Imagine you are stock analyst who wants to analyze the news associated with a particular company to better understand important trends and future performance.  Where do you turn to? Well entrenched tools such as a Google or Bloomberg immediately come to mind and these do a pretty good job of delivering relevant results.  That said, there is still a lot of effort involved in going from that initial information set to a synthesized view of it that can enable us to more easily find meaningful patterns and surface interesting signals. 

We still, and appropriately so, marvel at the ability of Google to quickly and efficiently help us find items that we are searching for.  However, in the context of a knowledge worker such as that analyst who has to distill large amounts of data into actionable insights, this experience doesn’t nearly go far enough.

Enter deep learning, a branch of machine learning. Deep learning has been setting the pace for the last several years within machine learning - largely as a result of the exponential increase in the right kind of cheap computing power that arrived with GPU based systems. My colleagues and I are big fans of deep learning because of its ability to help us attack unstructured data analytic problems and advance the knowledge cause. There are two big reasons underlying this:

  • Deep learning allows us to teach a computer by example. This enables a knowledge worker to throw examples at the computer and use them to teach it how they want their data to be analyzed and understood. In the context of unstructured data where it is often very difficult to codify all the underlying rules that make something interesting, this goes a long way. This approach allows us to approximate human intuition and is a big part of the reason why deep learning techniques deliver the best accuracies today.
  • Second, is the emergence of transfer learning.  Traditionally, the Achilles heel of deep learning has been the scale of the training effort required to teach the computer to do something well.   Transfer learning, a particular branch of deep learning, leverages the fact that it is possible to drastically shrink the training effort to do something if you can leverage learnings from related tasks.  So, for instance, if you have access to pre-trained deep learning models that have learned to analyze text in generalized ways, its possible to extend these models with only a handful of examples and teach them to handle your specific text analytic tasks with great efficiency.  Transfer learning makes deep learning practical and allows for a rapid level of customization that the knowledge distillation task, given the breadth of its users and uses, demands.

So how does this all translate into real world impact?  We wanted to run a quick experiment to showcase the possibilities. Imagine if you will, that you are that analyst that wants to hone in on what’s happening to a stock by looking at its associated news and commentary. Instead of starting from a long list of articles that just mention said company, what if you could quickly build a news article tagging service that classifies and organizes the information in these articles in a manner that you specify and makes it easier for you to quickly understand the underlying content? Let’s see what this would take.
First, we defined a classification system for articles to our liking.  For the purposes of this exercise we kept things simple and ended up with the following taxonomy:

  • Corporate News: Articles describing events that happen to a company (product launch, event announcement, etc.)
  • Stock News & Analysis: Articles focusing on analysis of stock price movement and future predictions
  • Fundamental Analysis: Articles looking at the underlying business issues affecting a company
  • Sector/Market News: News or analysis that focus on broad trends affecting the company
  • Other: Articles that don’t fall into any of other categories.

To build a model to organize information this way, we needed training data.  Next we gathered a dataset of about a thousand finance-specific news articles from various public websites and loaded these into our data labeling tool.  Long story short, it took five of us a couple of hours to attach labels to some of these articles and create a few hundred high quality examples that could serve as the training data for our model. Our reason for using multiple data labelers in this task was twofold. One, it helped us get things done faster. Second, it allowed us to compare the labels that different users attached to the same article and sub-select the ones where multiple users agreed with one another to improve the quality of our training data set.
A quick side note

Imagine you are stock analyst who wants to analyze the news associated with a particular company to better understand important trends and future performance.  Where do you turn to? Well entrenched tools such as a Google or Bloomberg immediately come to mind and these do a pretty good job of delivering relevant results.  That said, there is still a lot of effort involved in going from that initial information set to a synthesized view of it that can enable us to more easily find meaningful patterns and surface interesting signals. 

We still, and appropriately so, marvel at the ability of Google to quickly and efficiently help us find items that we are searching for.  However, in the context of a knowledge worker such as that analyst who has to distill large amounts of data into actionable insights, this experience doesn’t nearly go far enough.

Enter deep learning, a branch of machine learning. Deep learning has been setting the pace for the last several years within machine learning - largely as a result of the exponential increase in the right kind of cheap computing power that arrived with GPU based systems. My colleagues and I are big fans of deep learning because of its ability to help us attack unstructured data analytic problems and advance the knowledge cause. There are two big reasons underlying this:

  • Deep learning allows us to teach a computer by example. This enables a knowledge worker to throw examples at the computer and use them to teach it how they want their data to be analyzed and understood. In the context of unstructured data where it is often very difficult to codify all the underlying rules that make something interesting, this goes a long way. This approach allows us to approximate human intuition and is a big part of the reason why deep learning techniques deliver the best accuracies today.
  • Second, is the emergence of transfer learning.  Traditionally, the Achilles heel of deep learning has been the scale of the training effort required to teach the computer to do something well.   Transfer learning, a particular branch of deep learning, leverages the fact that it is possible to drastically shrink the training effort to do something if you can leverage learnings from related tasks.  So, for instance, if you have access to pre-trained deep learning models that have learned to analyze text in generalized ways, its possible to extend these models with only a handful of examples and teach them to handle your specific text analytic tasks with great efficiency.  Transfer learning makes deep learning practical and allows for a rapid level of customization that the knowledge distillation task, given the breadth of its users and uses, demands.

So how does this all translate into real world impact?  We wanted to run a quick experiment to showcase the possibilities. Imagine if you will, that you are that analyst that wants to hone in on what’s happening to a stock by looking at its associated news and commentary. Instead of starting from a long list of articles that just mention said company, what if you could quickly build a news article tagging service that classifies and organizes the information in these articles in a manner that you specify and makes it easier for you to quickly understand the underlying content? Let’s see what this would take.
First, we defined a classification system for articles to our liking.  For the purposes of this exercise we kept things simple and ended up with the following taxonomy:

Corporate News: Articles describing events that happen to a company (product launch, event announcement, etc.)
Stock News & Analysis: Articles focusing on analysis of stock price movement and future predictions
Fundamental Analysis: Articles looking at the underlying business issues affecting a company
Sector/Market News: News or analysis that focus on broad trends affecting the company
Other: Articles that don’t fall into any of other categories.

To build a model to organize information this way, we needed training data.  Next we gathered a dataset of about a thousand finance-specific news articles from various public websites and loaded these into our data labeling tool.  Long story short, it took five of us a couple of hours to attach labels to some of these articles and create a few hundred high quality examples that could serve as the training data for our model. Our reason for using multiple data labelers in this task was twofold. One, it helped us get things done faster. Second, it allowed us to compare the labels that different users attached to the same article and sub-select the ones where multiple users agreed with one another to improve the quality of our training data set.

A quick side note. Our labelers only agreed about 60% of the time in our experiment which speaks to both room for improvement in our simple classification scheme, as well as the difficulty of doing this task even from a human perspective.

We then trained a custom classification model with our training data and quickly had a custom model that had learned to organize articles just like us. Throughout the training process, we monitored the accuracy of our model and stopped training when we got to an acceptable level of accuracy (70%+ in the case of our quick experiment, compared to 20% accuracy of random guess).  The fact we could get to these levels of accuracies with just a few hundred articles highlights the power of transfer learning. To get higher accuracies we could have continued to add more training data and kept retraining the model.

So how does this all play out?  Once we had our trained model, we ran a simple new search for articles (that were not included in our training data) related to the stock – which happened to be Pepsi. We then used our model to organize these search results accordingly. 

We could keep going further and just as easily create additional models (e.g. filters to identify particular kinds of articles or language discussing specific concepts) to pull out more artifacts as well as bring our own private data into the mix such as emails, proprietary data, etc., to help us extract greater value from the information at our disposal.

Deep learning is a very powerful tool that is starting to put many possibilities within our reach. Transfer learning gives us a way to avoid the high cost of deep learning for many tasks, and with easy-to-use tools, it is now possible for anyone to quickly teach a machine to “pre-process” many things in a way not too dis-similar to what a skilled personal administrative assistant would do for you - but with added consistency and at far greater scale and speed. The net impact is that we can spend more of our time on value-added tasks and advance our goal of generating maximum insight from all the information we have access to.

(About the author: Vishal Daga is chief customer officer at indico.io, a start up in the machine learning and artificial intelligence space. In his role, Vishal works closely with indico’s enterprise customers to help them extract meaningful insight from unstructured data in applications such as sentiment analysis, social media monitoring, content filtering, content classification, recommendations, and personalization. Prior to indico, Vishal was in the data analytics space with IBM and Netezza.)

Register or login for access to this item and much more

All Information Management content is archived after seven days.

Community members receive:
  • All recent and archived articles
  • Conference offers and updates
  • A full menu of enewsletter options
  • Web seminars, white papers, ebooks

Don't have an account? Register for Free Unlimited Access