I was hired to work on a 8+ digit contract to apply Watson, office mate and I (only engineers on the team) were laid off 2 months into our new jobs. Do what that company did, put cancellation clauses in any contract should IBM not deliver.
The taxonomy basically determined the escalation process within the company. As an example, think about issues with a cell phone. "Sometimes my phone volume goes down automatically" isn't as critical as "Sometimes my phone gets really really hot in my pocket."
The Good: They felt Watson was doing a good job classifying the customer feedback. Overall, they felt they could rely on Watson and could defend using Watson instead of human raters in a legal situation.
The Bad: The amount of time/resource investment required to get it up and running had been way beyond what they initially expected. They also felt various parts of the system could have been more open (e.g. Watson's back end database) to inter-operate with downstream systems. They also were exploring if other solutions (e.g. Data Scientists + Python) could do this and had started to see positive results.
Watson seems to me like it's mostly leadgen to sell IBM consulting services to companies that don't know how to build software in-house.
This is exactly what it is. It's a marketing product designed to help IBM sell the company overall as a big data company. The only people the commercials are targeted at are C-level execs who don't know the first thing about data science but know they have to invest in it to stay competitive. I don't work for IBM, but I'd wager a lot that the idea for it was created by a sales team and not an engineer.
That said, IBM is certainly not the only business to do this and it's not even clear that it's not a smart business strategy. I'd wager that most enterprise software sales use deception as a major part of sales strategy.
I work at a large bank and we've seen the Watson sales pitch every year for four years now. Every time they visit we ask them this question.
It would be great to hear from folks using it.
Voila, now everybody is "using" it
where "using" = sitting dormant.
He told me that IBM's data service (ie Watson) was not at all impressive. At the time (about a year ago) their service was not worth the pay. He told me all the solutions they provided, and the techniques they used (in terms of algorithms and infrastructure) were very easy to implement themselves (I mean, it was an internal data research team).
But, in our conversation, he told me that there were reasons to go for some AWS dervices (such as integration and provided apis) or Google (the amount of data their stuff in trained on, the infrastructure they have) could be pointed out as reasons to hire AWS or Google. He couldn't find any that suggested hiring Watson.
I'm not actually sure if this is a compliment, now that I think about it.
1. A brand - almost everything that has anything to do with analytics, gets labelled Watson. From cloud services to healthcare co-developments, to statistical software (i.e. SPSS) to anything (e.g. an e-commerce suite).
1.1 There also is "Watson Machine Learning" which is nothing more than SPSS in the cloud. (But of course you can use it for doing ML stuff)
2. A research project, which developed Watson as the system for the Jeopardy gameshow event and is still continuing to develop this system/technology further
3. There is a Watson core system, nowadays codenamed "Watson Discovery Advisor", which is architecturally pretty close to what was the original research system. There are only a handful of systems built on this architecture. It has nothing to do with the cloud services on bluemix, but is a massive parallel pipline in which pretty decent NLP gets combined with a set of hypothesis evidence scorer (various algorithms, from grammatical reasoning like simple LAT to ontological reasoning, lookups, tuple- and triple searches, indri, fulltext and other search approaches). This architecture is used for example, to power the Watson Cancer product. Every project like this is super expensive... speaking of 20 million +
4. A set of APIs in IBM's cloud platform ("Bluemix").
4.1 There we have NLC, a semantic text classifier (I personally find it quite good), a chatbot-system which pretty much is NLC with continuous training plus a handwritten dialog tree ("Conversations")
4.2 a machine-learning-ranking based search system ("R&R", "Discovery"), which combines an Elastic Search with a query language, paragraph handling, an NLP pipeline (see 4.3) and a supervised machine learning based ranker. The ranker is used to train the search to find relevant answers for a given natural language question (i.e. long tail question answering)
4.3. A NLP pipeline, which is NOT UIMA based, but based on the AlchemyAPI aquisition (which brought in the public taxonomy) and a SIRE based statistical machine learning relationship extraction approach. There is a training environment called "Watson Knowledge Studio" which lets you either program (dictionaries, rules) or train (click instances on sample texts) the NLP to your type system. This is running on Bluemix as "NLU" or in "Discovery"
4.4 And other stuff like image classifiers (party brought through the AlchemyAPI acquisition, partly from IBM's research in Almaden), Speech 2 Text, Text 2 Speech, Translation
6. An on-premise software called "Watson Explorer" which is a mixture of the Vivissimo aquisition (a search engine + portal + data integration) and IBM's homegrown UIMA pipeline and a nice analytical interface (data science tool for textual analytics - correlational analysis on text, NLP features and metadata with deviations, timelines, etc). This pipeline can be programmed with a rule- and dictionary based dev environment AND with the supervised machine learning based training environment Watson Knowledge Studio (->5)
7. Very few "products" like Watson for Oncology or ICPA, which are based on a variation of the above technologies and sold through specific channels
8. A Watson Health brand, dedicated to applying the Watson ideas to the healthcare market. They "own" Watson for oncology. They also have a big set of custom cloud APIs (some quite good and complex), like body part and drug tagging, adverse event detection, ICD10 coding, etc.
There is a huge sales folk in IBM, who does not understand the complexity and the breadth of portfolio the company has. Very often this results in overselling. The big Watson projects are pitched as a "this is what you get" plus the Watson APIs on the Bluemix cloud are shown as a "this is how easy and cheap you can start"... sales "forgets" to mention, that there is a huge gap between playing around with APIs and developing large systems like Watson for oncology.
Or they are pitching the big Watson story and then selling Watson Explorer, just because this is, what they get their sales-quota for.
Also the stuff requires really skilled people to actually forge a customer specific Watson system. I have seen many projects gone bad, just because some product managers think "this is my product and we solve the problem with just this" - instead you would have needed a few good techies and an skilled architect to develop a solution, leveraging multiple different APIs and a combination of cloud and on-prem (Watson Explorer) to achieve a good result for the customer.
If you know what the technology can do, do not oversell it, and have skilled people working on it, you can built great things. I myself have lead the implementation of a good dozens of pilots and a handful of production projects (which are still active!)
It's actually easy to use the API and I'm hoping to use more features for the my next project, which is a mobile app to get voice alerts of distributed IoT data from the cloud. Not sure if this answers the question, but I think the API is usable by businesses.
Seems like they'd better align the marketing and the offering if Bluemix was the "Watson gateway" so to speak. They could still sell and market Watson as the custom consulting they do, with Bluemix branded as an entry point to basic Watson functions. There's some speak of this in their product copy, but it could probably be more obvious.
Mixed results, so far on how accurate it is, but I also recognize that there are likely a number of improvements that can be made to the parameters that I’m setting for each processing run. I also have only started to play with the custom modeling.
Right now it’s a toy, but promising for our small automation needs.
Really though, I don't think bluemix/watson has anything GCP/AWS/Azure don't have at this point. Plus it looks like a bunch of their machine learning research got the axe.
The one I used for NodeJS:
It was useful to read the source code to figure out some aspects I didn't find in the documentation.