A: Hello
B: Hello there
A: So, are you interested in this great TV? Honestly, I barely used it and decided to sell it because I don’t really watch much TV these days. I’m selling it for \$275
B: I am definitely interested in the TV, but it being 10 years old has me a bit skeptical. How does the TV look running movies and games, if you don’t mind me asking.
A: It’s full HD at 1080p and it looks great. The TV works like it is brand new. I’ll throw in a DVD player that was hooked up to it for the same price of \$275
B: The DVD player sounds nice, but unfortunately I’m on somewhat of a budget. Would you be willing to drop the price a tad, maybe \$230?
A: 230 is kind of low. I’ll tell ya what, if you come pick it up where it is located I’ll sell it for \$260
B: Throw in a couple of movies with that DVD player,and you have yourself a deal.
A: Deal.
B: OFFER \$260.00
A: ACCEPT
We threshold the probability by 0.5 to obtain the k-hot representation.
we bold the best results for a metric in all tables (several results are in bold if they have statistically insignificant differences).
This representation is enriched with the encodings of explicit strategy and dialogue act sequences using the structure encoders which is then used to condition the Utterance decoder. Please
使用图神经网络GAT（Graph Attention Networks）来处理谈判数据集CraigslistBargain（商家与消费者商议商品价格的一个数据集）
复制到markdown编辑器以获取更好的阅读体验
信息
1 数据集
数据集中的主体是卖家（即模型要学习的部分）和消费者之间的讨价还价的过程
大概是这样的：
A: Hello B: Hello there A: So, are you interested in this great TV? Honestly, I barely used it and decided to sell it because I don’t really watch much TV these days. I’m selling it for \$275 B: I am definitely interested in the TV, but it being 10 years old has me a bit skeptical. How does the TV look running movies and games, if you don’t mind me asking. A: It’s full HD at 1080p and it looks great. The TV works like it is brand new. I’ll throw in a DVD player that was hooked up to it for the same price of \$275 B: The DVD player sounds nice, but unfortunately I’m on somewhat of a budget. Would you be willing to drop the price a tad, maybe \$230? A: 230 is kind of low. I’ll tell ya what, if you come pick it up where it is located I’ll sell it for \$260 B: Throw in a couple of movies with that DVD player,and you have yourself a deal. A: Deal. B: OFFER \$260.00 A: ACCEPT
数据集中还包含着消费者想购买的商品的信息，包括介绍，价格，图片等等
2 模型
数据集一共有三个部分：（1 机器与消费者之间的对话；（2 双方的谈判策略 ；（3 对话的意图（dialogue action），分别用三个（两种）encoder来embed
hierarchical dialogue encoder：用于embed（1 ，利用BERT来embed对话，再将每一轮对话embed后作为GRU每个时间步的输入并将最后一个时间步的隐向量作为输出。
structure encoder：用于分别embed （2 和（3，即策略和对话意图，以策略为例，先将策略构建成有向图，然后再利用GAT以及Adaptive Structure Aware Pooling (ASAP) 层，来获得各个节点的向量表示，通过拼接平均和相加节点的向量表示来获得最后的整个图的向量表示。
decoder：利用前述的三个encoder的输出拼接后作为decoder的输入，decoder为一个GRU，将GRU每个时间步的隐状态通过softmax函数来生成与消费者的对话，值得注意的是协商的价格是用place holder来表示的，如\<现在的价格打八五折>
如何构建图：图为有向图，每一步机器或消费者都有一个或数个策略，且每个策略都有所有前面所有轮的策略指向它的一条边
模型的训练：模型的loss有四个部分，并分别有超参权重来控制这四个部分的loss，分别是预测下一步策略的对数损失函数，下一步对话意图的交叉熵损失，还有生成的对话的交叉熵损失，最后为交易的成功度等级分为五个等级，然后对于将模型预测出来成功度的等级和真正的等级计算交叉熵损失
3 评价方法
用于多分类问题的macro 和 micro F1，可见http://sofasofa.io/forum_main_post.php?postid=1001112
以及ROC AUC score，BLUE score
4 其他
1、除了把最好的结果标出来，这篇论文还把与最好结果在显著性上没有差别的结果也标了出来，这一点可以掩饰了有些结果比不上原来的模型的结果，显得结果栏里面都是超过前面模型的。
2、还有很多细节论文里都没写出来，但是这些细节对于不做这个任务的人来说也不是很重要：）
5 好的句子
We threshold the probability by 0.5 to obtain the k-hot representation.
we bold the best results for a metric in all tables (several results are in bold if they have statistically insignificant differences).
This representation is enriched with the encodings of explicit strategy and dialogue act sequences using the structure encoders which is then used to condition the Utterance decoder. Please