Text mining, voice mining and unstructured data analysis

Contributed by Tom Reamy, Kaps Group

As Chief Knowledge Architect and founder of KAPS Group, I lead a group of knowledge architecture, taxonomy, and text analytics consultants. This is my third year attending and second year as a speaker at the Semantic Technology Conference (#SemTech on Twitter).

Presentation themes revolve around evolution and maturity. The evolution theme is that there are no really big, revolutionary changes being touted this year. I like that, as I’m a little tired of and skeptical of “This Changes Everything!” mentality. The biggest buzz is about Schema.org – more about that later. Another evolutionary theme can be seen in the size of the conference – more people are attending this year than last.

The maturity theme provided most of the excitement instead of major new technologies or breakthroughs, the excitement was generated by real life applications with real business value. Three keynote talks were good examples of this:

• The first application was developed by Amdocs, a telecom provider, to add intelligence to customer support software.
• The second was the BBC World Cup web site that used a variety of Semantic and ontology elements to drive the web site.
• The third was a DoD presentation that was very slick – a video introduction with all sorts of special effects and a discussion of what they are up to – and they have more of everything than anyone else in the world.

The Amdocs example has a personal and a SAS connection. I did a proof-of-concept (POC) for Amdocs last year around choosing a text analytics solution to add to their intelligence platform. The outcome of that POC was that they chose SAS Enterprise Content Categorization and SAS Sentiment Analysis products.

At the conference, I gave a “Text Analytics Evaluation” talk which presented and argued for the approach that I used with the Amdocs project as well as other, earlier projects. So, the standing room only crowd (OK, it wasn’t a huge room) got to hear how to evaluate text analytics software in general and look into a case study in which SAS Text Analytics was clearly superior to any of the 30 other vendors, including IBM’s offerings.

It was heartening to see so many people attend a text analytics talk at a Semantic Web/Ontology focused conference. More people seem open to the idea that combining ontologies and text analytics gives you the best of both worlds. For a man with 20 years in information architecture, enterprise search, intranet management and text analytics consulting, that’s very good news.

The science of mimetics (or knowyourmeme.com) provides a framework for understanding and potentially predicting changes in the structures of our world including the structure of the language we use.

Collocations are words that frequently occur together, often in a particular fashion that is subject to the linguistic rules of the language of expression. Collocations often reflect the evolution of a language. For example, “high rise” becomes “high-rise” which in turn becomes “highrise” or even “HighRise”. In English composition the hyphen is often used in a transitionary phase as two separate words progress mimetically to become one word.

There are often unobserved and often strict rules in the development of acceptable collocations in English. For example, we accept “High Rise” but would not accept “Tall Rise”. On the other hand, we would understand a “Tall Drink” but would view a “High Drink” in a completely different fashion.

SAS Text Miner has a collocation detection mechanism built into its operation. The identification of Noun Phrases and proper nouns in the text mining linguistic parser is optimized to identify acceptable collocations, as discussed above. The concept link diagram is much more free wheeling: here collocations are identified simply on the basis of a calculation that computes the likelihood that the two identified terms will occur together. This likelihood calculation is a function of the number of times the words occur together.

It is essential to calculate the joint probability of two words occurring together since mimetically-inspired word combinations will only occur in specific conditions. Another way of saying this is: if the co-occurrence of two words is not sufficiently unusual it is highly unlikely that it is a salient term combination.

We used SAS Text Miner to look at the Vaccine Adverse Events Reporting data. Here we observed that the terms “Tylenol” and “Fever” have a salient collocation probability: In this example, the numbers show that the term Tylenol occurs in 173 documents with the term fever. When you point to these terms, the fever term at the center shows 1620/1620 while the term Tylenol connected to fever shows 173/399. This indicates that the term Tylenol is highly associated with the term fever. The term Tylenol appears with the term fever in 173 of the 399 documents containing the term Tylenol. This is a strong enough collocation to cause the two terms to be displayed in the concept link diagram that illustrates collocations in the Text Miner interface.

The Text Miner software is being exceptionally clever here: neither "Tylenol fever" nor "Fever Tylenol" would likely appear as acceptable noun groups from a parsing point of view. Yet Tylenol and Fever could easily evolve mimetically just as “Xeroxing” and “photocopying” have evolved mimetically (due to the strong marketing of the Xerox brand).

The Text Miner documentation describes the calculation of the collocation probability used in the concept link diagram as follows: For a given pair of terms, say a and b, the strength of association between these terms is computed using the binomial distribution.

Let
u be the number of documents containing term a
v be the number of documents in the collection
n be the number of documents containing term b
k be the number of documents containing both term a and term b

p = u/v be the probability that term a will occur when term b occurs assuming that they are independent of each other.

More information on collocation is given in Wikipedia.

A number of references, produced below, are cited on the Wikipedia link.

References
Dunning, T. (1993): "Accurate methods for the statistics of surprise and coincidence". Computational Linguistics 19, 1 (Mar. 1993), 61-74.
Gledhill C. (2000): Collocations in Science Writing, Narr, Tübingen
Firth J.R. (1957): Papers in Linguistics 1934–1951. Oxford: Oxford University Press.
Sinclair J. (1996): “The Search for Units of Meaning”, in Textus, IX, 75–106.
Smadja F. A & McKeown, K. R. (1990): “Automatically extracting and representing collocations for language generation”, Proceedings of ACL’90, 252–259, Pittsburgh, Pennsylvania.
Hunston S. & Francis G. (2000): Pattern Grammar — A Corpus-Driven Approach to the Lexical Grammar of English, Amsterdam, John Benjamins
Hausmann F. J. (1989): Le dictionnaire de collocations. In Hausmann F.J., Reichmann O., Wiegand H.E., Zgusta L.(eds), Wörterbücher : ein internationales Handbuch zur Lexicographie. Dictionaries. Dictionnaires. Berlin/New-York : De Gruyter. 1010-1019.
Moon R. (1998): Fixed Expressions and Idioms, a Corpus-Based Approach. Oxford, Oxford University Press.
Frath P. & Gledhill C. (2005): “Free-Range Clusters or Frozen Chunks? Reference as a Defining Criterion for Linguistic Units,” in Recherches anglaises et Nord-américaines, vol. 38 :25–43

The Sentiment Analysis Symposium (#sas11) in NY this week was even better than last year.

Focusing on practical uses of sentiment analysis, as this field continues to mature, was the overriding theme. The presentations and attendee comments stimulated great conversations.

The show’s customer-experience focus was equaled by financial investment firm interests represented in the audience. Both groups need to measure value from the analysis of sentiment.

Attendees and speakers confirmed that sentiment is not always impactful to an organization – nor its cousins of emotion, trust, feelings, and alike. How someone feels does not always affect what they do – but when it has an important impact, it is worthwhile to analyze it. Having done historic work in buyer behavior, there seems to be a number of parallels to the classic notions of utility theory in this – and I’m still working through the meaning of the ‘order of the good’ in today’s buying environments.

If I were to classify the over-arching types of applications discussed, it (perhaps to no surprise) seems to align with one item that I shared with the audience – that sentiment analysis can be used for:

• Validating
• Learning
• Changing

Validating data with Sentiment Analysis, often with other text analytics capabilities as well - can help inform and identify the accuracy of information that has been structured, say by the selection of a code identifying the severity of a problem heard by a call center representative. Going through this exercise can help ensure that the decisions based on reported "high priority items" are in fact, the most important issued to be focused on. Putting that into production can include scoring news feeds, web crawls and internal file systems scours with the sentiment models (applying any additional information, say through categorization), and updating search systems to improve relevancy retrieval.

Learning from sentiment analysis - many of the symposium talks described how new learning is gleaned by applying sentiment analysis. But as @KDPaine pointed out – it is important to focus on the so what factor. The reports are informative, but how do you make this information actionable? By integrating with the technologies that you may already have (dashboards, office productivity tools, adding to data stores as new elements or surfacing in mobile devices, etc.), you can share this new insight. In a production context, you can modify activities based on this learning by integrating it with customer conversation centers, campaign management systems or other relevant workflow applications.

Changing behavior with sentiment analysis - from a proactive standpoint is possible with the inclusion of the results from sentiment analysis into analytic models. It is only through the analytic application that you can change the behavior yet to occur, as you first must have the information on what will most likely happen before you can modify what you do. When sentiment affects that behavior, it is a key ingredient in reaping significant value from sentiment analysis.

I confess, the other reason for writing this is to finish saying what I didn’t in my symposium talk: that yes, Stevie Wonder said it best “We all have the ability. The difference is in how we use it.” {And thus the album reference in the bottom corner in the slide above}. The life of sentiment analysis in creating a competitive advantage is in how you use it differently than your competitors (as they can scrape the same web sites that you can for their sentiment monitoring).

Finally, another big Thank You to @SethGrimes for chairing the Sentiment Analysis Symposium. I hope to see you all at the Text Analytics Summit in May!

Till then, I’m interested in ways you measure the value of sentiment analysis in your organization. Write and tell me your most surprising finding.

All types of businesses and government agencies are taking advantage of information buried in previously untapped text documents (surveys, product reviews, on-line forums, emails, instant messages, articles, and social media).

Customers I work with are interested in protecting their brand equity, increasing customer satisfaction and loyalty, and reducing risks by carefully guarding their reputation. Top executives attempt to manage all kinds of risk within their organizations. Many of these risks are financial (credit risk, market risk, liquidity risk, insurance risk), but others are non-financial.

Reputational risk is one of the highest non-financial risks identified in the “Global Risk Management Survey – Sixth Edition” executed by Deloitte in 2009. That survey identified that 81 percent of the respondents were attempting to manage reputational risk.

A strong reputation can be a key competitive advantage. How many really understand what customers, analysts, and key opinion leaders are saying about them, their products and services so they can address issues? A good reputation can take decades to build, but it can be ruined very quickly. The actions taken directly after an incident occurs are the key to whether those involved effectively protect their reputation.‪ Creating an on-going dialogue with your customers or stakeholders to listen and learn from their feedback, in good and bad times, will help you maintain your good reputation. This dialogue can enable you to communicate with customers and prospects at a significantly lower cost than traditional marketing -- and with increased speed and effectiveness. It can also enable a more rapid response to perceived customer issues and competitive threats.‪

Executives responsible for managing their organization's reputation must answer some tough questions.

• What are the issues that people are discussing about my business compared to the top competitors in my industry?
• Are my marketing messages resonating with my customers?
• Am I spending my marketing dollars on the right channel to reach my target audience?
• What actions should be taken to improve my products, services, or employees based on the feedback I am hearing?
• What effect will the feedback have on my brand and my reputation in the marketplace?

Your stakeholders have spoken to all of these areas but are you listening? The answers to these questions are too often buried in massive volumes of stored text. Text is a largely unused asset in many organizations, but it offers far reaching impact on sales and marketing, customer service, and product development.

It's time for text analytics to give voice to ways for managing reputational risk. Unlike any other technology, text analytics can help organizations interpret, summarize, and report on information contained in virtual mounds of stored documents.

Are you ready to uncover sentiments; positive or negative opinions, attitudes, or feelings that could affect your reputation and ultimately translate into financial impacts? ‪

Let's bring the voice of the market into the boardroom‪‪.