![](data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAxoAAAN0CAIAAAC83MkvAAAACXBIWXMAABYlAAAWJQFJUiTwAAAgAElEQVR42u3dfXDV1YHw8YMNM4vBWZAJ3EdhQM26FcmVZ9xiS8Kok4ALWAuDAyS0zmqXrViC0tRpRZc0jKXT0aBNUOrSqTPs5gULog4RhGSAlaSF1i7ei7XTphoX5Am562JrLulubuPzx93Nur4ggbzz+fwVQnJ/13Nj5ss5555fCAAAnIcR77//vlEAADhnFxkCAAA5BQAgpwAA5BQAgJwCAEBOAQDIKQAAOQUAIKcAAJBTAAByCgBATgEAyCkAAOQUAICcAgCQUwAAcgoAADkFACCnAADkFACAnAIAkFMAAMgpAAA5BQAgpwAA5BQAAHIKAEBOAQDIKQAAOQUAgJwCAJBTAAByCgBATgEAIKcAAOQUAICcAgCQUwAAcgoAADkFACCnAADkFACAnAIAQE4BAMgpAAA5BQAgpwAAkFMAAHIKAEBOAQDIKQAA5BQAgJwCAJBTAAByCgAAOQUAIKcAAOQUAICcAgCQUwAAyCkAADkFACCnAADkFAAAcgoAQE4BAMgpAAA5BQCAnAIAkFMAAHIKAEBOAQAgpwAA5BQAgJwCAJBTAAByCgAAOQUAIKcAAOQUAICcAgBATgEAyCkAADkFACCnAACQUwAAcgoAQE4BAMgpAADkFACAnAIAkFMAAHIKAEBOAQAgpwAA5BQAgJwCAJBTAADIKQAAOQUAIKcAAOQUAAByCgBATgEAyCkAADkFAICcAgCQUwAAcgoAQE4BACCnAADkFACAnAIAkFMAAHIKAAA5BQAgpwAA5BQAgJwCAEBOAQDIKQAAOQUAIKcAAJBTAAByCgBATgEAyCkAAOQUAICcAgCQUwAAcgoAQE4BACCnAADkFACAnAIAkFMAAMgpAAA5BQAgpwAA5BQAAHIKAEBOAQDIKQAAOQUAgJwCAJBTAAByCgBATgEAyCkAAOQUAICcAgCQUwAAcgoAADkFACCnAADkFACAnAIAQE4BAMgpAAA5BQAgpwAAkFMAAHIKAEBOAQDIKQAA5BQAgJwCAJBTAAByCgBATgEAIKcAAOQUAICcAgCQUwAAyCkAADkFACCnAADkFAAAcgoAQE4BAMgpAAA5BQCAnAIAkFMAAHIKAEBOAQDIKQAA5BQAgJwCAJBTAAByCgAAOQUAIKcAAOQUAICcAgBATgEAyCkAADkFACCnAACQUwAAcgoAQE4BAMgpAAA5BQCAnAIAkFMAAHIKAEBOAQAgpwAA5BQAgJwCAJBTAADIKQAAOQUAIKcAAOQUAAByCgBATgEAyCkAADkFAICcAgCQUwAAcgoAQE4BAMgpAADkFACAnAIAkFMAAHIKAAA5BQAgpwAA5BQAgJwCAEBOAQDIKQAAOQUAIKcAAJBTAAByCgBATgEAyCkAADkFAICcAgCQUwAAcgoAQE4BACCnAADkFACAnAIAkFMAAMgpAAA5BQAgpwAA5BQAAHIKAEBOAQDIKQAAOQUAIKcMAQCAnAIAkFMAAHIKAEBOAQAgpwAA5BQAgJwCAJBTAADIKQAAOQUAIKcAAOQUAAByCgBATgEAyCkAADkFAICcAgCQUwAAcgoAQE4BAMgpAADkFACAnAIAkFMAAHIKAAA5BQAgpwAA5BQAgJwCAEBOAQDIKQAAOQUAIKcAAJBTAAByCgBATgEAyCkAADkFAICcAgCQUwAAcgoAQE4BACCnAADkFACAnAIAkFMAAMgpAAA5BQAgpwAA5BQAAHIKAEBOAQDIKQAAOQUAgJwCAJBTAAByCgBATgEAyCkAAOQUAICcAgCQUwAAcgoAADkFACCnAADkFACAnAIAQE4BAMgpAAA5BQAgpwAAkFMAAHIKAEBOAQDIKQAAOQUAgJwCAJBTAAByCgBATgEAIKcAAOQUAICcAgCQUwAAyCkAADkFACCnAADkFAAAcgoAQE4BAMgpAAA5BQAgpwAAkFMAAHIKAEBOAQDIKQAA5BQAgJwCAJBTAAByCgAAOQUAIKcAAOQUAICcAgBATgEAyCkAADkFACCnAACQUwAAcgoAQE4BAMgpAAA5BQCAnAIAkFMAAHIKAEBOAQAgpwAA5BQAgJwCAJBTAADIKQAAOQUAIKcAAOQUAAByCgBATgEAyCkAADkFACCnAACQUwAAcgoAQE4BAMgpAADkFACAnAIAkFMAAHIKAAA5BQAgpwAA5BQAgJwCAEBOAQDIKQAAOQUAIKcAAOQUAAByCgBATgEAyCkAADkFAICcAgCQUwAAcgoAQE4BACCnAADkFACAnAIAkFMAAMgpAAA5BQAgpwAA5BQAAHIKAEBOAQDIKQAAOQUAIKcAAJBTAAByCgBATgEAyCkAAOQUAICcAgCQUwAAcgoAADkFACCnAADkFACAnAIAQE4BAMgpAAA5BQAgpwAA5BQAAHIKAEBOAQDIKQAAOQUAgJwCAJBTAAByCgBATgEAIKcAAOQUAICcAgCQUwAAyCkAADkFACCnAADkFACAnAIAQE4BAMgpAAA5BQAgpwAAkFMAAHIKAEBOAQDIKQAA5BQAgJwCAJBTAAByCgAAOQUAIKcAAOQUAICcAgBATgEAyCkAADkFACCnAADkFAAAcgoAQE4BAMgpAAA5BQCAnAIAkFMAAHIKAEBOAQAgpwAA5BQAgJwCAJBTAADIKQAAOQUAIKcAAOQUAICcAgBATgEAyCkAADkFACCnAACQUwAAcgoAQE4BAMgpAADkFACAnAIAkFMAAHIKAAA5BQAgpwAA5BQAgJwCAJBThgAAQE4BAMgpAAA5BQAgpwAAkFMAAHIKAEBOAQDIKQAA5BQAgJwCAJBTAAByCgAAOQUAIKcAAOQUAICcAgBATgEAyCkAADkFACCnAADkFAAAcgoAQE4BAMgpAAA5BQCAnAIAkFMAAHIKAEBOAQAgpwAA5BQAgJwCAJBTAADIKQAAOQUAIKcAAOQUAICcAgBATgEAyCkAADkFACCnAACQUwAAcgoAQE4BAMgpAADkFACAnAIAkFMAAHIKAAA5BQAgpwAA5BQAgJwCAEBOAQDIKQAAOQUAIKcAAOQUAAByCgBATgEAyCkAADkFAICcAgCQUwAAcgoAQE4BACCnAADkFACAnAIAkFMAAMgpAAA5BQAgpwAA5BQAgJwCAEBOAQDIKQAAOQUAIKcAAJBTAAByCgBATgEAyCkAAOQUAICcAgCQUwAAcgoAADkFACCnAADkFACAnAIAkFMAAMgpAAA5BQAgpwAA5BQAAHIKAEBOAQDIKQAAOQUAgJwCAJBTAAByCgBATgEAIKcAAOQUAICcAgCQUwAAyCkAADkFACCnAADkFACAnAIAQE4BAMgpAAA5BQAgpwAAkFMAAHIKAEBOAQDIKQAA5BQAgJwCAJBTAAByCgAAOQUAIKcAAOQUAICcAgCQUwAAyCkAADkFACCnAADkFAAAcgoAQE4BAMgpAAA5BQCAnAIAkFMAAHIKAEBOAQAgpwAA5BQAgJwCAJBTAAByCgAAOQUAIKcAAOQUAICcAgBATgEAyCkAADkFACCnAACQUwAAcgoAQE4BAMgpAADkFACAnAIAkFMAAHIKAAA5BQAgpwAA5BQAgJwCAJBTAADIKQAAOQUAIKcAAOQUAAByCgBATgEAyCkAADkFAICcAgCQUwAAcgoAQE4BACCnAADkFADAgMswBDCcJBKJZDLZ/cesrKzMzEzDAiCngE+USqUOHTq0deszlZUVn/Q1EyZE7rlnxcKFC3NycowYQK8b8f777xsFGIoaGxtLS7/T0FAfQsjOji5cuPzP//zSaHRmCGHs2Kw//vF0R0fy9On3mpvjx441b9u26Z13WidMiCxevPjv//6hrKwsAwggp+DC1dLS8sADa2pra7Kzo3/zN9+64YbZY8d+eh41N8f37duR7qqysrKSkhLrgAByCi44qVTqjjvuqKmpGTcu8v3vb5s+PfccHmH79h8+8khxCKGqqqqoqMioAsgpuFAkEon8/Px4PP7ww1UFBYszMs5972NHR/KJJ9bU1lYUFhZu2bLlfB4KADkFQ0M8Hp89e04qFWprY2eztHc2Dh588b775ufn59fU1NhNBXDOnDsFQ0BjY2M0Gp08edpzzzX3VkuFEPLy5r3wwptHjryWkxNNJBLGGeDcmJ2CwS6RSIwfP37GjIKKil19sSp36lRi6dLo9OnX7t6926ofgJyC4SaVSk2cOCmVCnV1x/qudU6caLnttisKCwurq6uNOUBPWeyDQe2v/3puer9Un84bXXbZlMcfr6upqdm4caMxB+gps1MweDU2Nubl5T3+eF1e3rx+uNyjj95bW1vR3t7uPCqAHjE7BYNUKpW6++57Zswo6J+WCiF8/evrQwjl5eUGH0BOwXDwzDPPHD0ae+ihzf12xVGjMh9+uKq0tLSlpcX4A8gpGPK+8Y2SW24pvOyyKf150YKCxePGRTZseMz4A8gpGNri8fjJk6133vlAP183IyPjrrserKysSKVSXgUAOQVD2I4dO8aNi2Rn5/T/pWfNujWE8Prrr3sVAOQUDGE/+cn22bMXD8ilL7tsyrhxkR07dngVAOQUDFWJROLo0VhBweKBegK3377iySc3eSEA5BQMVclkMoQwefLVA/UEJk3KPnmy1QsBIKdgqGpqagoh9OKtjnsqGp0ZQnBTZAA5BZyX9CQZAJ/K3eNh0Glubp4xo2AAn0B6Yuwf/uEfPnSe52c/+9ns7OzuP+bk5FxyySUhhIkTJ/bpLQUB5BTQM++8884ArvSFEEaNygwhNDU1HThw4Oy/KxKJ3HzzzSGEW2+9dcyYMVOnTs3KynL7P+BC4BbIMOhUV1cvW7bsF78YsP83T5xoue22K958880pU6akUqnjx49/6Avefvvtt956K/3xzp07QwiJRKK+vv5jH62wsDA9rTVz5swpU6Z4fYHhx+wUcMbfERkZH22gKVOm5Obmpj8uKirq/ny6vd577714PN7c3PzrX/963759NTU1H/zegoKCWbNm5efnT58+3dwVMDyYnYJBJx6PR6PRl19uTy+69b+DB1+87775nZ2dvbIjKplMJhKJpqamQ4cO7d+/PxaLdf9VNBpdtGhRfn7+DTfcYPcVIKeAXtPS0nLFFVfU1sYG5CYzIYStWzc+8khxH/1ySKVSr7/++oEDB55//vkPrg8WFBSsXr36c5/7XFZWlp8BQE4B5xscI0eOvP/+yiVLVg7IE1i69Lq/+qtra2qq++G/9PXXX9+xY8f27du7Z62i0ejy5cuXLFmiqwA5BZy7Vavu3bVrf23tq/1/6Y6O5KxZo+vq6ubNm9ef100kElu3bv3glFV6vmrOnDnWAQE5BfRYY2NjXl7e3r1t/X9iQnrjVHt7+0DtE08mk88///z3v//97vmqVatWrV692rsCATkF9MAArvctXXrd5MlZn3TqQX9KJBKbNm3atGlTa2trCCEajT755JM2rQODkJvMwGCUkZFRVlb2yCPFHR39eqeXI0cam5tjZWVlg2EQsrKy1q5de+zYsYMHDxYUFMRisby8vEmTJlVXV6dSKT8kwOBhdgoGqWQyOXr06K99rWz58rX9dtGlS68bNSrEYq8OwgFpaWnZsmVLaWlpCCESiaxYsaKkpMTJVYCcAs5k48aNxcXF/baDKr1rKn0Y+mCuzPLy8u6oKi8vX7x4seU/QE4BHy+VSk2cOGny5GkVFbv6uhhOnUosXRqdPfvmfjgfQVQBcgroP+kT0pcuXfXNb/6gT7vty1++/r332n73u+YhtHz2wahKb1TvvvUNQH+yFR0GtZycnLq6utraiq1bN/bdVVatmtvcHIvHY0NrK1JmZubatWvb2toKCwvTG9WLiooSiYQfG0BOAf/LvHnz0u/y64uiSqVSDz5YdPhw/cGDB4foKeRZWVnV1dVvvvlmNBqtqakZP378xo0bvfUP6E8W+2AISKVS69evLy0tXbp01X33lffWJqFTpxLdLTU8lsmqq6tLSkpaW1uj0ejzzz/v5E9ATgEfboVly5bNmFFQXv7cqFHnuyrX3Bz/+tfnZGSEn/3sp8MpO5LJ5PLly2tqakIIlZWVd999ty3qQF+z2AdDRlFRUSwWe+uto7Nmjd68ed05n/B54kTLgw8WLV0anT792ng8NsymcDIzM6urq+vq6iKRSHFx8fXXX283FSCngP+Rk5Nz/PixsrKyp54qXbAge/fu6lOnetAKJ060PProvbfddsWRI/uqqqrq6+uH6H6pTzVv3rzm5ub0FvXx48dXV1f74QH6jsU+GJKSyeQDD6yprKwIIWRnRxcuXD5r1q3jx0/86MJWR0fy7bff2Ldvx7Ztm955p3XChMiGDRfQKU3pFdIQQmFh4ZYtWyz8AXIK+HBUHThwYPful9JdlZadHb3qqmtDCL/4xb533mlNf3LChMg996xYuHDhNddcc6ElRSKRSN/yLxKJxGKx4TohB8gp4LykUqnjx4//6le/evfdd//93/+9qalpxIgR8+fPDyFMnjz58ssvv8Df45ZKpUpKSioqKkIIw+ZtjICcAvpQWVlZCCF9XDjduhf+ysrK1q5da0CA3mIrOnChSL81MhKJlJaWFhUVOeoTkFMAPZaTkxOLxdLnp8+dO1dRAXIKoMeysrJeeeWVgoKC+vr6SZMmOZUKkFMAPZaRkbFr166ysrL07WgUFSCnAM6lqNauXauogF7hnX3ABW3dunXpt0A6QAGQUwDnqPsABUUFyCmAc9TY2JiXl6eogHNj7xRAyM3NPXjwYAghLy+vsbHRgAA9YnYK4L+k56jc2g/oKbNTMDx1dnYahJ7Kzc2trKz0Xj9ATgGhrKxs/fr1xuEcrFy50ukJgJwCOC/d51EVFBS4Cw0gpwDOsagKCwtjsZj7+gFyCuAcbdmyJX1fv5KSEqMByCmAHkvf1y8SiVRUVGzcuNGAAHIK4FyKKhaLRSKR4uJih1EBZ+DcKYAz6T4wva2tzWFUwMcyOwVwJrm5uVVVVSGEaDRqWzogpwDORVFR0apVq1pbW+fOnWs0ADkFcC7Ky8vTb/SzLR34KHunAM5KIpGIRqOtra2xWCwnJyf9yZaWlszMzE/dU5VKpY4fP979x4kTJ2ZkZBhSGDbMTsEwVFZWVlZWZhx6V1ZW1rZt20IIc+bMSSaT6Ui64oor9u7d+7H9FI/H161bF4n8nxEjRowcOfKKDxg5cuSIESNycq5bt25dPB43tjDU+ecRwNnKzc0tKysrLS1dsGDB3r179+zZE0LonqlKa2lp2bDhscrKihDCuHGR229fMWlSdiQyefz4y7u/pq3t7dbWt44ePVRRsam0tHTChMjixYv//u8f8s5BGKIs9sEwlJ6aKi0tNRS9LpVKzZ07t76+vqqq6umnn25ra3v11VfTf5VIJO69996amprs7OjChctnzbr1ssumfOoDNjfH9+3bsW3bpnfeaS0rKyspKcnMzDTOIKcAOTWcdW+iCiFUVlauXLkylUqtX7++tLR03LjI6tXlBQWLe7o1KpVKbd/+w0ceKe5+TOMMQ4jFPoCz0tjY+MQTT6Q/njFjxgsvvBBCuPTSS5PJ5Je+9KWGhob7769ctOjuc9tjnpGRsWTJyttuu/OJJ9YUFxc3NTVt2bLFdnWQUwDD1uHDh9Mf/OEPf7jqquxUKtTWxrKzc87zYUeNyvzmN39QULD4b/827+jRow0NDXZTwZBgsQ/gXMTj8RdffPHb3/52dnZ006b6sWN7s3tOnGi5884vZGSEeDymqEBOAQxPiUQiJyc6efK0iopdfbEqd+pUYunS6GWXZf3yl7+06geDnHOnAHoslUoVFhamUuG7363uo9YZOzartjYWj8fd2QbkFMAwdMcddzQ0NDz99E97d43vo0X1ox8drK+vr6x0ZxsY1Cz2AfRMS0vLFVdc8fjjdXl58/rhco8+em9tbUV7e7vzqEBOAQwT0eh1HR2htvbV/rlcKpWaP3/SnDk3V1dXG3wYnCz2wTDknn19p7GxMR6PffvbT/bbFTMyMlavLq+pqWlpaTH+IKcAhry7775nxoyC6dNz+/OiBQWLx42LbNjwmPEHOQUwtCUSiaNHY3/3d9/p5+tmZGTcddeDlZUVqVTKqwByCmAI27p1awhh2rQb+v/Sc+YsCSEcOnTIqwByCmAIe+655265pXBADtUcOzYrOzu6deszXgWQUwBDVTKZbGhomDv3ywP1BBYuXF5ZWeGFgEHIjQtgGCotLTUIvS6RSIQQrrxy6kA9gb/8y/+brjoHUMFgY3YK4Ky89957IYQ+PQb9zMaPv7y76gA5BTD0xOPxEMKoUQM2MzSAlwbkFMBwkJ4Ya2pqMhQgpwCGquzsqEEA5BTAuWtujg3g1Ts6kiGEnJwcLwTIKaDPuWdfXxgzZszAPoFTpxIhhEsuucRrAXIKYEiaOnVqCOHEiZaBfRpOSQA5BTC0nT793kBdOhZrCiFkZWV5FUBOAQxJU6ZMmTAhsm/fjoF6Ai+88HR+fr4XAuQUwBB2zz0rtm3bNCCXTqVShw/X33XXXV4FkFMAQ9jChQvfead1QLZPHT16KIQwe/ZsrwIMQiPef/99owBwNlKp1MiRI7/2tbLly9f286UffLDo7bdfi8df9SrAIGR2CuBsZWRklJWVPfVUafoIqH5z4kTLSy/VPPDAt7wEMDiZnQLogWQyedVV2TffvPib3/xBv130nntmnz7d9i//8kpGRoaXAAYhs1MAPZCZmblhQ3ltbUW/7aA6cqTx8OH6H/7wSS0Fg5bZKYCeSaVSEydOmjx5WkXFrr5OnI6O5IIF2ZddNj4Ws2sKBi+zUwA9k5GR8bOf/fTw4frS0jv6uttKShZkZISGhnrDDnIK6Ffu2dfXpkyZUldX99JLNZs3r+u7qzz+eMnhw/V79+5xEjrIKYBhaObMmWvWrNm8uWzr1o29/uCpVGrz5nW1tRV1dXU5OTlGGwY5GxsBzjZx9uzZM3Xq1K6urh/84AfV1dX/9m//dtdddz3ySPGxY7+9777y3tpHlUqlVq2ae/hwfVVV1bx584w8yCmAYeKZZ55ZtmxZ9x8/85nPhBBuvvnmRYsWzZ8//403fvXd71aPHXu+q3InTrTceecXMjLCwYMHc3NzDTsMCRb7AM5KUVHRm/8tNzf3T3/6UwjhoosumjdvXiwWe+uto7Nnj9+8ed05n/DZ0ZF89NF7b7vtioyMEI/HtBTIKYBhoqurq62tLf3xlClTpkyZsnPnzsbGxoKCgkgksnPnzhBCTk7O8ePH0gemL1iQvXt3dY+i6tSpxObN62bNGr1v3zNVVVXHjx+z9xyGFudOAXyiY8eO1dXVnTx58v7777/44otDCPF4PBqNRiKRWCy2adOm0tLS9vb2zMzM9Ncnk8kHHlhTWVkRQsjOji5cuHzOnCWftALY3Bx/5ZUDO3Zsbm6OhRDKysrWrFnjrE6QUwDDLaRCCNOmTZs7d+7FF1+cTCazs7NbW1vTG5sSicT48ePr6uo+tGE8mUweOHBg9+6X0l2VNmNGQbqrfve719L9lFZcvGrJksU33HCDkAI5BTB8HDhwYP/+/emQys/PHzNmTPrzs2fPrq+vLysrW7t2bfoz69atW7hw4SedZZBKpY4fP97U1BRCaG5ufv31X48YEWbOnHnppZeOGTNm6tSpU6ZMMdogpwCGofTU1Pz58ydNmtT9yXXr1pWWlhYUFOzatctMEiCnAHqmsbExLy8vvWXKPnHgQ7yzD7igdXZ2Hjhw4N133z3D1yQSidtvvz2EsGeP+70AH8N8NQxD6Rv2lZaWGooz6Orqeu211/bs2dPe3v6v//qvX/nKVz72y1KpVDQabW1traqqcr8XQE4BfDikQgg33XTTzJkzP6ml5s6d29raumrVqqKiIkMHyCmAEEL4wx/+8Oyzz3aH1MiRIz/pK9evX19fX19QUFBeXm7cADkF8F/GjBkzd+7cq6++uvsEhI+1cePG0tLSSCTy3HPPeSsfIKeAC1pXV9dFF/2vd97MmDHjzN/S2NhYXFycfitf96HnAB/LO/uA4ez06dPbt29/7LHHurq6zv670scihBC2bdvmrXzAp3LuFDA8dXZ2NjU1pQ83nzBhwle/+tUz7JH6oEQikX4rX/pOMkYS+FQW+4BhqPsuMaNHj54zZ8611177ocW+T22pyspKLQXIKeCC1tOQ+mBLlZWVrVy50hgCZ8liHzAMdXV1/elPfzrL1b2PtlT3HY4Bzoat6MCQd+zYsc7Ozv/1q+2ii7QU0G8s9gFDWFtb27PPPnvy5MmbbrrpxhtvPLcH0VKAnAI+7EK4Z9+7777b0NBw9OjREMK0adOuu+46LQXIKYCzdezYsR//+MchhAkTJsyfP3/SpElaCpBTAD1w+eWXX3nllTfddNM5h1QIobGx8fbbb9dSgJwCLkQXXXTRV77ylfN5hO5zz6uqqoqKigwpcF6/lAwBMJh1dXXF4/Hdu3f34mOuW7cu3VIHDx7UUsD5MzsFDF7xeHzPnj3t7e0hhPz8/B6dffCxUqlUSUlJRUVFJBLZs2dPTk6OQQbOn2M8gcHo2LFjdXV1J0+eDCHcdNNNM2fOPP+WSiQSRUVF9fX1kUgkFou5tzHQW8xOwXDT0tKS/iAzM3PoFkO6paZNm5afnz9mzJjzf8B4PD5nzpzW1taCgoJdu3ZlZPjtB/Qas1Mw5MXj8R07dmz/yU9iR49+7Bfk5+cvWLBgyZIlQ6iuTp8+/Z//+Z+9ElIhhI0bNxYXF4cQvIkPkFPA/2hpaXlsw4aKysoQQmTcuBW33549adKY0aOnXnll99c0xWIhhENHj1bU1oYQIhMmrLjnnpKSkszMzEH139LZ2Xn+a3kfK5lMLliwIL3At23bttzcXD85gJwCQiKRuPfee2tqaqLZ2csXLrzx+utzsrPP/C2pVOrQ0aMNP/956VNPhRAqKyvvvvvuwbDg1dnZ2dTUtHND5tIAAAisSURBVH///rvuuut8DpH6WN0nSxUUFFRXV9ssBcgpIKRSqfXr15eWlkbGjStfvXpxQUFPkyjZ0VH+T/9U+tRTkQkTyjdsGMBjArq6ul5++eX9+/eHEEaPHn3bbbf9xV/8RS8OVPodfIOqHQE5BQywRCJRWFjY0NBQef/9dy9adD59kOzoWP7wwzUvvVRWVrZmzZr+T422trZ//Md/bG9vHz169Jw5c6699tqLLuq1Y/C6d507DQHoH/65BkNDS0vLFz7/+ZBKxWprP3Vp71NljhpV/d3v3jpr1rKHHvrnf/7n559/vp93U40ePbq9vb23TkD4n0xMJtesWZOelFq1alV5eblJKaAfmJ2CISAej0ej0Wh2dv2mTVljx/biIzceOXL7t74VMjKaf/e7fi6qrq6uXpyRCiFUV1eXlJSYlALkFPBhiUQimpMzbfLkXRUVfTHXkjh1Krp06bXTp+/evbuP5nLa2tp+//vf9+LWqI/m5pe//OVYLBbslAIGgnv2waCWSqXy8/NDKvVcn61bZY0du+eJJxoaGkq+8Y1ef/B33313+/btmzZteuGFF7q6unr98VtaWoqKiqLRaCwWKywsbGtrW7lypZYC+plfOjCozZ07Nx6Pt+3dmzlqVN9dJSc7u+rhh5c99NANn/98b73X7/Tp07t27Tp69GgIYcKECfPnz+/dpb1kMrl8+fKampoQQjQaffLJJ50pBQwUi30weDU2Nubl5dU9/vi8vLx+uNy9jz5aUVvb3t7eK5uoysvL02/cW7x4ce8eKNXS0rJly5bS0tIQQiQSKS8vH8DjHgDkFAxq1+XkhD/+8dXa2v65XLKjY/SsWb11G5bf/va3f/zjH3v3BISWlpY1a9akZ6TSIbV48WJLe4CcAj7eiy++OH/+/F45FuHsbdy6tfiRR3prgqq3pFKpQ4cO3XPPPenN5tFo9Fvf+paQAgYPW9FhkPrqXXcV3nJLf7ZUCOHuRYsi48ateeCBHn1XPB4vLy/v7Ozs9eeTSCTWrVs3adKkvLy8WCwWjUYPHjz46quvFhUVaSlg8DA7BYNR+qCpfp6aSktPUJ3lb4Zjx47V1dWdPHkyhFBUVNRbRyEkk8kDBw489thj9fX16c+sWrVq9erVU6ZM8bMBDEL+eQeD0YEDB0II/d9SIYQbr78+3XNnPgazra3t2WefTYfUtGnT8vPzx4wZc/4VdeTIkWeeeSZ9rHkIIRqNfu9737vxxhsH1eIjgJyCIWDzU0+tWrp0QC6dk50dGTdux44dZ86p//iP/zh58uSVV175xS9+8TxDKpFI7N279+mnn+6ei4pEIitWrFixYkVWVpYfBmDws9gHg04ikRg/fvzBH/0od/r0AXkC6zZv3vTcc/+vtfXMX3bs2LFzPgEhkUj85je/aWho2LRpU+t/XygajS5atGjhwoXuDwMMLWanYNBJJpMhhKsnTx6oJ5A9aVLryZMf/ExnZ2dnZ+fFF1/8wU/2qKVSqdTx48ebmpoOHTr0zDPPtH6g1QoLC2+99dbZs2ebiwLkFNA7mpqaQgi9e6vjHpkZjYYQEolEVlZWV1fXyy+/vH///mnTpi1atOjsH6SlpeXtt99+6623du7c+dprr6XPOOhWUFDwpS996cYbb7zmmmu8Rw+QU8Dw1N7e3traumfPnvb29hDC1Vdf/bHNFEJ477334vF4CGHnzp0hhH379rV+ZKEwGo3edNNNt9xyy9SpU71BD5BTQN9qbm4umDFjAJ9A1tixI0eOvO+++954440Qwvjx47Oysr73ve+l/zZ9KPmZFRQUZGVl3XrrrZMnT7788sv1EyCngH7161//egBX+kIImaNGdXZ2Hj9+PH0D44+KRCI333xz+uOZM2deeuml6Q9CCBMnTrR4B1xovLMPBp3q6uply5a9/4tfDNQTaDlx4orbbnvllVfSnaSQAM7Mr0jg41166aUW6QDOhnv2waCTPnUp2dExUE/g7ba2EMLEiRO9FgByCoakSy65JISQOHVqoJ7AW62tIQQLfAByCoaq9LTQzpdfHqgn8PQLLxTk53shAOQUDFUZGRmrios379gxIFdPdnTUHz68+hvf8EIAyCkYwhYvWRJrbh6Q9b4Dr7wSQrjxxhu9CgByCoaw6dOnhxD2HjrU/5d+rLo6Pz8/MzPTqwAgp2AIy8zMXFVcXPLYY6lUqj+vG29urj98+BtW+gB6wjGeMEglk8nRo0eXfe1ra5cv77eLXrd0afizP3s1Hjf+AGfP7BQMUpmZmWVlZaVPPdVvB1A1HjkSa25+8oc/NPgAPWJ2CgavZDKZfdVV0yZP3lVR0denQCVOnYouXXrz7NnVZ3GHYwA+yOwUDF6ZmZnbtm+vP3x4/dNP9+mFUqlU0YMPhoyMH1RUGHYAOQXDSm5ubmVlZelTT1Xv3t13V7mjtLT+8OGf/uxnWVlZxhygp9xEAga7lStX/vY3v1n20ENjRo+el5fXuw+eSqXWP/10zUsvHTx40A2PAc7NZ77zne8YBRjkZs+e/ft337133bqLRoy48frre+thkx0dX1y9+scvvFBVVTV//nzjDHBubEWHIWPjxo3FxcUFM2Y8V16eOWrUeT5ay4kTX7jzzpCRsW379tzcXMMLIKfggtDY2JiXlxdCqHr44cUFBef2dr9kR8fyhx+ueemlnJychoYG+6UAzpOt6DCU5Obmtre3l5WVLXvooUnz51fv3t2jU6kSp06t27x59KxZ+44cqaqq+uUvf6mlAM6f2SkYklpaWtY88EBNbW0IoWDGjNVFRZ+79tqssWM//otPnNj58subd+yINTeHEMrKytasWdPXB1kByClgCEgkEj//+c83bNjQ0NDQ/cnCW275r789dar+8OHuz68qLl68ZMkNN9wgpADkFPBhyWTyjTfeiMfjIYRDhw4lTp4MI0Z89pprsrOzQwgzZ86cOHGiigLoI/8fmnR8Hb/7zwsAAAAASUVORK5CYII=)
562 Chapter 16. 低资源神经机器翻译 肖桐 朱靖波
要两次翻译,时间开销较大。而且在两次翻译中,翻译错误会进行累积从而产生错
误传播问题,导致模型翻译准确性降低。此外,基于枢轴语言的方法仍然假设源语
言和枢轴语言(或者目标语言和枢轴语言)之间存在一定规模的双语平行数据,但
是这个假设在很多情况下并不成立。比如,对于一些资源极度稀缺的语言,其到英
语或者汉语的双语数据仍然十分匮乏,这时使用基于枢轴语言的方法的效果往往也
并不理想。虽然存在以上问题,基于枢轴语言的方法仍然受到工业界的青睐,很多
在线翻译引擎也在大量使用这种方法进行多语言的翻译。
16.3.2 基于知识蒸馏的方法
为了缓解基于枢轴语言的方法中存在的错误传播等问题,可以采用基于知识蒸
馏的方法
[551, 953]
。知识蒸馏是一种常用的模型压缩方法
[549]
,基于教师-学生框架,在
第十三章已经进行了详细介绍。针对低资源翻译任务,基于教师-学生框架的方法基
本思想如图16.12所示。其中,虚线表示具有平行语料库的语言对,带有箭头的实线
表示翻译方向。这里,将枢轴语言(p)到目标语言(y)的翻译模型 P (y|p) 当作教
师模型,源语言(x)到目标语言(y)的翻译模型 P (y|x) 当作学生模型。然后,用
教师模型来指导学生模型的训练,这个过程中学习的目标就是让 P (y|x) 尽可能接近
P (y|p),这样学生模型就可以学习到源语言到目标语言的翻译知识。举个例子,假设
图16.12中 x 为源语言德语“hallo”,p 为中间语言英语“hello”,y 为目标语言法语
“bonjour”,则德语“hallo”翻译为法语“bonjour”的概率应该与英语“hello”翻译
为法语“bonjour”的概率相近。
x
p
y
P (y|x)
P (y|p)
图 16.12 基于教师-学生框架的翻译过程
需要注意的是,基于知识蒸馏的方法基于一个假设:如果源语言句子 x、枢轴语
言句子 p 和目标语言句子 y 这三者互译,则 P (y|x) 应接近 P (y|p) ,即:
P (y|x) ≈ P(y | p) (16.7)
和基于枢轴语言的方法相比,基于知识蒸馏的方法无需训练源语言到枢轴语言
的翻译模型,也就无需经历两次翻译过程。不过,基于知识蒸馏的方法仍然需要显性
地使用枢轴语言进行桥接,因此仍然面临着“源语言 → 枢轴语言 → 目标语言”转