Virtual Assistant Update

 

We recently published “Virtual Assistant Update.” It’s a broad and not too deep update on virtual assistant technologies, products, suppliers, and markets from the perspective of the five leading suppliers: [24]7, Creative Virtual, IBM, Next IT, and Nuance. These are the leaders because they:

• Have been in the virtual assistant business for some time (from 16 years for [24]7 via its acquisition of IntelliResponse to four years for IBM).
• Have attractive and useful virtual assistant technology
• Offer virtual assistant products that are widely used and well proven.
• Want to be in the virtual assistant business and have company plans and product plans to continue.

The five suppliers are quite diverse. There’s the public $80 billion IBM and the public $2 billion Nuance. Then there are the private [24]7, a venture backed company big on acquisitions and the more closely held Creative Virtual and Next IT. Despite these big corporate-level differences, the five’s virtual assistant businesses are quite similar. Roughly they’re all about same size and the five compete as equals to acquire and retain virtual assistant business.

By the way, across the past 12 to 24 months, business has been good for all of the five suppliers. Customer growth has been very good across the board. Our suppliers have expanded into new markets and have introduced new and/or improved products.

Natural Language Processing and Machine Learning

Technologies are quite similar, too. All five have built their virtual assistant offerings with the same core technologies: Natural Language Processing (NLP) and machine learning.

Virtual Assistants use NLP to recognize intents of customer requests. NLP implementations usually comprise an engine that processes customer requests using an assortment of algorithms to parse and understand the words and phrases in a customer’s request. An NLP engine’s processing is guided by customizable and/or configurable deployment-specific mechanisms such as language models, grammars, and rules. These mechanisms accommodate the vocabularies of a deployment’s business, products, and customers.

Virtual assistants use machine learning technology to match actual customer requests with anticipated customer requests and then to select the content or execute the logic associated with the anticipated requests. (Machine learning algorithms learn from and then make predictions on data. Algorithms learn from training. Analysts/scientists train them with sample, example, or typical deployment-specific input then with feedback or supervision on correct and incorrect predictions. A trained algorithm is a deployment-specific machine learning model. The accuracy of models can improve with additional and continuing training. Some machine learning implementations are self-learning.)

Complex and Sophisticated Work: Consultant-led or Consultant-assisted

The work to adapt NLP and machine learning technology implementations for virtual assistant deployments is sophisticated and complex. This is work for experts: scientists, analysts, and developers in languages, data, and algorithms. The approach to this is work differentiates virtual assistant suppliers and products. The approach drives virtual assistant product selection. Here’s what we mean.

All the virtual assistant suppliers have built tools and package predefined resources to make the work simpler, faster, and more consistent. Some suppliers have built tools for the experts and these suppliers have also built consulting organizations with the expertise to use their tools. Successful deployments of their virtual assistant offerings are consultant-led. They require the services of the suppliers’ (or the suppliers’ partners’) consulting organizations.

Some suppliers have built tools that further abstract the work and make it possible for analysts, business users, and IT developers to deploy. While these suppliers have also built consulting organization with expertise in virtual assistant technologies and in their tools, successful deployments of their virtual assistant offerings are consultant-assisted and may even approach self-service.

So, a key factor in the selection of a virtual assistant product is deployment approach: consultant-led or consultant-assisted. Creative Virtual, Next IT, and Nuance offer consultant-led virtual assistant deployments. [24]7 and IBM offer consultant-assisted deployments. For example, IBM Watson Virtual Agent includes tools that make it easy to deploy virtual assistants. In the Figure below, we show the workspace wherein analysts specify the virtual assistant’s response to the customer request to make a payment. Note that the possible responses leverage content, tool, and facilities packaged with the product.

© 2017 IBM Corporation

Illustration 7. This Illustration shows the Watson Virtual Agent workspace for specifying responses from the bot/virtual assistant.

 

Which is the better approach? Consultant-assisted is our preference, but we’ve learned over our long years of research and consulting that deployment approach is a function of corporate, style, personality, and culture. Some businesses and organizations give consultants the responsibility for initial and ongoing technology deployments. Some businesses want to do it themselves. For virtual assistant software, corporate style could very well be a key factor in product selection.

 

 

 

 

Analytics in Radian6

This week’s report is our evaluation of Radian6, the component of Salesforce Marketing Cloud that does social monitoring, analysis, and interaction. Its tight integration with Salesforce Service Cloud—automatic creation of Cases and Contacts—makes it the obvious social-service choice to add to the customer service application portfolio of Salesforce CRM users

Customer social-service is all about monitoring customers’ conversations in the social cloud, identifying customers with questions, problems, and issues, and then interacting with those customers to answer questions, solve problems, and address issues. The number of customer posts and conversations in the social cloud that may be relevant to a business can be very large, ranging to thousands or even tens of thousands per week and, in the extreme, hundreds of thousands per day. Monitoring and analyzing all of them, identifying the (few) posts that require attention, and then handling each one individually and handling all of them consistently are daunting and complex tasks, daunting because of the sheer volume and complex by the diversity and nuance of language, breadth of topics, and depth of emotion (sentiment).

Most social-service products use third parties to monitor social posts, to crawl and search the key social networks and the hundreds of millions of blogs and forums where customers ask questions, get answers, and make comments.  The value-add of these products is in their analytic capabilities, capabilities that can “understand” the content of social posts. Natural Language Processing (NLP), sometimes called text analytics, is the technology that they most commonly use. And, also most commonly, each of them is built its own NLP implementation. Their companies are built on it, too. These NLP implementations are frequently patented and almost always proprietary. They’re the crown jewels of analytics companies. So, the selection of a social-service application usually involves the evaluation and comparison of NLP implementations, a difficult selection of sophisticated and complex technology.

Not the case for Radian6. It takes the opposite approach. Rather than leverage the data collection capabilities of third parties and apply its own analytics, Radian6 does its own data collection (The current version searches and crawls over 650 million social sources.) and leverages the analytic capabilities of third-party analytics suppliers to understand the content of social posts. (Radian does a bit of its own analytics, too, although its analytics are a bit basic and are not built on NLP.) These 14, third-party analytics suppliers comprise what Salesforce.com calls the Radian6 Insights Ecosystem, Insights for short. They apply their analytic technologies to the social posts collected by Radian6.

The 14 are:

• Bitext
• Communication Explorer
• Clarabridge
• EpiAnalytics
• Hottolink
• Klout
• LeadSift
• Lymbix
• OpenAmplify
• Open Calais
• PeekAnalytics
• Soshio
• The SelfService Company
• Trendspottr

Let’s take a little closer look at three Insights to get an idea of their capabilities.

• The Bitext Sentiment analytic perform Entity extraction and sentiment analysis for posts in Spanish (European and Latin American), Portuguese (Brazilian and European), Italian, and English using natural language processing technology (NLP).
• Clarabridge provides two analytics. Clarabridge Link Sentiment provides sentiment analysis of social posts in Chinese, Dutch, English, French, German, Italian, Portuguese, Russian, and Spanish using NLP; Clarabridge Link Classification applies a Universal Category and Classification model to social posts in Chinese, Dutch, English, French, German, Italian, Portuguese, Russian, and Spanish using NLP.
• OpenAmplify also provides two analytics. OpenAmplify Cust Svc uses NLP to identify social posts containing potential customer service issues and the topics of those potential issues. OpenAmplify uses NLP to identify sentiment, intention, and topics of social posts.

Salesforce.com offers these Insights like usage-priced cell phone minutes within the subscription licenses and their monthly fees for Radian6 Editions. (Editions are licensing tiers that bundle applications resources.) More specifically, Radian6 Editions include blocks of Insights partner credits. The analysis of one social post by a one analytic application from one partner costs one partner credit. At the low end, Marketing Cloud Radian Basic Edition includes 1,000 Insights partner credits. At the high end, Marketing Cloud Radian Enterprise Edition includes 500,000 Insights partner credits. Blocks of 10,000 additional Insight partner credits are available for a fee of $100 per month. Credits are expire every month (like cell phone minutes).

Insights’ suppliers set up pre-configured deployments of their analytic applications for access and usage by Radian6 licensees at runtime. That approach can be a disadvantage. For NLP based Insights, runtime access means that language models and processing configurations are those implemented by their suppliers for general-purpose usage, not language models and configurations of deployments tailored to the applications and vocabularies of specific businesses and their customers. For example, the Clarabridge Link Classification Insight uses a “Universal Category and Classification” to classify social posts. Analytic processing will still be quite useful, just not custom tailored.

There are also advantages to Radian6’s Insights approach of runtime access to analytic applications. Most significantly, Radian6 lets businesses easily combine and nest these analytics. For example, analysts might use the entity, fact, and event extraction capabilities of Open Calais to find posts relevant to a product launch and then use PeekAnalytics to identify the demographics of those posters. Also, specifying language models and processing configurations for NLP-based analytic applications is complex work, work that Radian6 users do not have to do to get much of the benefits of these sophisticated applications.

The approach to analysis in Radian6 is a significant differentiator and a key factor for selection. Radian6 delivers most of the power of a wide array of third-party analytic applications and the flexibility to use them separately or to combine their processing. Pricing is based on usage. Value is very good.

IntelliResponse VA

Accurate Answers with Fast and Easy Deployment

We’ve just published our Product Review of IntelliResponse Virtual Agent (IntelliResponse VA), the virtual assisted-service offering from IntelliResponse Systems, Inc., a privately held supplier founded in 2000 and based in Toronto, ON Canada. The report completes our latest research series on virtual agents/virtual assisted-service.

We’ve published evaluations of the four leading virtual agent offerings:

• Creative Virtual V-Person
• IntelliResponse VA
• Next IT Active Agent
• Nuance Nina Web (VirtuOz Intelligent Virtual Agent when we published. Nuance acquired VirtuOz earlier this year.)

Virtual agents implemented on all four can deliver a single answer to a customer’s question on web, mobile, and social channels. Expect the answer to be correct about 90 percent of the time.

Virtual agents deliver bottom line benefits. They can lower cost to serve as compared to live agents and they can improve customer sat by improving the speed, accuracy, and consistency of the answers to customers’ questions.

 Contrast virtual agents with search and knowledgebase approaches that deliver many answers and leave it to the customer to pick the correct one. This single correct answer makes virtual agents useful for answering many kinds of customers’ questions, certainly customer service questions but also questions about your business and about your business policies, processes, and practices, about your products, and everything about your customers’ relationships with you—accounts, orders, bills, and passwords, for example. They can be your agents for marketing, for sales, and for service.

 Like your live agents, it takes time and effort to get virtual agents ready to engage with your customers. You have to give them the knowledge about the business areas that they support. You have to train them to understand your customers’ questions and to correlate or match those questions with the correct answers. The knowledge is contained in/represented by the items in their knowledgebases, their store of predefined answers. Anticipate the questions that your customers will ask, specify the answers, and store them in the virtual agent’s knowledgebase. All four virtual agent products have knowledgebases and provide tools and facilities for creating and managing answers. Your customers’ questions will change and evolve with their relationships and with changes to your offerings of products and services and to your business. Virtual agent’s knowledge has to keep up with those changes (just like live agents’ knowledge).

Training virtual agents to understand your customers’ questions and to correlate/match them to correct answers is the harder part. Virtual agents use very sophisticated and complex technology to analyze customers’ questions and to match them with the answers in their knowledgebases. Analysis and matching is the core processing that virtual agents perform. Analysis and matching technology is the virtual agent supplier’s core IP, its secret sauce. Each of the four has patented some or all of this technology. The suppliers want you to appreciate the sophistication and power of the technology. They don’t give you much detail of what it does or how it works.

(We describe and evaluate how a virtual agent analyzes and matches questions in our product review. We actually read many of the suppliers’ patents to help us understand the technology. In our reports, we describe it a bit, but we focus on what you’ll have to do to use it effectively.)

Creative Virtual V-Person, Next IT Active Agent, and Nuance Nina Web use Natural Language Processing (NLP) technology for their analysis and matching. Each has its own NLP implementation. NLPs perform computational linguistic analyses on customers’ questions, parsing for subjects, verbs, object, and qualifiers, extracting entities, identifying actors and roles, and codifying relationships. This is sophisticated and complex processing.

For a successful virtual agent deployment, you provide critical input to your virtual agent supplier’s NLPs, for example:

• Words that your customers will likely include in their questions

• Misspellings, typos, slang, idioms, ad stems for those words
• Conditions/rules/expressions for how your customers combine words into phrases
• Parameters for configuring the NLP processing

If you don’t specify the actual words and their various alternative forms that your customers use in their questions, then your virtual agent cannot deliver answers. Complete specification of your customers’ vocabularies is critical. Virtual agent products help considerably with packaged dictionaries of common industry and application terms, but it’s on you to provide the vocabulary specific to your business and your products.

Virtual agent suppliers also provide consulting services to help the NLP specification. These services are essential for a successful virtual agent deployment. These services are also essential for ongoing management of your virtual agent. Remember that customers’ questions are always changing. So is your business.

IntelliResponse VA uses machine learning technology for its analysis and matching. Machine learning is an algorithmic approach. The algorithm learns by training it with sample data in a controlled environment. It applies its learning when it goes live. The sample data that you provide to train an IntelliResponse VA virtual agent are the typical questions that you want it to answer, not the words and phrases in those questions, not various forms of those words, not the relationships between them, just the questions. IntelliResponse VA can do the rest of the work, even to accommodate the ongoing changes in customers’ questions and in your business. IntelliResponse VA virtual agent deployment is easier and faster than NLP-based deployments and delivers answers with the same level of accuracy. Read our report for the details and note that IntelliResponse can also provide those consulting services to help you deploy and manage virtual agents. The details of the work will be a bit different and there will be less work to do, but the objective will be the same—deploying virtual agents that answer customers’ questions.

 

 

NLP in Social Monitoring, Analysis, and Interaction

Natural Language Processing (NLP) technology, frequently called text analytics technology, is key to analyzing what customers are saying about your company. So leveraging NLP in a big part of the best social-service products. NLP input is the content of customers’ posts, messages, and feedback, while NLP output is a tree structure that represents and contains their syntax, semantics, context, and intent. NLP processing performs tasks such as:

• Corrects spelling
• Parses content to determine parts of speech and their relationships
• Extracts entities and facts
• Resolves vague pronoun antecedents (anaphora)
• Determines the meaning of unknown words using their morphological attributes
• Identifies phrases
• Identifies relationships between words and phrases

This week’s report is about Clarabridge Analyze, Clarabridge Collaborate, and Clarabridge Engage—the Voice of the Customer/ social-service offering from Clarabridge, Inc., a privately-held software supplier based in Reston, VA. Clarabridge Analyze listens, analyzes, reports, and alerts on customer conversations on social and on internal channels. Analyze’s alerts are sent to Collaborate for their assignment and management. From within Collaborate, facilities of Engage let agents respond to and interact with customers.

NLP is a core analysis component of Clarabridge Analyze and the key IP of Clarabridge, Inc. Clarabridge did not supply the details of the functionality of its NLP, certainly not its internals and not very much about its externals.

Clarabridge characterizes its NLP as “proprietary.” While the company owns a few patents, none of them is for its NLP. It protects this IP through minimal disclosure. Other VoC and social monitoring, analysis, and interaction products that we’ve evaluated have taken the steps to patent their NLP technology or have been willing to discuss the details of the technologies that they’ve used to analyze customer conversations. Patents protect the technology from competitors who might copy it but, at the same time, patents reveal the technology to those, like us, who evaluate the products that contain it and those who purchase and use the products. This revealing enables product comparisons and helps ensure that selection decisions address requirements.  (For evaluations of other text analytics-based social monitoring, analysis, and interaction products, we really have read the patents and the patent applications.)

On one hand, understanding what a product does and how it does it are critical to actionable evaluations. For VoC and social monitoring, analysis, and interaction products these are important factors for a product’s performance, throughput and scalability, accuracy, and consistency. Without detailed information on NLP, you’ll be buying a black box. That can be risky. Selection will rely on demonstrations, limited trials, and references.

On the other hand, we’re not language scientists. Our evaluations do not consider the internals of the algorithms that an NLP implementation uses for parsing customer verbatims or for extracting entities, facts, and relationships from them. But, we sure want to know that these products have facilities for automating the analysis of the huge and ever increasing volumes of customer conversations, for performing these analyses quickly and consistently, and for identifying which verbatims need follow-up actions. Clarabridge Analyze does have these facilities. Along with Clarabridge Engage, Clarabridge Analyze can help businesses deliver effective social-service.