A Colloquial Treatment of the Product Performance and Outcomes Aspect of Mining of the Blogosphere for Brand Services Monitoring.
The author’s recently posted monograph, “ Employing Advanced Natural Language Text Processing to Provide Guidance to Mid-Market Multiline Dealers and Product Distributors”, has generated a great deal of blog traffic, but some have asked for a more brief and colloquial treatment.
The following is a basic description of where the industry is, or has, led itself, and where the author thinks the industry should be going, based on research recently conducted for a telecom industry client. The resulting conclusions do not betray any verbatim strategic conclusions that the author provided to the client.
The Text Mining Industry, and in particular those specializing in brand monitoring services, have come to focus on ‘sentiment’ as the metric du jour. We all know the typical meaning of sentiment, it’s how we feel, or how we express how we feel. The problem occurs when we create machine scored metrics of a very human thing, such as sentiment, and expect the mined results from the blogosphere to make sense.
Other than the linguistic problems of generation (That is so bad [in one generation] means bad, in another it means really great – Pimp, is (a derogatory) in one generation and quite amazingly, good in another [ we be each other’s pimp] (advocates), [pimp my ride] (make fancy)). One can find even more subtle examples of how linguistic subtleties might confound an algorithm. There are many more robust social media metrics that far exceed the reliability of sentiment, or that can augment sentiment in order to strengthen the ultimate guidance that is being sought via the mining of the corpus in question, i.e., the blogosphere and/or public user forums.
But, linguistic problems aside, it is the role that the early text mining sector entrants have cast themselves in as ‘brand monitoring surrogates’, that is really a point of contention. Who is concerned with monitoring brand? Why, it’s the brand owners of the Fortune 1000. These large companies, mostly in the durable goods sector (for who blogs about toothpaste and other consumer packaged goods?), have had unfettered access to the best brand monitoring and consulting practices. AC Neilsen, Arbitron, and Gallup are the giants of this industry, but there are others.
Creating text mining services for the Fortune 1000 has been a high latency, fussy business. Whereas professional brand monitoring is based on actuarially sound models and standardized methods of sampling, these recently innovated ‘social media text mining services’, often have account reps, and sometimes, computational linguists (gasp), work with clients to identify verbiage that that does, or does not, express sentiments concerning products of interest and brand issues of concern. But there is a problem:
Catering to these brand owners is, as previously stated, a high latency business; contracts from the leaders sometimes take 2-4 weeks in the sales cycle, and 2+ weeks to setup the query and dashboard reporting. Furthermore, brand ownership is limited to the relatively small group of brand owners who are used to sampling brand awareness, and regularly avail themselves of brand equity practices, such as consulting. Such brand consulting businesses often make their nut by getting a percentage of the ad buy. This brave new world of ‘text mining of the blogosphere’ is a curiosity that has not made significant inroads into the brand monitoring business – maybe to the tune of a very optimistic $100M, compared to the entrenched brand services billions. There is also ample evidence that the clientele served, and the investors at equity in such new age ventures, are tiring quickly of the model and results.
So, what is needed for text mining of the public corpus to succeed? First of all, to turn the attention of these services from the brand equity owners, to the branding recipients – those who must deal with the customer’s perception of branding, product performance outcomes, and interactions with the entire spectrum of the product’s touch points – service, warranty, dealerships, etc.
Who are these prime recipients of brand decisions? Customers, certainly, but from the point of view of a web based service to analyze the public corpus, the true targets are the brand intermediaries. These intermediaries are multiline retailers and distributors that span the gamut of local shops, regional retailers, and national department stores and distributors. These are the true recipients of branding decisions, and have had very little guidance to steer their decisions as to whether or not to add or drop product lines, take advantage of ‘spiffs’ (incentives that cover cooperative advertising and floor-plan financing), or any decision affecting what brands to carry and promote.
Brand intermediaries, or the mid-market, make up a vastly more numerous and influential group than the brand owners. In the author’s previously sponsored research, a local and telephone survey, augmented by interviews with the current text mining leaders and their clients, found that the likely period of closure for a services contract would be far less for brand intermediaries. Whereas the brand owners have been reluctant participants in the fledgling electronic brand monitoring sector, the mid-market composed of the brand intermediaries is much more likely to subscribe to web-based subscription service that monitors customer perceptions, service, and dealer performance outcomes, as well as all things impacting brand and product perceptions.
The mid-market has different needs than the brand owners; intermediaries need quantified guidance as opposed to a vague ‘sentiment bar-graph’. Brand intermediaries of the mid-market also need a way to take into account the aforementioned spiffs – is it really a good idea to increase inventory of a certain product just because the manufacturer is offering, say, $1,500.00 more in coop advertising?
Now, the question turns to what technologies and methods can be applied to measure the customer’s perceptions of product performance outcomes, across all aspects of the experience? Certainly, the current sentiment offerings are woefully inadequate. What is needed is a way to detect archetypal ‘models’ of declarations stating redress sought, pronouncements of value relating to a particular experience, and public advisories that are brandished by the disaffected. We see these linguistic models all the time in customers service Business Intelligence databases, so why not find a way to model these interactions, mine them from (detect within) the public corpora, potentially match them against private corpora for statistical refinement, and finally, integrate them with product retailer program incentives in a pleasing, visual way that supports the multiline retailer’s decision making process.
A discussion as to which technologies are appropriate for conversation dynamics detection is a long, technical, and speculative diatribe. The author has plumbed the depths of current research on modular ontologies, applied semantic database tools, advances text mining, and all types of academic research. There is work being done. There is prior art. These research topics have not, however, surfaced within the current crop of text mining products…as of yet.
All of the foregoing, technicalities aside, need to be put into the perspective of customers – how they compose requests for redress within formal, structured interactions, as well as within self-published media. Properly detecting the dynamics of a conversation, applying proper metrics, quantitative weighting, and making the presentation of such an analysis useful, is analogous to the use of stochastic charts serving commodities traders on the NYMEX or Chicago BOT.
Imagine, if you will, that you are a large retailer of higher end mountain and trail bikes – you would, with this service, have the ability to specify standard brand and product queries via a web interface, with the additional ability to input variables such as program spiffs. The resulting report would look like a chart, with pull-down customizations covering time vs. dynamics. Against this data, the customer’s overall perceptions against time and program incentives would lead to a method, as yet only speculative, of advising the decision makers of a potential problem emerging in a soft market, and whether or not these incentives are a good bet, as opposed to a reduction in inventory, or outright abandonment of the line.
The author’s previous reporting on this blog are a great deal more technical – and much more needs to be done regarding the overall product strategy of such a service.
Such a service would greatly exceed the state of the current art for brand monitoring by re-targeting the services away from the brand owners (who already have alternatives in the form of established consultancies), to the brand intermediaries that make up a much large market, and who are more likely to avail themselves of practical text mining services that lead to bedrock advisories.
The author and his allied research counterpart, Dr. Charles Martin, are available to carry on this research.