-
摘要: 目前自然语言推理（Natural language inference，NLI）模型存在严重依赖词信息进行推理的现象.虽然词相关的判别信息在推理中占有重要的地位，但是推理模型更应该去关注连续文本的内在含义和语言的表达，通过整体把握句子含义进行推理，而不是仅仅根据个别词之间的对立或相似关系进行浅层推理.另外，传统有监督学习方法使得模型过分依赖于训练集的语言先验，而缺乏对语言逻辑的理解.为了显式地强调句子序列编码学习的重要性，并降低语言偏置的影响，本文提出一种基于对抗正则化的自然语言推理方法.该方法首先引入一个基于词编码的推理模型，该模型以标准推理模型中的词编码作为输入，并且只有利用语言偏置才能推理成功；再通过两个模型间的对抗训练，避免标准推理模型过多依赖语言偏置.在SNLI和Breaking-NLI两个公开的标准数据集上进行实验，该方法在SNLI数据集已有的基于句子嵌入的推理模型中达到最佳性能，在测试集上取得了87.60%的准确率；并且在Breaking-NLI数据集上也取得了目前公开的最佳结果.Abstract: At present, natural language inference (NLI) models rely heavily on word information. Although the discriminant information related to the words plays an important role in inference, the inference models should pay more attention to the internal meaning of continuous text and the expression of language, and carry out inference through an overall grasp of sentence meaning rather than make shallow inference based on the opposition or similarity between individual words. In addition, the traditional supervised learning method makes the model rely too much on the language priori of the training set, and lacks the understanding of the language logic. In order to explicitly emphasize the importance of the learning sequence encoding and reduce the impact of language bias, this paper proposes a natural language inference method based on adversarial regularization. This method firstly introduces an inference model based on word encoding, which takes the word encoding in the standard inference model as input, and it can infer successfully only by using language bias. Then, through the adversarial training between the two models, the standard inference model can avoid relying too much on language bias. Experiments were carried out on two open standard datasets, SNLI and Breaking-NLI. On the SNLI dataset, the method achieves the best performance in existing inference models based on sentence embedding, and achieves 87.60% accuracy in test set. And the inference model has achieved state-of-the-art result on the Breaking-NLI dataset.
-
状态估计在信号处理、计算机视觉、自动控制、目标跟踪、导航、金融、通信等领域[1-6]有着广泛应用.在高斯噪声环境下, 卡尔曼滤波(Kalman fllter, KF)[7]及其次优滤波算法可以很好解决该问题.在非高斯噪声环境下, KF算法及其次优滤波算法不再适用, 因此存在着粒子滤波(Particle filter, PF)[8]及其衍生滤波算法来解决状态估计问题.
基于无迹变换(Unscented transform, UT)的无迹卡尔曼滤波(Unscented Kalman fllter, UKF)[9-11]是一种计算非线性变换均值和协方差的次优卡尔曼滤波算法.相比于扩展卡尔曼滤波(Extended Kalman fllter, EKF), UKF不需要计算雅可比矩阵, 且其可以达到非线性函数二阶泰勒展开式的精度[9].因此其在导航制导、目标跟踪、信号处理和图像跟踪等方面有着很广泛应用.但UKF算法也存在着在某些情况下估计效果差等问题.
目前, 针对UKF算法估计值不准确的问题, 有众多改进方法.为了解决UKF在工程应用中因舍入误差导致数值不稳定的问题, 提出了求根UKF (Square-root unscented Kalman fllter, SRUKF)[12]算法.在加性噪声条件下, 为了降低UKF算法的计算复杂度, 提出了简化UKF (Simplified unscented Kalman fllter, SUKF)[13]算法.在先验信息不确定性大而量测精度高的情况下, 只用一次量测值的UKF算法的估计效果较差.因此, 提出了多次利用量测值的迭代UKF (Iterated unscented Kalman fllter, IUKF)[14], 递归更新滤波器(递归更新扩展卡尔曼滤波(Recursive update extended Kalman fllter, RUEKF)[15]、递归更新容积卡尔曼滤波(Recursive update cubature Kalman fllter, RUCKF)[16])等算法.基于二阶UT变换的UKF算法滤波估计精度只能达到二阶, 为了提高滤波精度, 提出了基于高阶UT变换和高阶容积变换(Cubature transform, CT)的高阶UKF[17-18]和高阶容积卡尔曼滤波(Cubature Kalman fllter, CKF)[19-21]等算法.
UKF及其改进算法虽然可以较好处理UKF算法的估计不准确的问题, 但其仍然存在在非线性程度高的环境下估计效果差等问题, 文献[22-23]中提出将UKF算法作为PF算法建议分布, 将UKF算法估计值作为重要性密度函数, 这就是无迹粒子滤波(Unscented particle fllter, UPF)[22-23]算法.从理论上讲, 随着随机采样粒子数量提高, UPF算法的精度可以逐渐提高.但UPF算法也存在一些问题, 如其运算时间很长, 时效性较差.且UPF算法效果不总是好于UKF算法, 在量测噪声较大时, UPF算法估计精度会不如UKF算法.
为了在低计算负载的情况下获得高的滤波估计精度, 本文提出了双层无迹卡尔曼滤波器(Double layer unscented Kalman filter, DLUKF)算法.其核心思想是用带有权值的采样点表示前一个时刻的后验密度函数; 而后用内层的UKF算法对每个带权值的采样点进行更新, 并用最新的量测值对采样点的权值进行更新; 然后将各个采样点进行加权融合, 得到了初始的估计值; 最后用外层UKF算法的更新机制对初始估计值进行更新得到最终的估计值.
1. 无迹卡尔曼滤波(UKF)算法
1.1 UT变换
假设非线性函数为$\mathit{\boldsymbol{y}} = \mathit{\boldsymbol{f}}(\mathit{\boldsymbol{x}}) $, UT变换是通过近似非线性函数的概率密度分布来近似非线性函数.其在得到先验均值$\bar {\mathit{\boldsymbol{x}}}$和协方差$ {\mathit{\boldsymbol{P}}_{xx}}$的基础上, 用采样策略选取一组确定性采样点集.而后得到这些采样点集经非线性变换后的采样点集, 进而求得经非线性变换后的均值$\bar {\mathit{\boldsymbol{y}}}$和协方差${\mathit{\boldsymbol{P}}_{yy}}$.
UT变换算法可以归纳为以下三步:
1) 根据先验均值$\bar {\mathit{\boldsymbol{x}}}$和协方差${\mathit{\boldsymbol{P}}_{xx}}$, 用采样策略得到$N$个确定性采样点$\{ {\mathit{\boldsymbol{x}}_i}\} _{i = 1}^N$.定义$w_i^m$为均值加权作用的权值, $w_i^c$为协方差加权所用的权值.
2) 将确定性采样点$\{ {\mathit{\boldsymbol{x}}_i}\} _{i = 1}^N$进行非线性$\mathit{\boldsymbol{f}}(\cdot) $变换, 得到$N$个经非线性变换后的采样点集$\{ {\mathit{\boldsymbol{y}}_i}\} _{i = 1}^N = \mathit{\boldsymbol{f}}(\{ {\mathit{\boldsymbol{x}}_i}\} _{i = 1}^N) $.
3) 通过对采样点集$\{ {\mathit{\boldsymbol{y}}_i}\} _{i = 1}^N$进行加权的形式得到经非线性变换后的均值和协方差为.
1.2 UKF算法
考虑典型的非线性系统, 其状态方程和量测方程分别为:
$ \begin{equation} {\mathit{\boldsymbol{x}}_{k + 1}} = \mathit{\boldsymbol{f}}({\mathit{\boldsymbol{x}}_k}) + {\mathit{\boldsymbol{w}}_k} \end{equation} $
(1) $ \begin{equation} {\mathit{\boldsymbol{z}}_{k + 1}} = \mathit{\boldsymbol{h}}({\mathit{\boldsymbol{x}}_{k + 1}}) + {\mathit{\boldsymbol{v}}_{k + 1}} \end{equation} $
(2) ${\mathit{\boldsymbol{x}}_k}$为$k$时刻$n$维的状态向量, ${\mathit{\boldsymbol{z}}_{k + 1}}$为$k + 1$时刻的量测向量. ${\mathit{\boldsymbol{w}}_k}$为$m$维的过程噪声, 其服从均值为0方差为$\mathit{\boldsymbol{Q}}$的高斯分布. ${\mathit{\boldsymbol{v}}_{k + 1}}$为$q$维的量测噪声, 其服从均值为0方差为$\mathit{\boldsymbol{R}}$的高斯分布.滤波算法的目的就是从带有噪声的量测值${\mathit{\boldsymbol{z}}_{k + 1}}$中估计出真实值${\mathit{\boldsymbol{x}}_{k + 1}}$.
UKF[9-10]算法是基于UT变换的一种滤波算法, 其思想是在一步预测的时候, 用UT变换来进行均值和协方差传递.在UKF算法中, 因为存在噪声项, 需要对状态进行扩维.因此状态向量可以表示为. UKF算法流程为:
1) 在$k$时刻由UT变换中的采样策略得到$N$个采样点集$\{ \mathit{\boldsymbol{x}}_k^i\} _{i = 1}^N$.
2) 采样点集$\{ \mathit{\boldsymbol{x}}_k^i\} _{i = 1}^N$经非线性变换$\mathit{\boldsymbol{f}}(\cdot) $后得到采样点集$\{ \mathit{\boldsymbol{x}}_{k + 1|k}^i\} _{i = 1}^N$.
3) 由采样点集$\{ \mathit{\boldsymbol{x}}_{k + 1|k}^i\} _{i = 1}^N$加权求得预测值${\hat {\mathit{\boldsymbol{x}}}_{k + 1|k}}$和预测协方差${\hat {\mathit{\boldsymbol{P}}}_{k + 1|k}}$.
4) 采样点集$\{ \mathit{\boldsymbol{x}}_{k + 1|k}^i\} _{i = 1}^N$经非线性变换$\mathit{\boldsymbol{h}}(\cdot) $后得到采样点集$\{ \mathit{\boldsymbol{z}}_{k + 1|k}^i\} _{i = 1}^N$.
5) 由采样点集$\{ \mathit{\boldsymbol{z}}_{k + 1|k}^i\} _{i = 1}^N$加权求得预测的量测值${\hat {\mathit{\boldsymbol{z}}}_{k + 1|k}}$及其协方差${\mathit{\boldsymbol{P}}_{zz}}$和互协方差${\mathit{\boldsymbol{P}}_{xz}}$.
6) 求得$k + 1$时刻的估计值${\hat {\mathit{\boldsymbol{x}}}_{k + 1}}$及和协方差${\hat {\mathit{\boldsymbol{P}}}_{k + 1}}$.
在实际应用中, 受初始误差的影响, UKF算法存在着收敛速度慢, 精度不高等问题.基于此, 文献[14]提出了IUKF算法, 文献[15]提出了RUEKF算法, 文献[16]出了RUCKF算法.这三种算法的核心思想都是多次利用量测值对估计值进行更形, 以获得更好的滤波估计效果.
2. 无迹粒子滤波(UPF)算法
UPF[22-23]算法是在PF算法的基础上, 用UKF算法的滤波估计值作为PF算法的建议密度函数.这虽然可以解决UKF算法不适用于非高斯环境等问题, 但其由于要选取大量的随机性采样点来逼近密度函数, 所以UPF算法会临着计算量大的问题. UPF算法具体步骤如下:
1) 由$p({\mathit{\boldsymbol{x}}_0}) $得到$N$个粒子点$\{ \mathit{\boldsymbol{x}}_0^{(i)}\} _{i = 1}^N$, 初始权值为$\mathit{\boldsymbol{w}}_0^{(i)} = 1/N$.
2) 用UKF算法对每一粒子进行状态更新.
3) 计算粒子点对应的权值$\mathit{\boldsymbol{w}}_k^{(i)} = \mathit{\boldsymbol{w}}_{k - 1}^{(i)}\frac{{p({\mathit{\boldsymbol{z}}_k}|\mathit{\boldsymbol{x}}_k^{(i)})p(\mathit{\boldsymbol{x}}_k^{(i)}|\mathit{\boldsymbol{x}}_{k - 1}^{(i)})}}{{q(\mathit{\boldsymbol{x}}_k^{(i)}|{\mathit{\boldsymbol{z}}_{1:k}})}}$并对其归一化.
4) 当粒子退化严重时, 对粒子进行重采样.
5) 计算每个粒子点${\mathit{\boldsymbol{x}}^{(i)}}$对应的协方差.
6) 重复步骤2)~5).
最后得到$k$时刻状态量的估计为${\hat {\mathit{\boldsymbol{x}}}_k} = \sum\nolimits_{i = 1}^N {\tilde {\mathit{\boldsymbol{w}}}_k^{(i)}\mathit{\boldsymbol{x}}_k^{(i)}} $.
3. 双层无迹卡尔曼滤波(DLUKF)算法
UPF需要用大量的粒子点去逼近状态的后验密度函数, 因此其有着运算量大的问题.本文所提的DLUKF算法用带权值的采样点去表征状态的后验密度函数, 其核心思想为用内层的UKF对每个带权值的采样点进行更新, 而后用最新的量测值对每个采样点的权值进行更新, 并对更新后的采样点进行加权求和得到下一时刻初始估计值, 然后将该初始估计值作为预测值运行外层UKF算法, 从而得到最终估计值.
3.1 DLUKF算法框架
DLUKF算法由外层UKF算法和内层UKF算法组成, 其算法流程如下:
状态初始条件为初始值${\hat {\mathit{\boldsymbol{x}}}_0} = {\rm E}({\mathit{\boldsymbol{x}}_0}) $, 初始协方差${\hat {\mathit{\boldsymbol{P}}}_0} = {\rm E}(({\mathit{\boldsymbol{x}}_0} - {\hat {\mathit{\boldsymbol{x}}}_0}){({\mathit{\boldsymbol{x}}_0} - {\hat {\mathit{\boldsymbol{x}}}_0})^{\rm T}}) $.因为存在噪声项, 需要对初始的状态进行扩维处理.其可以表示为
$ \begin{equation} \hat {\mathit{\boldsymbol{x}}}_0^a = {\left[ {\begin{array}{*{20}{c}} {{{\hat {\mathit{\boldsymbol{x}}}}_0}}&0&0 \end{array}} \right]^{\rm T}} \end{equation} $
(3) $ \begin{equation} \mathit{\boldsymbol{P}}_0^a = {\left[ {\begin{array}{*{20}{c}} {{\mathit{\boldsymbol{P}}_0}}&0&0\\ 0&\mathit{\boldsymbol{Q}}&0\\ 0&0&\mathit{\boldsymbol{R}} \end{array}} \right]^{\rm T}} \end{equation} $
(4) 内层UKF算法:
在$k$时刻, 用采样策略选取$N$个采样点$\{ {\hat {\mathit{\boldsymbol{x}}}_{i, k}}\} _{i = 1}^N$, 并求取其权值对应的一阶矩$w_{i, k}^m$和二阶矩$w_{i, k}^c$.而后用内层UKF算法对每个采样点进行更新.
对每个采样点, 用采样策略选取$M$个采样点$ \{ {\hat {\mathit{\boldsymbol{x}}}_{j, i, k}}\} _{j = 1}^M$, 并取其对应的一阶矩$w_{j, i, k}^m$和二阶矩$w_{j, i, k}^c$.
时间更新:
$ \begin{equation} \hat {\mathit{\boldsymbol{x}}}_{j, i, k + 1|k}^x = \mathit{\boldsymbol{f}}(\hat {\mathit{\boldsymbol{x}}}_{j, i, k}^x, \hat {\mathit{\boldsymbol{x}}}_{j, i, k}^w) \end{equation} $
(5) $ \begin{equation} {\hat {\mathit{\boldsymbol{x}}}_{i, k + 1|k}} = \sum\limits_{j = 1}^M { w_{j, i, k}^m\hat {\mathit{\boldsymbol{x}}}_{j, i, k + 1|k}^x} \end{equation} $
(6) $ \begin{align} {{\hat {\mathit{\boldsymbol{P}}}}_{i, k + 1|k}} = &\sum\limits_{j = 1}^M {w_{j, i}^c(\hat {\mathit{\boldsymbol{x}}}_{j, i, k + 1|k}^x - {{\hat {\mathit{\boldsymbol{x}}}}_{i, k + 1|k}})}\times\nonumber\\& {(\hat {\mathit{\boldsymbol{x}}}_{j, i, k + 1|k}^x - {{\hat {\mathit{\boldsymbol{x}}}}_{i, k + 1|k}})^{\rm T}} + \mathit{\boldsymbol{Q}} \end{align} $
(7) 量测更新:
基于预测值${\hat {\mathit{\boldsymbol{x}}}_{i, k + 1|k}}$和预测协方差${\hat {\mathit{\boldsymbol{P}}}_{i, k + 1|k}}$产生新的$M$个带权值的采样点.
$ \begin{equation} {\mathit{\boldsymbol{z}}_{j, i, k + 1|k}} = \mathit{\boldsymbol{h}}(\mathit{\boldsymbol{x}}_{j, i, k + 1|k}^x, \mathit{\boldsymbol{x}}_{j, i, k + 1|k}^v) \end{equation} $
(8) $ \begin{equation} {\hat {\mathit{\boldsymbol{z}}}_{i, k + 1|k}} = \sum\limits_{j = 1}^M {w_{j, i}^m{\mathit{\boldsymbol{z}}_{j, i, k + 1|k}}} \end{equation} $
(9) $ \begin{align} {\mathit{\boldsymbol{P}}_{i, zz}} = & \sum\limits_{j = 1}^M {w_{j, i}^c({\mathit{\boldsymbol{z}}_{j, i, k + 1|k}} - {{\hat {\mathit{\boldsymbol{z}}}}_{i, k + 1|k}})}\times \nonumber\\& {({\mathit{\boldsymbol{z}}_{j, i, k + 1|k}} - {{\hat {\mathit{\boldsymbol{z}}}}_{i, k + 1|k}})^{\rm T}} + \mathit{\boldsymbol{R}} \end{align} $
(10) $ \begin{align} {\mathit{\boldsymbol{P}}_{i, xz}} = &\sum\limits_{j = 1}^M w_{j, i}^c(\hat {\mathit{\boldsymbol{x}}}_{j, i, k + 1|k}^x - {{\hat {\mathit{\boldsymbol{x}}}}_{i, k + 1|k}})\times \nonumber\\&{{({\mathit{\boldsymbol{z}}_{j, i, k + 1|k}} - {{\hat {\mathit{\boldsymbol{z}}}}_{i, k + 1|k}})}^{\rm T}} \end{align} $
(11) $ \begin{equation} {\mathit{\boldsymbol{K}}_{i, k + 1}} = {\mathit{\boldsymbol{P}}_{i, xz}}\mathit{\boldsymbol{P}}_{i, zz}^{ - 1} \end{equation} $
(12) $ \begin{equation} {\hat {\mathit{\boldsymbol{x}}}_{i, k + 1}} = {\hat {\mathit{\boldsymbol{x}}}_{i, k + 1|k}} + {\mathit{\boldsymbol{K}}_{i, k + 1}}({\mathit{\boldsymbol{z}}_{k + 1}} - {\hat {\mathit{\boldsymbol{z}}}_{i, k + 1|k}}) \end{equation} $
(13) $ \begin{equation} {\hat {\mathit{\boldsymbol{P}}}_{i, k + 1}} = {\hat {\mathit{\boldsymbol{P}}}_{i, k + 1|k}} - {\mathit{\boldsymbol{K}}_{i, k + 1}}{\mathit{\boldsymbol{P}}_{i, zz}}\mathit{\boldsymbol{K}}_{i, k + 1}^{\rm T} \end{equation} $
(14) 在采样点用内层UKF算法更新后, 类似于UPF算法, 表示一阶矩的权值和表示二阶矩的权值的更新可以表示为:
$ \begin{equation} \left\{ {\begin{array}{*{20}{c}} {w_i^m = w_i^m\frac{{p({\mathit{\boldsymbol{z}}_{k + 1}}|{{\hat {\mathit{\boldsymbol{x}}}}_{i, k + 1}})p({{\hat {\mathit{\boldsymbol{x}}}}_{i, k + 1}}|{{\hat {\mathit{\boldsymbol{x}}}}_{i, k}})}}{{q({{\hat {\mathit{\boldsymbol{x}}}}_{i, k + 1}}|{\mathit{\boldsymbol{z}}_{1:k}})}}}\\ {w_i^c = w_i^c\frac{{p({\mathit{\boldsymbol{z}}_{k + 1}}|{{\hat {\mathit{\boldsymbol{x}}}}_{i, k + 1}})p({{\hat {\mathit{\boldsymbol{x}}}}_{i, k + 1}}|{{\hat {\mathit{\boldsymbol{x}}}}_{i, k}})}}{{q({{\hat {\mathit{\boldsymbol{x}}}}_{i, k + 1}}|{\mathit{\boldsymbol{z}}_{1:k}})}}} \end{array}} \right. \end{equation} $
(15) 在得到权值更新的基础上, 对权值进行归一化处理, 有
$ \begin{equation} \left\{ {\begin{array}{*{20}{c}} {w_i^m = \frac{{w_i^m}}{{\sum\limits_{i = 1}^N {w_i^m} }}}\\ {w_i^c = \frac{{w_i^c}}{{\sum\limits_{i = 1}^N {w_i^c} }}} \end{array}} \right. \end{equation} $
(16) $k + 1$时刻的初始估计值及其协方差可以表示为
$ \begin{equation} \hat {\mathit{\boldsymbol{x}}}_{k + 1}^I = \sum\limits_{i = 1}^N {w_i^m{{\hat {\mathit{\boldsymbol{x}}}}_{i, k + 1}}} \end{equation} $
(17) $ \begin{equation} \hat {\mathit{\boldsymbol{P}}}_{k + 1}^I = \sum\limits_{i = 1}^N {w_i^c({{\hat {\mathit{\boldsymbol{x}}}}_{i, k + 1}} - \hat {\mathit{\boldsymbol{x}}}_{k + 1}^I){{({{\hat {\mathit{\boldsymbol{x}}}}_{i, k + 1}} - \hat {\mathit{\boldsymbol{x}}}_{k + 1}^I)}^{\rm T}}} + \mathit{\boldsymbol{Q}} \end{equation} $
(18) 外层UKF算法:
基于$\hat {\mathit{\boldsymbol{x}}}_{k + 1}^I$和$\hat {\mathit{\boldsymbol{P}}}_{k + 1}^I$, 用采样策略选取$N$个带权值的采样点$ \{ \mathit{\boldsymbol{x}}_{i, k + 1}^I\} _{i = 1}^N$.而后再次对粒子点进行量测更新, 可以表示为:
$ \begin{equation} \mathit{\boldsymbol{z}}_{i, k + 1}^I = \mathit{\boldsymbol{h}}(\mathit{\boldsymbol{x}}_{i, k + 1}^{I, x}, \mathit{\boldsymbol{x}}_{i, k + 1}^{I, v}) \end{equation} $
(19) $ \begin{equation} \hat {\mathit{\boldsymbol{z}}}_{k + 1}^I = \sum\limits_{j = 1}^N {w_i^m\mathit{\boldsymbol{z}}_{i, k + 1}^I} \end{equation} $
(20) $ \begin{equation} \mathit{\boldsymbol{P}}_{zz}^I = \sum\limits_{i = 1}^N {w_i^c(\mathit{\boldsymbol{z}}_{i, k + 1}^I - \hat {\mathit{\boldsymbol{z}}}_{k + 1}^I){{(\mathit{\boldsymbol{z}}_{i, k + 1}^I - \hat {\mathit{\boldsymbol{z}}}_{k + 1}^I)}^{\rm T}}} + \mathit{\boldsymbol{R}} \end{equation} $
(21) $ \begin{equation} \mathit{\boldsymbol{P}}_{xz}^I = \sum\limits_{i = 1}^N {w_i^c(\mathit{\boldsymbol{x}}_{i, k + 1}^{I, x} - \hat {\mathit{\boldsymbol{x}}}_{k + 1}^I){{(\mathit{\boldsymbol{z}}_{i, k + 1}^I - \hat {\mathit{\boldsymbol{z}}}_{k + 1}^I)}^{\rm T}}} \end{equation} $
(22) $ \begin{equation} \mathit{\boldsymbol{K}}_{k + 1}^I = \frac{{\mathit{\boldsymbol{P}}_{xz}^I}}{{\mathit{\boldsymbol{P}}_{zz}^I}} \end{equation} $
(23) $ \begin{equation} {\hat {\mathit{\boldsymbol{x}}}_{k + 1}} = \hat {\mathit{\boldsymbol{x}}}_{k + 1}^I + \mathit{\boldsymbol{K}}_{k + 1}^I({\mathit{\boldsymbol{z}}_{k + 1}} - \hat {\mathit{\boldsymbol{z}}}_{k + 1}^I) \end{equation} $
(24) $ \begin{equation} {\hat {\mathit{\boldsymbol{P}}}_{k + 1}} = \hat {\mathit{\boldsymbol{P}}}_{k + 1}^I - \mathit{\boldsymbol{K}}_{k + 1}^I\mathit{\boldsymbol{P}}_{zz}^I{(\mathit{\boldsymbol{K}}_{k + 1}^I)^{\rm T}} \end{equation} $
(25) 不断重复方程(5)~(25), 即可求得DLUKF算法在每个时刻的估计值${\hat x_k}$.
DLUKF算法的流程图如图 1所示.
3.2 基于对称采样策略的DLUKF算法
根据选取粒子点的采样策略不同, 又可以得到多种DLUKF算法.在UT变换中, 目前的采样策略方法包括对称采样、单形采样、3阶矩偏度采样和高斯分布4阶矩对称采样[8]等.还有为了保证经过非线性变换后协方差${\mathit{\boldsymbol{P}}_{yy}}$的正定性而提出的对基本采样策略进行比例修正的算法框架.
下面主要详细介绍对称采样策略.
考虑均值$\bar {\mathit{\boldsymbol{x}}}$和协方差${\mathit{\boldsymbol{P}}_{xx}}$的情况下, 通过对称采样的策略选取$N = 2n + 1$个采样点.采样点及其权值可以表示为:
$ \begin{equation} \left\{ \begin{array}{l} {\mathit{\boldsymbol{x}}^{(1)}} = \bar {\mathit{\boldsymbol{x}}}\\ \{ {\mathit{\boldsymbol{x}}^{(i)}}\} _{i = 2}^{N - n} = \bar {\mathit{\boldsymbol{x}}} + \sqrt {(n + \kappa )} {(\sqrt {{\mathit{\boldsymbol{P}}_{xx}}} )_{i - 1}}\\ \{ {\mathit{\boldsymbol{x}}^{(i)}}\} _{i = N - n + 1}^N = \bar {\mathit{\boldsymbol{x}}} - \sqrt {(n + \kappa )} {(\sqrt {{\mathit{\boldsymbol{P}}_{xx}}} )_{i - n + 1}} \end{array} \right. \end{equation} $
(26) $ \begin{equation} \left\{ \begin{array}{l} w_1^m = w_1^c = \frac{{\kappa }}{{n + \kappa }}\\ \{ w_i^m\} _{i = 2}^N = \{ w_i^c\} _{i = 2}^N = \frac{{1}}{{2n + 2\kappa }} \end{array} \right. \end{equation} $
(27) 式(26)中的$n$表示均值$\bar {\mathit{\boldsymbol{x}}}$维数. $\kappa $为比例参数, 可调节采样点与均值$\bar x$间的距离, 仅影响二阶以后高阶矩带来的误差. $ {(\sqrt {{\mathit{\boldsymbol{P}}_{xx}}})_i}$表示平方根矩阵的第$i$列或行.
在对称采样策略中, 采样点除了中心点外, 其他的采样点的权值是相同的.这说明除中心点外, 其他采样点的重要性是相同的.从采样点的分布可以看出, 采样点是关于中心点呈中心对称的.
基于对称采样的DLUKF算法就是在产生粒子点时用对称采样策略产生粒子点, 其具体的算法流程如下:
1) $k$时刻的估计值为${\hat {\mathit{\boldsymbol{x}}}_k}$, 协方差为${\hat {\mathit{\boldsymbol{P}}}_k}$.
2) 基于${\hat {\mathit{\boldsymbol{x}}}_k}$和${\hat {\mathit{\boldsymbol{P}}}_k}$, 通过式(26)和式(27)求得外层UKF算法$N$个采样点$\{ {\hat {\mathit{\boldsymbol{x}}}_{i, k}}\} _{i = 1}^N$, 及其权值对应的一阶矩$w_{i, k}^m$和二阶矩$w_{i, k}^c$.
3) 通过方程(5) $ \sim $ (14)得到每个粒子经内层UKF更新后的粒子点$\{ {\hat {\mathit{\boldsymbol{x}}}_{i, k + 1}}\} _{i = 1}^N$及其协方差$\{ {\hat {\mathit{\boldsymbol{P}}}_{i, k + 1}}\} _{i = 1}^N$.
4) 通过方程(15) $ \sim $ (16)得到外层UKF更新后的权值$\{ w_i^m\} _{i = 1}^N$和$\{ w_i^c\} _{i = 1}^N$.
5) 通过方程(17) $ \sim $ (18), 得到$k + 1$时刻的初始估计值$\hat {\mathit{\boldsymbol{x}}}_{k + 1}^I$及其协方差$\hat {\mathit{\boldsymbol{P}}}_{k + 1}^I$.
6) 基于$\hat {\mathit{\boldsymbol{x}}}_{k + 1}^I$和$\hat {\mathit{\boldsymbol{P}}}_{k + 1}^I$, 通过式(26)和式(27)求得$N$个采样点$\{ \mathit{\boldsymbol{x}}_{i, k + 1}^{I}\} _{i = 1}^N$.
7) 通过方程(19) $ \sim $ (25), 得到$k + 1$时刻的估计值为${\hat {\mathit{\boldsymbol{x}}}_{k + 1}}$, 协方差为${\hat {\mathit{\boldsymbol{P}}}_{k + 1}}$.
4. 仿真分析
将本文所提的基于对称采样策略DLUKF算法与UKF算法、IUKF算法、RUEKF算法、RUCKF算法、高阶UKF算法、高阶UKF算法和UPF算法分别在一维和二维仿真场景下进行仿真对比分析, 用滤波算法估计值与真实值间的均方根误差(Root mean square error, RMSE)来表示滤波算法估计效果.
4.1 单维仿真
假设有下述状态空间模型, 其状态方程和量测方程分别可以表示为:
$ \begin{equation} {x_{k + 1}} = 0.5{x_k} + \sin (0.04\pi k) + 1 + {w_k} \end{equation} $
(28) $ \begin{equation} {z_{k + 1}} = 0.2x_{k + 1}^2 + {v_{k + 1}} \end{equation} $
(29) 式(28)中${w_k}$表示过程噪声, 其服从$Ga(3, 2) $的伽马分布.式(29)中的${v_{k + 1}}$表示量测噪声, 其服从均值为0, 方差为的高斯分布.初始位置为${x_0} = 3$, IUKF算法, RUEKF算法和RUCKF算法的迭代次数都为10次. UPF算法粒子数量为100个, DLUKF算法产生粒子的方法是对称采样策略.仿真时间为30 s, 蒙特卡洛仿真次数为100次.其仿真结果如图 2所示.
通过图 2可以看出, IUKF算法、RUEKF算法、RUCKF算法、高阶UKF算法、高阶UKF算法和UPF算法滤波估计效果都略好于UKF算法.这是因为IUKF算法、RUEKF算法、RUCKF算法、高阶UKF算法、高阶UKF算法和UPF算法都对UKF算法进行了改进, 所以其效果是好于UKF算法的.本文所提的DLUKF算法在每个时刻的估计效果都好于其他的滤波算法.这说明, DLUKF算法对于UKF算法的改进效果比其他经典算法更加显著.且因为DLUKF算法用两层UKF算法对状态进行估计, 所以可以有着很好的滤波估计效果.
将UPF算法的粒子数由100逐渐增加到500, 其与UKF算法、IUKF算法、RUEKF算法、RUCKF算法、高阶UKF算法、高阶UKF算法和DLUKF算法的单次运行时间以及平均RMSE如表 1所示.
表 1 各算法计算时间及RMSE对比分析表Table 1 The calculation time and RMSE of each algorithm算法 运行时间(s) 平均RMSE UKF 0.0002 0.1566 IUKF 0.0014 0.0881 RUEKF 0.0006 0.0378 RUCKF 0.0031 0.0337 高阶UKF 0.0006 0.1434 高阶CKF 0.0006 0.1437 UPF (100) 0.1032 0.1153 UPF (200) 0.2097 0.0714 UPF (300) 0.3200 0.0626 UPF (400) 0.4296 0.0564 UPF (500) 0.5416 0.0476 DLUKF 0.0016 0.0297 通过表 1可以看出, UKF算法、RUEKF算法、高阶UKF算法和高阶CKF算法的用时都很少. IUKF算法、RUCKF和DLUKF算法的用时略长, 这是由于这三种算法都进行了多次滤波计算. UPF算法用时最长.在UPF算法中, 随着粒子数目的增多, 用时也是逐渐增加.在RMSE方面, DLUKF算法比另外7种方法小很多.在UPF算法中, 随着粒子数增多, RMSE也是逐渐变小的.但当500个粒子点时, UPF算法RMSE依然是DLUKF算法的两倍.这说明了基于带权值的采样点表征后验分布的方法是优于随机点表征后验分布的.
4.2 多维仿真
考虑一个二维匀速直线运动的例子, 其状态方程和量测方程分别为:
$ \begin{equation} {\mathit{\boldsymbol{{\rm{X}}}}_{k + 1}} = \mathit{\boldsymbol{F}}{\mathit{\boldsymbol{X}}_k} + {\mathit{\boldsymbol{w}}_k} \end{equation} $
(30) $ \begin{equation} {\mathit{\boldsymbol{Z}}_{k + 1}} = \mathit{\boldsymbol{h}}({\mathit{\boldsymbol{X}}_{k + 1}}) + {\mathit{\boldsymbol{v}}_{k + 1}} \end{equation} $
(31) 式(30)中, ${{\mathit{\boldsymbol{X}}_k} = [{x_k}, {\dot x_k}, {y_k}, {\dot y_k}]^{\rm T}}$是状态变量, 分别表示$x$轴和$y$轴方向的位置和速度. ${\mathit{\boldsymbol{w}}_k}$为过程噪声, 其服从均值为0, 方差为$\mathit{\boldsymbol{Q}}$的高斯分布.其中, $\mathit{\boldsymbol{F}}$和$\mathit{\boldsymbol{Q}}$分别可以表示为
$ \begin{equation} \mathit{\boldsymbol{F}} = \left[ {\begin{array}{*{20}{c}} 1&T&0&0\\ 0&1&0&0\\ 0&0&1&T\\ 0&0&0&1 \end{array}} \right] \end{equation} $
(32) $ \begin{equation} \mathit{\boldsymbol{Q}} = {q^2}\left[ {\begin{array}{*{20}{c}} {\frac{T^3}{3}}&{\frac{T^2}{2}}&0&0\\ {\frac{T^2}{2}}&T&0&0\\ 0&0&{\frac{T^3}{3}}&{\frac{T^2}{2}}\\ 0&0&{\frac{T^2}{2}}&T \end{array}} \right] \end{equation} $
(33) 式(31)中, ${\mathit{\boldsymbol{Z}}_{k + 1}} = {[{r_{k + 1}}, {\theta _{k + 1}}]^{\rm T}}$为观测变量, 分别表示对目标的径向距和方位角. ${\mathit{\boldsymbol{v}}_{k + 1}}$为量测噪声, 其为闪烁噪声, 可以表示为:
$ \begin{align} p({\mathit{\boldsymbol{v}}_{k + 1}}) = &(1 - \varepsilon ){p_1}({\mathit{\boldsymbol{v}}_{k + 1}}) + \varepsilon {p_2}({\mathit{\boldsymbol{v}}_{k + 1}}) = \nonumber\\& (1 - \varepsilon )N({\mathit{\boldsymbol{v}}_{k + 1}};0, {\mathit{\boldsymbol{R}}_1}) + \varepsilon N({\mathit{\boldsymbol{v}}_{k + 1}};0, {\mathit{\boldsymbol{R}}_2}) \end{align} $
(34) 量测方程$\mathit{\boldsymbol{h}}(\cdot) $可以表示为:
$ \begin{equation} \mathit{\boldsymbol{h}}({\mathit{\boldsymbol{X}}_{k + 1}}) = {\left[ {\begin{array}{*{20}{c}} {\sqrt {x_{k + 1}^2 + y_{k + 1}^2} }&{\arctan (\frac{{{y_{k + 1}}}}{{{x_{k + 1}}}})} \end{array}} \right]^{\rm{T}}} \end{equation} $
(35) 式(34)中, ${\mathit{\boldsymbol{R}}_1}$和${\mathit{\boldsymbol{R}}_2}$分别可以表示为
$ \begin{equation} {\mathit{\boldsymbol{R}}_1} = \left[ {\begin{array}{*{20}{c}} {\sigma _{1r}^2}&0\\ 0&{\sigma _{1\varepsilon }^2} \end{array}} \right] \end{equation} $
(36) $ \begin{equation} {\mathit{\boldsymbol{R}}_2} = \left[ {\begin{array}{*{20}{c}} {\sigma _{2r}^2}&0\\ 0&{\sigma _{2\varepsilon }^2} \end{array}} \right] \end{equation} $
(37) 仿真中, 仿真时间为100 s, 蒙特卡洛仿真次数为300次.目标初始位置为(20 000 m, 40 000 m), 初始速度为(-160 m/s, -150 m/s). IUKF算法、RUEKF算法和RUCKF算法的迭代次数都为10次. UPF算法粒子数量为300个, DLUKF算法产生粒子的方法是对称采样策略.
其他的参数设置为:
表 2 仿真参数设置Table 2 The Simulation parameters参数 $T$ $q$ ${\sigma _{1r}}$ ${\sigma _{1\varepsilon }}$ ${\sigma _{2r}}$ ${\sigma _{2\varepsilon }}$ $\varepsilon $ 数值 1 1 20 m 0.2$^{o}$ 200 m 0.2$^{o}$ 0.1 位置的RMSE公式可以表示为$RMS{E_{}} = \sqrt {RMSE_{x}^2 + RMSE_{y}^2} $, 进行仿真分析, 其效果如图 3所示.
图 3是各个算法在位置方面的RMSE, 可以看出, RUEKF算法、RUCKF算法、高阶UKF算法、高阶CKF算法和UKF算法的估计效果基本相同, 而IUKF算法和UPF算法的估计效果优于UKF算法.而本文算法性能是最好的, 这是因为本文算法用带权值的采样点表征后验分布, 这比随机的粒子点表征后验分布更有优势, 故DLUKF算法的RMSE是好于其他滤波算法的.这也说明所提的DLIKF算法对匀速直线运动可以有着很好的滤波估计效果.
在匀速直线运动中, 将UPF算法的粒子数由300逐渐增加到1 000, 其与其他算法的单次运行时间以及位置和速度的平均RMSE如表 3所示.
表 3 各个算法的性能Table 3 The performance of each algorithm算法 运行时间(s) 平均RMSE UKF 0.0059 99.8709 IUKF 0.0424 85.0107 RUEKF 0.0150 100.2616 RUCKF 0.0397 99.8704 高阶UKF 0.0193 100.4763 高阶CKF 0.0191 99.7558 UPF (300) 3.5953 88.2638 UPF (400) 4.8406 86.5004 UPF (500) 6.0552 85.8206 UPF (600) 7.2596 85.1056 UPF (700) 8.4211 84.6700 UPF (800) 9.6178 83.2706 UPF (900) 10.8389 82.9057 UPF (1 000) 12.0105 82.4258 DLUKF 0.0757 78.5559 从表 3可以看出, 本文算法的运算时间虽然略长于UKF算法、IUKF算法、REUKF算法、RUCKF算法, 高阶UKF算法和高阶CKF算法, 却远远小于UPF算法.且由于DLUKF算法的外层UKF算法选取了9个确定性采样点, 所以其运算时间是大约是UKF算法的9倍.在UPF算法中, 随着粒子数目的增多, 运算时间也是逐渐增加的.在各个方面的RMSE、DLUKF算法都是最好的.在UPF算法中, 随着粒子数目的增多, RMSE也是逐渐减小的, 但比起DLUKF算法、UPF算法的RMSE依然是很大的.这说明基于双层采样的的DLUKF算法在多维目标跟踪中有着很好的滤波估计效果.
5. 结论
本文所提的DLUKF算法是在双层UKF算法的基础上, 用采样策略选取带权值的采样点, 而后用内层UKF算法对每个采样点进行更新, 同时用最新的量测对采样点的权值进行更新, 最后通过外层UKF算法的更新机制得到每个时刻的滤波估计值.仿真结果表明, 在一维和二维的仿真场景中, 相比于存在的经典算法, 本文所提的DLUKF算法可以在较短的时间内获得很好的滤波估计效果.
-
图 2 基于对抗正则化的自然语言推理模型结构框图
Fig. 2 The structure of natural language inference model based on adversarial regularization
表 1 SNLI数据集上的三个例子
Table 1 Three examples from the SNLI dataset
Premise (前提) Hypothesis (假设) Label (标签) A soccer game with multiple males playing. Some men are playing a sport. Entailment (译文) 一场有多名男子参加的足球比赛. 有些男人在做运动. 蕴涵 A person on a horse jumps over a broken down airplane. A person is training his horse for a competition. Neutral (译文) 一个人骑着马跳过了一架坏掉的飞机. 为了参加比赛, 一个人正在训练他的马. 中立 A black race car starts up in front of a crowd of people. A man is driving down a lonely road. Contradiction (译文) 一辆黑色赛车在一群人面前启动. 一个男人开着车行驶在荒凉的路上. 矛盾 
下载: 导出CSV
表 2 不同方法在SNLI上的实验结果(%)
Table 2 Experimental results for different methods on SNLI (%)
对比方法 模型 训练准确率 测试准确率 Mou等[13] (2015) 300D Tree-based CNN encoders 83.3 82.1 Liu等[12] (2016) 600D (300 + 300) BiLSTM encoders 86.4 83.3 Liu等[12] (2016) 600D BiLSTM encoders with intra-attention 84.5 84.2 Conneau等[34] (2017) 4096D BiLSTM with max-pooling 85.6 84.5 Shen等[6] (2017) Directional self-attention network encoders 91.1 85.6 Yi等[7] (2018) 300D CAFE (no cross-sentence attention) 87.3 85.9 Im等[16] (2017) Distance-based Self-Attention Network 89.6 86.3 Kim等[35] (2018) DRCN (-Attn, -Flag) 91.4 86.5 Talman等[36] (2018) 600D HBMP 89.9 86.6 Chen等[37] (2018) 600D BiLSTM with generalized pooling 94.9 86.6 Kiela等[38] (2018) 512D Dynamic Meta-Embeddings 91.6 86.7 Yoon等[17] (2018) 600D Dynamic Self-Attention Model 87.3 86.8 Yoon等[17] (2018) Multiple-Dynamic Self-Attention Model 89.0 87.4 本文方法 BiLSTM_MP 89.46 86.51 本文方法 EMRIM 92.71 87.36 本文方法 BiLSTM_MP + AR 89.02 86.73 本文方法 EMRIM + AR 93.26 $\textbf{87.60}$
下载: 导出CSV
表 3 不同方法在Breaking-NLI上的测试结果
Table 3 Experimental results for different methods on Breaking-NLI
下载: 导出CSV
表 4 权重$\lambda$对NLI准确率的影响
Table 4 Impact of weight $\lambda$ on NLI accuracy
权重值 测试准确率(%) 0.5 86.90 0.25 87.14 0.10 87.60 0.05 87.35 0.01 87.39
下载: 导出CSV
-
[1] 郭茂盛, 张宇, 刘挺.文本蕴含关系识别与知识获取研究进展及展望.计算机学报, 2017, 40(4):889-910 http://d.old.wanfangdata.com.cn/Periodical/jsjxb201704008Guo Mao-Sheng, Zhang Yu, Liu Ting. Research advances and prospect of recognizing textual entailment and knowledge acquisition. Acta Automatica Sinica, 2017, 40(4):889 -910 http://d.old.wanfangdata.com.cn/Periodical/jsjxb201704008 [2] 任函, 冯文贺, 刘茂福, 万菁.基于语言现象的文本蕴涵识别.中文信息学报, 2017, 31(1):184-191 http://d.old.wanfangdata.com.cn/Periodical/zwxxxb201701024Ren Han, Feng Wen-He, Liu Mao-Fu, Wan Jing. Recognizing textual entailment based on inference phenomena. Journal of Chinese Information Processing, 2017, 31(1):184-191 http://d.old.wanfangdata.com.cn/Periodical/zwxxxb201701024 [3] Bowman S R, Angeli G, Potts C, Manning C D. A large annotated corpus for learning natural language inference. In: Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing. Lisbon, Portugal: Association for Computational Linguistics, 2015.632-642 [4] Glockner M, Shwartz V, Goldberg Y. Breaking nli systems with sentences that require simple lexical inferences. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Short Papers). Melbourne, Australia: Association for Computational Linguistics, 2018.650-655 https://arxiv.org/abs/1805.02266 [5] Gong Y, Luo H, Zhang J. Natural language inference over interaction space. In: Proceedings of the 6th International Conference on Learning Representations. 2018. [6] Shen T, Zhou T, Long G, Jiang J, Pan S, Zhang C. DiSAN: directional self-attention network for RNN/CNN-free language understanding. In: Proceedings of the 32nd AAAI Conference on Articifical Intelligence. New Orleans, USA: AAAI, 2018.5446-5445 [7] Yi T, Luu A T, Siu C H. Compare, compress and propagate: enhancing neural architectures with alignment factorization for natural language inference. In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. Brussels, Belgium: Association for Computational Linguistics, 2018.1565-1575 https://arxiv.org/abs/1801.00102 [8] Chen Q, Zhu X, Ling Z H, Wei S, Jiang H, Inkpen D. Enhanced LSTM for natural language inference. In: Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics. Vancouver, Canada: Association for Computational Linguistics, 2017.1657-1668 [9] Zhang K, Lv G, Wu L, Chen E, Liu Q, Wu H, Wu F. Image-enhanced multi-level sentence representation net for natural language inference. In Proceedings of the 2018 IEEE International Conference on Data Mining (ICDM). Singapore, Singapore: IEEE, 2018.747-756 [10] Camburu O M, Rocktäschel T, Lukasiewicz T, Blunsom P. e-SNLI: natural language inference with natural language explanations. In: Proceedings of the 32nd Conference on Neural Information Processing Systems. Montréal, Canada: NeurIPS, 2018.9560-9572 [11] Poliak A, Naradowsky J, Haldar A, Rudinger R, Van Durme B. Hypothesis only baselines in natural language inference. In: Proceedings of the 7th Joint Conference on Lexical and Computational Semantics. Vancouver, Canada, 2018.180-191 https://arxiv.org/abs/1805.01042 [12] Liu Y, Sun C, Lin L, Wang X. Learning natural language inference using bidirectional LSTM model and inner-attention. arXiv preprint arXiv: 1605.09090, 2016. [13] Mou L, Men R, Li G, Xu Y, Zhang L, Yan R, et al. Natural language inference by tree-based convolution and heuristic matching. In: Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics. Berlin, Germany: Association for Computational Linguistics, 2016.130 https://arxiv.org/abs/1512.08422 [14] Munkhdalai T, Yu H. Neural semantic encoders. In: Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics. Valencia, Spain, 2017.397-407 [15] Shen T, Zhou T, Long G, Jiang J, Wang S, Zhang C. Reinforced self-attention network: a hybrid of hard and soft attention for sequence modeling. In: Proceedings of the 27th International Joint Conference on Artificial Intelligence. New Orleans, USA: AAAI, 2018.4345-4352 [16] Im J, Cho S. Distance-based self-attention network for natural language inference. arXiv preprint arXiv: 1712.02047, 2017. [17] Yoon D, Lee D, Lee S K. Dynamic self-attention: computing attention over words dynamically for sentence embedding. arXiv preprint arXiv: 1808.07383, 2018. [18] Goodfellow I, Pouget-Abadie J, Mirza M, Xu B, WardeFarley D, Ozair S, et al. Generative adversarial nets. In: Proceedings of the 2014 Conference on Advances in Neural Information Processing Systems. Montreal, Canada: Curran Associates, Inc., 2014.2672-2680 [19] Mao X, Li Q, Xie H, Lau R Y, Wang Z, Smolley S P. Least squares generative adversarial networks. In: Proceedings of the 2017 IEEE International Conference on Computer Vision (ICCV). Venice, Italy: IEEE, 2017.2813-2821 https://arxiv.org/abs/1611.04076 [20] Arjovsky M, Chintala S, Bottou L. Wasserstein GAN. In: Proceedings of the 34th International Conference on Machine Learning. Sydney, Australia: ICML, 2017.214-233 [21] Zhu J Y, Zhang R, Pathak D, Darrell T, Efros A A, Wang O, et al. Toward multimodal image-to-image translation. In: Proceedings of the 31st Conference on Neural Information Processing Systems. Long Beach, USA: NIPS, 2017.465-476 https://arxiv.org/abs/1711.11586 [22] Yi Z, Zhang H, Tan P, Gong M. Dualgan: unsupervised dual learning for image-to-image translation. In: Proceedings of the 2017 IEEE International Conference on Computer Vision. Venice, Italy: IEEE, 2017.2849-2857 [23] Yu L T, Zhang W N, Wang J, Yu Y. Seqgan: sequence generative adversarial nets with policy gradient. In: Proceedings of the 31st AAAI Conference on Artificial Intelligence. San Francisco, USA: AAAI 2017.2852-2858 [24] Lin K, Li D, He X, Zhang Z, Sun M T. Adversarial ranking for language generation. In: Proceedings of the 31st Conference on Neural Information Processing Systems. Long Beach, CA, USA: NIPS, 2017.3155-3165 https://arxiv.org/abs/1705.11001 [25] Tzeng E, Hoffman J, Saenko K, Darrell T. Adversarial discriminative domain adaptation. In: Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, Hawaii, USA: IEEE, 2017.7167-7176 [26] Lample G, Zeghidour N, Usunier N, Bordes A, Denoyer L. Fader networks: manipulating images by sliding attributes. In: Proceedings of the 31st Conference in Neural Information Processing Systems. Long Beach, CA, USA: NIPS, 2017.5967-5976 [27] Pennington J, Socher R, Manning C. Glove: global vectors for word representation. In: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). Doha, Qatar: Association for Computational Linguistics, 2014.1532-1543 [28] Yang Z, Dhingra B, Yuan Y, Hu J, Cohen W W, Salakhutdinov R. Words or characters? fine-grained gating for reading comprehension. In: Proceedings of the 5th International Conference on Learning Representations. 2017. [29] McCann B, Bradbury J, Xiong C, Socher R. Learned in translation: contextualized word vectors. In: Proceedings of the 31st Conference in Neural Information Processing Systems. Long Beach, CA, USA: NIPS, 2017.6294-6305 [30] Anastasopoulos A, Chiang D. Tied multitask learning for neural speech translation. In: Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. New Orleans, USA: Association for Computational Linguistics, 2018.82-91 https://arxiv.org/abs/1802.06655 [31] Hinton G E, Srivastava N, Krizhevsky A, Sutskever I, Salakhutdinov R R. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv: 1207.0580, 2012. [32] Ba J L, Kiros J R, Hinton G E. Layer normalization. arXiv preprint arXiv: 1607.06450, 2016. [33] Kingma D P, Ba J. Adam: a method for stochastic optimization. arXiv preprint arXiv: 1412.6980, 2014. [34] Conneau A, Kiela D, Schwenk H, Barrault L, Bordes A. Supervised learning of universal sentence representations from natural language inference data. In: Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing. Copenhagen, Denmark: Association for Computational Linguistics, 2017.670-680 [35] Kim S, Hong J H, Kang I, Kwak N. Semantic sentence matching with densely-connected recurrent and co-attentive information. arXiv preprint arXiv: 1805.11360, 2018. [36] Talman A, Yli-Jyrä A, Tiedemann J. Natural language inference with hierarchical bilstm max pooling architecture. arXiv preprint arXiv: 1808.08762, 2018. [37] Chen Q, Ling Z H, Zhu X. Enhancing sentence embedding with generalized pooling. In: Proceedings of the 27th International Conference on Computational Linguistics. Santa Fe, New Mexico, USA: The COLING 2018 Organizing Committee, 2018.1815-1826 https://arxiv.org/abs/1806.09828 [38] Kiela D, Wang C, Cho K. Dynamic meta-embeddings for improved sentence representations. In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. Brussels, Belgium: Association for Computational Linguistics, 2018.1466-1477 https://arxiv.org/abs/1804.07983 [39] Parikh A, Täckström O, Das D, Uszkoreit J. A decomposable attention model for natural language inference. In: Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing. Austin, Texas, USA: Association for Computational Linguistics, 2016.2249-2255 https://arxiv.org/abs/1606.01933 [40] Nie Y, Bansal M. Shortcut-stacked sentence encoders for multi-domain inference. In: Proceedings of the 2nd Workshop on Evaluating Vector Space Representations for NLP. Copenhagen, Denmark: Association for Computational Linguistics, 2017.41-45 https://arxiv.org/abs/1708.02312 [41] Chen Q, Zhu X, Ling Z H, Inkpen D, Wei S. Neural natural language inference models enhanced with external knowledge. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics. Melbourne, Australia: Association for Computational Linguistics, 2018.2406-2417 https://arxiv.org/abs/1711.04289 期刊类型引用(27)
1. 孙惠平，周进爵，刘澍轩，陈钟. BFT问责机制综述. 信息网络安全. 2024(01): 14-23 . 
百度学术2. 王旭东，符精晶，王赟. 基于PBFT算法的分片技术的研究. 计算机与数字工程. 2024(01): 213-218+246 . 百度学术3. 付晶，倪旭. 区块链赋能战略性关键矿产资源高效利用：应用、挑战与对策. 学术探索. 2023(01): 94-102 . 百度学术4. 牛淑芬，韩松，谢亚亚，王彩芬. 基于区块链的多关键字属性基可搜索加密方案. 信息安全学报. 2023(01): 131-143 . 百度学术5. 于谦，李志淮，田娜. 针对区块链状态分片合谋攻击的改进方案. 计算机应用与软件. 2023(09): 341-349 . 百度学术6. 白兵，李志淮，李敏. 降低跨分片交易回滚概率的多轮验证方案. 计算机工程与应用. 2022(02): 129-136 . 百度学术7. 倪旭，付晶，张海亮. 区块链赋能战略性矿产资源生态治理研究. 云南社会科学. 2022(01): 91-97 . 百度学术8. 乔蕊，刘敖迪，陈迪，王清贤. 复杂物联网联盟链系统通信机制研究. 自动化学报. 2022(07): 1847-1860 . 本站查看9. 牛淑芬，陈俐霞，李文婷，王彩芬，杜小妮. 基于区块链的电子病历数据共享方案. 自动化学报. 2022(08): 2028-2038 . 本站查看10. 袁东彤，张桥云，顾瑞涛，曾建光. 金融科技的知识图谱分析——基于大数据与文献计量的视角. 产业经济评论. 2022(06): 153-171 . 百度学术11. 刘海洋，曹永生，陈彦清，井福荣，方沩. 农作物种质资源登记区块链模型研究. 植物遗传资源学报. 2021(01): 28-37 . 百度学术12. 张长鲁，张健. 国内区块链研究主题挖掘、热点分析及趋势探究. 统计与信息论坛. 2021(02): 119-128 . 百度学术13. 高昊昱，李雷孝，林浩，李杰，邓丹，李少旭. 区块链在数据完整性保护领域的研究与应用进展. 计算机应用. 2021(03): 745-755 . 百度学术14. 袁勇，欧阳丽炜，王晓，王飞跃. 基于区块链的智能组件：一种分布式人工智能研究新范式. 数据与计算发展前沿. 2021(01): 1-14 . 百度学术15. 解岩凯，魏凌波，张驰，王庆涛，孙启彬. 面向区块链轻节点的支付通道瞭望塔技术研究. 密码学报. 2021(05): 778-794 . 百度学术16. 李政，肖冰冰，李笑若，祝丙南，金晨光. Roundabout：一种基于燃烧证明的比特币隐私保护方法. 计算机与数字工程. 2021(12): 2538-2543 . 百度学术17. 王旭，甘国华，吴凌云. 区块链性能的量化分析研究. 计算机工程与应用. 2020(03): 55-60 . 百度学术18. 傅易文晋，陈华辉，钱江波，董一鸿. 面向时空数据的区块链研究综述. 计算机工程. 2020(03): 1-10 . 百度学术19. 毛志来，刘亚楠，孙惠平，陈钟. 区块链性能扩展与安全研究. 信息网络安全. 2020(03): 56-64 . 百度学术20. 朱建明，张沁楠，高胜. 区块链关键技术及其应用研究进展. 太原理工大学学报. 2020(03): 321-330 . 百度学术21. 袁勇，王飞跃. 可编辑区块链:模型、技术与方法. 自动化学报. 2020(05): 831-846 . 本站查看22. 刘海房，吴雨芯. 比特币系统综述. 现代计算机. 2020(19): 45-51 . 百度学术23. 李洋，门进宝，余晗，王思宁，范金刚，郭艳来. 区块链扩容技术研究综述. 电力信息与通信技术. 2020(06): 1-9 . 百度学术24. 刘海洋，方沩，陈彦清，曹永生. 区块链在农作物种质资源数据管理中的应用初探. 农业大数据学报. 2019(02): 105-113 . 百度学术25. 孙国梓，王纪涛，谷宇. 区块链技术安全威胁分析. 南京邮电大学学报(自然科学版). 2019(05): 48-62 . 百度学术26. 张朝栋，王宝生，邓文平. 基于侧链技术的供应链溯源系统设计. 计算机工程. 2019(11): 1-8 . 百度学术27. 杨林瑶，陈思远，王晓，张俊，王成红. 数字孪生与平行系统:发展现状、对比及展望. 自动化学报. 2019(11): 2001-2031 . 本站查看其他类型引用(50)
-
下载:
计量
- 文章访问数: 2317
- HTML全文浏览量: 331
- PDF下载量: 143
- 被引次数: 77
