Language and the human brain

How many geese in this picture?  You didn’t need to count. Your answer probably contained the words “approximately”, “average”, “almost”, “sort of”, “guesstimate” or “about”.  One of the most powerful features of your brain is that it does not treat language as math, a series of binary yes or no formal constructs.  Humans are masters of writing the same idea in many ways, understanding what you meant even if you didn’t say it perfectly.  You also know when someone is being so careful with their words, they’re lying (we’re all tested on this one daily). This critical skill is used by analysts all the time.

Many technical approaches to textual analysis try to convert sentences into math or use the presence of specific words to return a yes or no result.  Our approach at BEA overcomes this issue with software agents that compare arrays (models) of patterns and return “similarity scores”.  These normalized scores allow us to determine whether a paragraph is about the same as another and by how much. 

In my last post, I discussed the nature and definition of concepts and how our solution is built to find them. Whether it’s a concept you’ve created or an example you’ve found in a document, if you can’t describe it you can’t find it.

Describing concepts

You can spot a concept when you read one, but learning to describe one isn’t as easy as it sounds and its rarely exact. Our AI platform enables us to create and teach “intelligence agents” to “read” documents to score similarity to concepts.

The first step in the process is to provide examples in a simple text box. This is surprisingly difficult the first time because our brain is never reading text without injecting our own education, bias and assumptions into the process.  Unfortunately the software only sees exactly what you provide it, no more, no less. When one user asked us how to teach the concept of “fraud” to our agents, we had to take a step back.

Risk and opportunity in legal issues

In the financial filings we focus on, concepts are often expressed indirectly, through a description of the act or its consequences.  The objective of the analyst is to find indications of wrongdoing or their cover up in company filings, earnings calls, news and social media. The critical evidence is never a clear statement like “I just gave my friend inside information on our earnings announcement so they could trade ahead of our disclosure.”

People don’t recite the definition of the crime when they talk about it. The language used is always more subtle and disguised. It also varies dramatically from one context to another (email vs. interview transcript for example). Tweets and text messages are full of acronyms, slang, phrases and partial sentences.

The guilty party is usually aware of the act and tries to avoid being discovered. He is more likely to say, “Hi Dave, here are some stats you might find interesting.” Is he talking about the company earnings report or his fantasy football team? If he says, “we’ll have a big surprise for you tomorrow”, is it a surprise birthday party or a merger announcement?

Kant’s tree concept (again)

To underscore this point, let’s revisit the example of the tree concept from my last post. The concept of a tree abstracted from descriptions of many trees is clear enough. The challenge in language analytics and many (most) language processing problems is that we are looking for the indirect effect or consequences resulting from the existence or actions of the tree.

This is easier to understand with examples. “The fall colors in New England are beautiful this time of year”.  Or “We need to get some shade for the yard at the nursery.” The concept of the tree is there but if I was searching for trunks, roots, branches and leaves, the “criminal” tree would escape detection!

Context is key

The context of the concept is critical.  Defining the context for the software can be complex.  But in most cases the patterns are there once you clear away the noise.

Capturing context can be as simple as setting filters.  For example, documents between specific individuals during specific time periods during which they had access to the information, resources and counterparts needed to commit wrongdoing.  Metadata from documents and entity attributes like organization, location and titles are commonly used for this purpose.

Where possible, we extract context from the source documents (and the source) to ensure that the more complex contextual factors are incorporated automatically into the intelligence agent. For example, the same person in the same period does not use the same language on twitter as they do in email.  Our powerful “almost engine” is what makes our system resilient and with our user in charge of how tight to set the “almost” meter, it adapts to a range of problems.

Is “big data” the answer?

We are all guilty of injecting our own bias and filters into understanding language. Good technical solutions capture the richness and subtlety in the context of the communications and ensure consistency of review.  This is not a problem solved by more and bigger data. 

We believe analysts have more than enough information.  They just don’t have enough time to find what matters.  BEA can help with that.

Want to know how we would use “almost” to solve your problems?  Drop us a line.

Tom @tom_semantic

KDD Analytics and Boulder Equity Analytics are partnering to deliver collaborative artificial intelligence to the financial and competitive analysis industries.

The post Good Concept Detection Requires an “Almost Engine” appeared first on KDD Analytics.

Our partner at BEA takes textual analysis a step further and teaches artificial intelligence (AI) software to find concepts and themes, not just words.  It’s one thing to find all occurrences of the word “decline” in an earnings call transcript.  It is another thing altogether to understand the concept of “decline” within the context of a paragraph.  Is it a decline in sales?  or a decline in bad accounts?

In another guest article, Tom Marsh, CTO at Boulder Equity Analytics (BEA), talks about keyword vs. theme detection and why “concepts are key, not words”.

A critical skill for the analyst during earnings season is detecting changes in the key indicators or themes for a company and its peers. Keyword detection is often passed off as theme detection but it’s not and the difference is critical.  Here at BEA, teaching software (AI) to find themes buried in SEC filings, earnings calls and press coverage from investor relations is a critical technology advantage.

First, understand the terms. Analysts tell us that with all the buzzwords and claims by vendors, it’s hard to understand the difference between real and apparent performance.  For us, the goal is to replicate an expert analyst’s ability to read and understand a document, whether its a filing, earnings call or interview.

While there’s more, in this post I want to make sure we understand each other when we use the term “theme”, “topic” or “concept”.

What is a “concept”?

Since we claim to teach software agents to find “concepts”, let’s check the definition of the term “concept” to make sure we are using it correctly. While I didn’t expect this to lead me all the way back to Philosophy class with references to Kant, Locke, Mill etc., our approach and use of this term are fundamentally consistent with the excerpts below from Wikipedia.

Concept– definition from Wikipedia

A concept is a general idea, or something conceived in the mind.

Notable definitions:

John Locke‘s description of a general idea corresponds to a description of a concept. According to Locke, a general idea is created by abstracting, drawing away, or removing the uncommon characteristic or characteristics from several particular ideas. The remaining common characteristic is that which is similar to all of the different individuals.

John Stuart Mill argued that general conceptions are formed through abstraction. A general conception is the common element among the many images of members of a class. “…When we form a set of phenomena into a class, that is, when we compare them with one another to ascertain in what they agree, some general conception is implied in this mental operation” (A System of Logic, Book IV, Ch. II).

Philosopher Arthur Schopenhauer argued that concepts are “mere abstractions from what is known through intuitive perception, and they have arisen from our arbitrarily thinking away or dropping of some qualities and our retention of others.” (Parerga and Paralipomena, Vol. I, “Sketch of a History of the Ideal and the Real”).

By contrast to the above philosophers, Immanuel Kant held that the account of the concept as an abstraction of experience is only partly correct. He called those concepts that result from abstraction “a posteriori concepts”.

A concept is a common feature or characteristic. Kant investigated the way that empirical a posteriori concepts are created.

“The logical acts of the understanding by which concepts are generated as to their form are:

1. comparison, i.e., the likening of mental images to one another in relation to the unity of consciousness;

2. reflection, i.e., the going back over different mental images, how they can be comprehended in one consciousness; and finally

3. abstraction or the segregation of everything else by which the mental images differ …

In order to make our mental images into concepts, one must thus be able to compare, reflect, and abstract, for these three logical operations of the understanding are essential and general conditions of generating any concept whatever. For example, I see a fir, a willow, and a linden. In firstly comparing these objects, I notice that they are different from one another in respect of trunk, branches, leaves, and the like; further, however, I reflect only on what they have in common, the trunk, the branches, the leaves themselves, and abstract from their size, shape, and so forth; thus I gain a concept of a tree.”

— Logic, §6

Optimization of AI software

We worked with our ai-one partner to optimize their AI for this task. 

At the core, our application processes each line of text much the way our brains do it, learning the patterns of language, the “key” words, their importance and the words most closely associated with them. The AI provides commands to extract as an array those key words and associations, their direction and values (strengths).

While the ability to score the similarity of concepts is important, my observation from years of applying it to problems from NASA (Topic Mapping pg.198) to SwissRe is that it’s even more proficient at filtering out the noise, giving lowest values to the unimportant words and associations.

Filtering is fundamental to our brain’s ability to find the topic that’s important.

Locke describes a concept as an idea “created by abstracting, drawing away, or removing the uncommon characteristic or characteristics”. This is very close to the way our solution builds a model of a concept after learning from the examples provided to teach it. The fingerprint we extract from that text is an array that represents the concept in the same way. The similarity score for that comparison is a powerful attribute we use in a number of ways to deliver a great user experience.

Building powerful qualitative analytics for financial analysts and investors starts with the right core technology. Finding concepts buried inside documents is the first part and foundation of extracting actionable insight.

Now that I think about it, maybe Philosophy 101 wasn’t a liberal arts waste of money after all.

Tom

@tom_semantic

KDD Analytics and Boulder Equity Analytics are partnering to deliver collaborative artificial intelligence to the financial and competitive analysis industries.

The post Concepts are Key, Not Words appeared first on KDD Analytics.

AI, artificial intelligence, has been through several boom and bust cycles.  Today the pronouncements are everywhere with AI coming soon to everything from medicine to your underwear.  For those of us laboring in the dark for years, it feels good to be the most popular kid on the block.  Warning!  What the marketing gods grant, they can take away overnight.

 I have been lucky to work with AI technology in language for the last six years.  As we worked to find traction, I was exposed to a huge variety of application problems.  Now focused on financial reporting and compliance, my favorite inbound call was from a fund in the UK with $11B AUM.  

“We have been reading about all the wonderful things AI could do and we want to purchase some ‘AI’”. 

They did not have a data repository, everything was on analyst PCs and not the cloud and they had no analytics capability.  While we completed a small proof of concept, it failed to get approval because “our analysts say they already read everything anyway”.  Could have seen that coming (good news is that it was part of the inspiration to start Boulder Equity Analytics).

AI is not just software

If you’re in the field, you’ve probably run into the same request, either from a prospective customer or from your own management.  Maybe your marketing department has already asked you for some AI to put into a product or, worse yet, added to the website. Now you’re supposed to figure out what sort of machine learning algorithm you can put in the product.  It’s just software, right?

This feels like the ’90s when the internet came online and everyone had to have an internet “strategy”.  Executives that didn’t even know how to turn on a PC were jumping on the bandwagon; most thought a website was an internet strategy.  

Keep it grounded

My request to AI practitioners is to keep it grounded and fight back. Don’t be co-opted into the marketing BS and don’t let management that has no experience with AI try to build an “AI” strategy from a 15 minute YouTube session. Read the technical experts that are debunking the exaggerated claims (like this piece by Yoav Goldberg) so you know the real state of the field. Talk to experts and ask questions, challenge them on the fail modes, the testing and the risks.

Walk before you run

If you’re in management, know that this is going to change your life, your organization, your markets and your customers.  The claims are real but do your homework.  Implementing an “AI” strategy is not a software project or something you can buy.  You need to start small, build a small team and execute some very small projects.  You do not need a big budget but you do need AI technology, domain expertise, access to the data and good UX people. Before attempting this, you should already have a good analytics foundation and culture (at least in the business you’re starting with).  

That said, don’t hold back.  Everything you’ve read about the exponential nature of the progress and its impact is true and coming fast.  The impact will be personal. If you don’t start right now, you will get run over by the freight train that is AI powered innovation. 

Tom (@tom_semantic)

KDD Analytics and Boulder Equity Analytics are partnering to deliver collaborative artificial intelligence to the financial and competitive analysis industries.

The post Buzzword Overkill – AI is Not a Thing People, it’s a Discipline appeared first on KDD Analytics.