Abstract

Imagine a beer advertisement next to an article about drunk driving, or a coupon for a free dinner embedded in an article about food poisoning. While humans are quite good at seeing the error in these examples, the machine-learning algorithms that place advertisements online continue to struggle with this type of contextual nuance. We argue that this shortcoming stems from the manner in which these machines are taught about semantic relatedness—the conceptual distance between words in the human mind. Specifically, we hypothesize that there is a difference in how humans view semantic relatedness when context is present versus when it is absent and that this difference is missing from the data used by machines to place advertisements online. To test this hypothesis, we adapt existing best practices to create a new, context-aware database and then compare it to the current state of the art. We find substantial differences in the distribution of semantic relatedness scores for context-aware versus context-free databases. We also find that the nature and scope of these differences are likely to lead to the types of mistakes observed in practice.

摘要

想象一下在有关酒后驾车的文章旁边的啤酒广告，或在有关食物中毒的文章中嵌入免费晚餐的优惠券。尽管人们很擅长于在这些示例中看到错误，但是将广告在线放置的机器学习算法仍在与这种上下文细微差别作斗争。我们认为，这种缺陷源于教授这些机器有关语义相关性的方式，即语义上单词之间的概念距离。具体来说，我们假设人们在存在上下文时和不存在上下文时如何看待语义相关性，并且机器用于在线投放广告的数据中缺少这种区别。为了验证这一假设，我们采用了现有的最佳做法来创建新的，上下文感知的数据库，然后将其与最新技术进行比较。我们发现上下文感知数据库和上下文无关数据库的语义相关性分数分布存在很大差异。我们还发现，这些差异的性质和范围很可能导致实践中观察到的错误类型。