What is that intelligence we have in Watson Conversation service

Question

I've been working on watson conversation for sometime.Mainly conversation service responds for inputs given by the user depending on the way we train and it does that based on the intents,entities and dialog flow we define.

Internally it would look for some keywords that we define in intents and if it matches it would respond with the text that we give in dialog flow.

Even same kind of logic I can put in my application saying if has these few particular keywords,I can respond with this particular text and go to next stage and look for some words so on.

But what is that special intelligence that we have in this watson conversation service,to use this?

Can someone tell me this who has good knowledge on watson

Answer 1

But what is that special intelligence that we have in this watson conversation service,to use this?

I take it this is the crux of your question?

First thing to realise that the Intents component uses machine learning, it is not a rule based engine. This gives you huge advantages.

You need a fraction of training examples that you would versus a rule based engine. To give a real world example. The was a service called "Watson Dialog" (now discontinued). This was an NLP rule based engine.

In Conversation I created an intent with 20 example questions from real world end users. To get the same level of accuracy in Dialog required over 7,000 permutations of examples (even with pattern matching used).

Also because Conversation is machine learning, it can give good responses to questions it has never seen before. In a rule/word lookup system if it sees a question it has never been trained on, it can never answer it. It can also understand when a question is not related to the topic, when keyword trapping in a rule based engine would try to answer.

Of course this is all dependant on correctly training the intents.

---

The Entities component on the other hand used to be keyword matching, there is some added intelligence there (and more coming).

---

The Dialog component, you are correct. You can certainly create your own code to do simple logic flows (again more coming on this). There are points regarding this.

1. "Watson" is about democratising AI. It's target audience is non-AI developers and subject matter experts (SME). So it is designed to make it as easy as possible for SMEs, while make it easy to extend for developers).

2. Separating your dialog logic out of conversation and into your code, makes it considerably harder to maintain. You are causing a tight coupling with your code. So if you want to use a different language, channel then it means you have to update/convert your code.

Answer 2

I have been working on Symbl, and specifically Trackers as a new intelligence capability to solve some of pain points of creating intents on human conversations.

Watson Conversation and other services work great for chatbots but for human conversations I haven't seen anything else working out of the box. Curious to hear others here.

Answer 3

Watson employs rule-based AI and a broad range of other mechanisms (e.g., information retrieval systems) to identify and assign feature values to candidate answers. The machine learning system then learns how to combine the values of these features into a final score for each candidate answer (thus allowing Watson to select one as its top answer and to determine how confident it should be in that top answer). What you describe just is one of the many classical AI algorithms that are used to analyze candidate answers in Watson, and whose results are considered by the machine learning algorithms when assigning a score to those answers, like confidence in Watson Conversation Service.

According to your ask, Watson will understand and categorize with a confidence level.

How IBM Watson works here. Official Video from IBM.

Reference: here.

Answer 4

https://github.com/joe4k/wdcutils/ has some tools to measure WCS's performance. WDC Jupyter notebooks report commonly used machine learning performance metrics to judge the quality of a trained model. Specifically, the WDC Jupyter notebooks report machine learning metrics that include accuracy, precision, recall, f1-score, and confusion matrix. If you’re interested in more details on these various metrics, please consult the “Is your chatbot ready for primetime?” blog on https://developer.ibm.com/dwblog/2016/chatbot-cognitive-performance-metrics-accuracy-precision-recall-confusion-matrix/

Answer 5

I think this is a very good question, and I really do not understand the downvotes here (have been -2 on Feb 13th).

As shown in Watson Conversation: What is lost when restoring a Workspace from a JSON "dump"-file? it seems that whatever is trained/improved in a WCS system, is only written into the current Workspace, leaving the WCS system dumb as it is, when this workspace is removed again.

Given that I have understood all this correctly, creating a workspace is just building up the STATIC logical base upon which a rather simple "AI" is just pattern-matching additional utterance to already labeled examples and intents.

So for me, the term "AI" is overrated, as it turns out for me, that Watson Conversation is just a machine, that is "just" matching incomplete or disturbed input (utterance) to predefined example sentences (intent attached examples and counter-examples) - something we have done with "fuzzy logic" and "pattern matching" algorithms. So, my definition of "intelligence" is another one.

All the dialog part is outside the AI scope anyway. The Dialog-part is just programming for non-programmers upon visual aids.

Maybe my comment here is a rant. I do not know. I am not an expert. I am just watching the system in real life. And THIS is my perception I have got so far. Eager to learn that Watson is smarter than it currently appears to me.