0.991 | 0.488 |
Valence | Arithmetic | True | 1.509 | 1.002 | 0.475 |
Arousal | Geometric | False | 1.273 | 0.892 | 0.136 |
Arousal | Arithmetic | False | 1.286 | 0.897 | 0.143 |
Arousal | Geometric | True | 0.801 | 0.726 | 0.183 |
Arousal | Arithmetic | True | 0.798 | 0.726 | 0.191 |
sigma = 2.0 using CVAW lexicon
Dims | Mean Method | Regression | MSE | MAE | r |
---|---|---|---|---|---|
Valence | Geometric | False | 1.619 | 1.026 | 0.511 |
Valence | Arithmetic | False | 1.671 | 1.041 | 0.508 |
Valence | Geometric | True | 1.483 | 0.991 | 0.487 |
Valence | Arithmetic | True | 1.510 | 1.002 | 0.474 |
Arousal | Geometric | False | 1.274 | 0.891 | 0.135 |
Arousal | Arithmetic | False | 1.287 | 0.896 | 0.142 |
Arousal | Geometric | True | 0.802 | 0.728 | 0.180 |
Arousal | Arithmetic | True | 0.800 | 0.727 | 0.188 |
从上述实验可以发现:
- 使用回归时MSE、MSE结果更好
- 不使用回归时的Valence的相关系数更好
上述实验在不使用回归时MSE和MAE指标已经超出最坏实验的结果,使用回归总体上会更好,另一方面可以发现几何平均值的方法都比算数平均值在预测Valence结果好,而在预测arousal时算数平均值结果更好。
Valence-Arousal 在幾何平均值和算數平均值+迴歸方式下試驗結果
Dims | sigma | Mean Method | Regression | MSE | MAE | r |
---|---|---|---|---|---|---|
Valence | 1.0 | Geometric | True | 1.434 | 0.988 | 0.506 |
Valence | 1.5 | Geometric | True | 1.460 | 0.984 | 0.503 |
Valence | 2.0 | Geometric | True | 1.456 | 0.979 | 0.505 |
Arousal | 1.0 | Arithmetic | True | 0.839 | 0.751 | 0.216 |
Arousal | 1.5 | Arithmetic | True | 0.793 | 0.724 | 0.217 |
Arousal | 2.0 | Arithmetic | True | 0.793 | 0.725 | 0.216 |
另外,可以发现不同标准差情况下得到的语料库,Valence, Arousal结果变化不大,sigma=1时Valence结果总体上较好,而sigma=1.5時Arousal結果總體上稍好。
最后,使用扩展词典情况下,对于非回归方法的MAE、MSE提升较多,而对于回归类方法相關係數性能略下降。
Dims | sigma | Mean Method | Regression | MSE | MAE | r |
---|---|---|---|---|---|---|
Valence | 1.0 | Geometric | True | 1.466 | 0.987 | 0.486 |
Valence | 1.5 | Geometric | True | 1.481 | 0.991 | 0.488 |
Valence | 2.0 | Geometric | True | 1.483 | 0.991 | 0.487 |
Note: The segmentation tool used above is Jieba, use jieba for valence and traditional version jieba for arousal.
In this experiment, we use CKIP as the segmentation method instead of Jieba for valence and arousal prediction.
Valence
Dims | sigma | Mean Method | Regression | MSE | MAE | r |
---|---|---|---|---|---|---|
Valence | 1.0 | Geometric | False | 2.213 | 1.161 | 0.545 |
Valence | 1.0 | Geometric | True | 1.361 | 0.961 | 0.540 |
Valence | 1.5 | Geometric | False | 2.216 | 1.164 | 0.546 |
Valence | 1.5 | Geometric | True | 1.378 | 0.960 | 0.541 |
Valence | 2.0 | Geometric | False | 2.205 | 1.162 | 0.546 |
Valence | 2.0 | Geometric | True | 1.387 | 0.952 | 0.536 |
Arousal
Dims | sigma | Mean Method | Regression | MSE | MAE | r |
---|---|---|---|---|---|---|
Arousal | 1.0 | Arithmetic | True | 0.837 | 0.749 | 0.222 |
Arousal | 1.5 | Arithmetic | True | 0.791 | 0.723 | 0.220 |
Arousal | 2.0 | Arithmetic | True | 0.792 | 0.725 | 0.218 |
**Use extended lexicon**
|Dims|sigma|Mean Method|Regression|MSE|MAE|r|
|-------|-------|-------|-------|-------|-------|-------|
|Valence|1.0|Geometric|True|1.434|0.988|0.506|
|Valence|1.5|Geometric|True|1.460|0.984|0.503|
|Valence|2.0|Geometric|True|1.456|0.979|0.505|
|Arousal|1.0|Arithmetic|True|0.839|0.751|0.216|
|Arousal|1.5|Arithmetic|True|0.793|0.724|0.217|
|Arousal|2.0|Arithmetic|True|0.793|0.725|0.216|
Results of using the CVAW lexicon to predict the VA ratings of the CVAT corpus.
#texts | #tokens | Avg. tokens | MAE | RMSE | r | MAE | RMSE | r | |
---|---|---|---|---|---|---|---|---|---|
ANEW vs Forum | 20 | 15,035 | 751.75 | 1.20 | 1.55 | 0.77 | 0.72 | 0.85 | 0.27 |
CVAW vs CVAT | 2,009 | 111,559 | 55.53 | 0.960 | 1.173 | 0.541 | 0.723 | 0.889 | 0.220 |
Book Review | 266 | 11,330 | 42.59 | 0.874 | 1.077 | 0.272 | 0.592 | 0.723 | 0.155 |
Car Forum | 276 | 19,485 | 70.59 | 0.876 | 1.067 | 0.299 | 0.814 | 0.970 | 0.183 |
Computer Review | 191 | 8,653 | 45.30 | 0.860 | 0.082 | 0.362 | 0.803 | 1.134 | 0.210 |
Hotel Review | 305 | 12,120 | 39.74 | 1.082 | 1.314 | 0.588 | 0.753 | 0.895 | 0.445 |
News Article | 567 | 37,185 | 65.58 | 0.854 | 1.068 | 0.582 | 0.744 | 0.908 | 0.176 |
Politics Forum | 403 | 22,761 | 56.48 | 0.770 | 0.946 | 0.581 | 0.733 | 0.941 | 0.091 |
Note: In this table, tokens mean different characters. and Geometric Mean is used.
In the following table, we count the word information after CKIP segmentation. And we use Arithmetic Mean method.
#texts | #tokens | Avg. tokens | MAE | RMSE | r | MAE | RMSE | r | |
---|---|---|---|---|---|---|---|---|---|
ANEW vs Forum | 20 | 15,035 | 751.75 | 1.20 | 1.55 | 0.77 | 0.72 | 0.85 | 0.27 |
CVAW vs CVAT | 2,009 | 65,559 | 32.63 | 0.974 | 1.187 | 0.525 | 0.724 | 0.891 | 0.212 |
Book Review | 266 | 7,112 | 26.74 | 0.872 | 1.081 | 0.251 | 0.595 | 0.727 | 0.139 |
Car Forum | 276 | 12,120 | 43.91 | 0.892 | 1.083 | 0.254 | 0.812 | 0.967 | 0.200 |
Computer Review | 191 | 5,054 | 26.46 | 0.853 | 1.067 | 0.394 | 0.805 | 1.136 | 0.208 |
Hotel Review | 305 | 7,261 | 23.81 | 1.098 | 1.328 | 0.577 | 0.750 | 0.888 | 0.459 |
News Article | 567 | 20,881 | 36.83 | 0.866 | 1.070 | 0.579 | 0.744 | 0.907 | 0.183 |
Politics Forum | 403 | 13,115 | 32.54 | 0.785 | 0.954 | 0.574 | 0.732 | 0.940 | 0.105 |
- [Yunchao He] (https://plus.google.com/+YunchaoHe)
- yunchaohe@gmail.com
- YZU at Taiwan
- @元智大学资讯工程学系1608B 民105年1月