AI Primer for Marketers Part 5: AI-Powered Content Creation

营销策划 2017-06-12 阅读原文

The most innovative marketers routinely pick up new things, try them out, and succeed or fail. Why are marketers struggling so much to adapt to artificial intelligence and machine learning? In this series, we’ll explore machine learning and artificial intelligence to build a foundation for understanding the field – and how it applies to marketing.

AI Applications in Marketing: Content Creation

In the last post, we reviewed the different types of artificial intelligence and machine learning technologies, and how they fit together in a landscape of development. However, everything we have discussed up to this point has been largely theoretical. Let’s change perspectives and look at practical applications of many of these technologies in use today.

While artificial intelligence technologies have not reached a point where they can create content autonomously, they are exceptionally useful in helping us determine what content to create. The first of these technologies is called topic modeling, which uses natural language processing and natural language understanding. Topic modeling helps us digest enormous amounts of content to understand what the key topics are within a body of text.

For example, suppose we are trying to understand how people talk about our company. We could skim news articles or social media updates about our company comma but once we reach a certain volume, there’s simply no way for us to keep up. No one human could read all of the articles or conversations about Ford Motor cCompany or IBM. We would struggle to even read a day’s worth of conversations. Yet, if we want to create content that is relevant to our audiences, we must have an understanding of what it is they want to hear about.

Topic Modeling

Enter topic modeling. We can feed large quantities of text to machine learning software to extract out the most relevant topics to the conversation. Topic modeling has its origins in simple mathematics; beginning programmers often make software which delivers a raw number of word frequencies, the number of times a word appears within a piece of text. While frequency is important, so is what is known as saliency, or the relevance of a particular word or phrase to the body of text as a whole. For example, I use coffee shops in many of my marketing examples. The word coffee may appear frequently within my text, but it is not necessarily salient to the topic of digital marketing.

Topic modeling allows us to determine not only which topics are most frequent, but also most salient, so that we have a better understanding of what contents are create that uses not only popular, high-frequency terms, but also those most relevant to our audiences.

Let’s look at a small-scale example of topic modeling in action. Suppose I want to understand the key topics that make up the topsearch results for a given search term. We know from previous revelations that Google uses topic modeling as part of understanding the intent of a searcher. What if we were to use topic modeling to effectively reverse-engineer Google’s own algorithm and better understand what contents create in order to rank well?

For this example, we’ll be very meta and look at the search term content marketing:

Using topic modeling, we’d extract the contents of every top search result on page 1 – news articles, OneBox summaries, and regular web pages – into a single, large corpus of text. We’d then create a lexicon of known specialty terms so that our machine learning software has a solid starting point to know what words and phrases are jargon. With that lexicon in place, we’d run the corpus of text through a popular topic modeling application such as NLTK for Python or MALLET for Java, then map it with a network mapping tool:

With this topic model, we now know what our content about content marketing should be about. From here, we’d build our web pages, news stories, and social media content using this map and its contents to stand a better chance of ranking well for the content marketing search term.

A second example of machine learning and AI in content production today is clustering and classification. We teach machine learning software to recognize, for example, top performing content such as social media posts, blog posts, images, etc. and build a corpus of knowledge about what things work best. Based on this training library, we then assess content we create to identify how similar our content is to “winning content” and make adjustments accordingly.

Next: Content Distribution

In the next post in this series, we’ll look at how artificial intelligence and machine learning help us to accelerate distribution of content. Stay tuned!

If you enjoyed this, please share it with your network!

Want to read more like this from Christopher Penn ? Get updates here:

Christopher S. Penn

责编内容by:Christopher S. Penn阅读原文】。感谢您的支持!


2 latent methods for dimension reduction and topic... Photo: Before the stat...
百度开源其NLP主题模型工具包,文本分类等场景可直接使用... 2017年7月4日,百度开源了一款主题模型项目,名曰:Familia。 InfoQ记者第一时间联系到百度Familia项目负责人姜迪并对他进行采访,在本文中,他将为我们解析Familia项目的技术细节。 什么是Familia F...
还敢说自己是TED粉吗?连哪个演讲最爆款都不知道!... 作者:Teresa Alvarez 编译:元元、sunflower 科学VS心理,哪类演讲播放量能相差20倍! 当我看到Kaggle上的TED数据集时,竟然发现观看次数极为分散:从五万到四千七百万多(平均数为100万左右...
用scikit-learn学习LDA主题模型 在LDA模型原理篇我们总结了LDA主题模型的原理,这里我们就从应用的角度来使用scikit-learn来学习LDA主题模型。除了scikit-learn, 还有spark MLlib和gensim库也有LDA主题模型的类库,使用的...
主题模型(概率潜语义分析PLSA、隐含狄利克雷分布LDA)... 一、pLSA模型 1、朴素贝叶斯的分析 (1)可以胜任许多文本分类问题。 (2)无法解决语料中一词多义和多词一义的问题——它更像是词法分析,而非语义分析。 (3)如果使用词向量作为文档的特征,一词多义和多词一义会造成计...