Natural Language Processing (NLP) - Strategy for Enterprise Architecture
Real time data processing and analytics help businesses become more Agile. To address market demands, data is an essential part and lifeline for the businesses to run efficiently. Manual processing of the data points and mining legacy data models lead to a retrogressed decision process. Data at rest has a time bound depreciation. Business strategies are deeply integrated to predictive models and analytics. Automation of data structuring, data processing coupled with artificial intelligence help build a rich echo system along with the direct customer interaction into enterprise platform.
Voice is the purest form of data that an enterprise can leverage to build a happy customer base. Voice provides the data that is rich with context along with the human emotions. In order to handle the Voice and Chat communication channels with the customers; Enterprise Architecture design needs to make a measured and balanced decision to go with one or more NLP engines that solve all the channel customer integrations (IVR, Chatbot, SMS, Voice Messages etc).
NLP as a medium for Customer Communications:-
Natural Language Understanding, Processing and Response are the 3 different features of the language automation in Channels Architecture. Below is a graphical representation of channel capabilities associated with NLU, NLP, and NLG.
Customer Communication Channels Capabilities
1. IVR with Natural Language understanding
2. Intelligent Response CHAT
3. Voice Assistants
4. Natural Language based Search
IVR uses (NLU) of the engine to understand the voice of the customer and to act on it rather than relying on the touch pad for the choices or options.
- NLU provides an extension to the customer input criteria (e.g. capturing an address without typing.)
- An NLU-based system allows for easier voice bio-metric authentication by comparing the voice print captured during one-time setup.
- IVR’s technology allows for a natural conversation by providing more human-like interactions using advanced dialog strategies.
Intelligent Response uses (NLU, NLP, NLG) to communicate with customers and improve on its interactions.
1. Text to Intelligent Text & Voice to Text conversion for the applications consumption response
2. The task automation via an intelligent response system alleviates live agent workload and associated costs with scaling, training, and management.
3. Average call time with an intelligent response system is lower than that with a Live Agent. Reduce the average call time with a Live Agent to handle it offline via an intelligent response system
4. Customer self-service is expanded as solutions are available 24/7 via the response system.
5. Intelligent response systems can fit into an omni-channel experience by integrating with multiple channels such as sms, mobile, chat and IVR.
6. The automation that’s delivered by the intelligent response system can improve and fill the gaps in the knowledge base of the Live Agents.
7. The intelligent response system drives revenue growth and cost optimization by online interactive conversation.
8. Visualize and monitor real-time information flow on the chat system to handle and scale the systems based on needs
9. Leverage cloud-based systems NLP engines to address reliability and scalability requirements.
Technology Stack:
Natural Language Processing considers linear relations between words; in simple cases of understanding, we can design feature extractors and select the entities to match for the given intent. The complex part of large volume tedious feature extraction involves neural networks and often more accurate than the simple feature extraction. The usage of multi-layered neural networks for machine learning is called deep learning. The collection of the algorithms implemented under deep learning has close similarities with the relationship between stimuli and neurons in human brain.
Natural Language automation consists of :-
Understanding: NLU (Natural Language Understanding)
Decisioning: NLP (Natural Language Processing)
Response: NLG (Natural Language Generation)
Voice detection and text understanding uses different algorithms.
An NLP system has a lexicon (vocabulary) and a set of grammar rules coded into the system. Natural Language Processing uses machines to run deep learning algorithms to understand users’ text communications and to intelligently respond to them based on the flow configurations. Intelligent Response applications are built with these NLP systems to automate Agent communications and to substitute for human beings, thereby handling costs associated with Agent scalability.
NLU & NLP are considered to be a part of the foundation fabric of the customer relationship management architecture. Natural Language Processing systems provide text-based conversation and other features like below :-
1) Speech to Text conversion facilitating Voice interface for Chat based Intelligent Response systems
2) Content Classification, Categorization, and Filtering across a large volume of content generated from structured data
3) Speech Recognition which leads to voice authentication
4) Sentiment Analysis of the conversation to identify the customer mood and patterns to figure out the correct response
5) Telephony Gateway to recognize IVR based integration into Intelligent Response systems
6) Knowledge Connectors to connect the enterprise API for product Knowledge Systems, connectivity and resolution to correct FAQ and respond with support articles
7) Topic modeling, a technique from NLP, to extract unique topics from a group of documents
Challenges: How to choose an NLP engine provider?
A sample decision tree helping to make the comparison between an ‘Agnostic Bot Provider’ vs Cloud Provider algorithm.
a. Choosing an Agnostic NLP interface can add complexity for channels various channels or specifically for the ‘Intelligent Response’ channel adds more complexity
b. Along with maintenance of the Agnostic Bot Controller, considerable amount of custom development work is needed to configure and maintain the NLP engine.
c. NLP engines are supposed to get better as time passes as the content that it hears, or listens work as a feedback loop to perform better in Deep Learning algorithm-based technologies. This helps the AI for better calibration and helps with higher accuracy of the response.
d. When we use Bot Controller to route traffic, we are creating artificial layers of data that gets fed into the AI engine. The disparity on the data gets larger over time and any given point of time, a single AI engine will not have overall context information and misses the overall theme of customer interaction.
e. The same customer interacting with a different channel using a different NLP engine will be isolated and treated differently with the response outcome. Even though this is a subtle change and not noticeable, we are injecting this kind of behavior into our implementations. Also, the response tone might not be consistent as the bot controller does not guarantee(unless we design rightly to avoid this complexity) the same engine for a given customer.
Solution: An Agnostic Bot Controller should only be chosen if the ‘Intelligent Response’ system that needs to be built is ahead of a cloud provider decision. Otherwise make the decision to stick with a single cloud provider; and feeding enough data to make the algorithm right is the best approach.
Note: The efficiencies of the cloud provider getting better based on new performance upgrades. The underlying implementation and its associated algorithms remains almost identical with all the cloud service providers.
Strategy for Leadership:
- NLP engines are the new class of software that all top cloud players (Google, Microsoft, Amazon) have been heavily investing into to build an echo system around the language processing engines. Speech Recognition, Text processing, Chat to Service automation, IVR, Voice authentication, etc. are all becoming the core applications in enterprise customer services. Using an exclusive NLP engine might be not optimal with the large cloud providers as the eco system supporting the NLP engine usually provides a complementary and complete solution.
- The cloud providers have all the same algorithms using Deep Learning technologies of Machine Learning. Some pure players, such as Nuance Communications, are heavily invested into voice recognition, and have been collaborating with big companies like Apple in their Siri voice engine and thus developed a deep core competency. Big box system providers provide better service and usage in enterprise applications in terms of an optimal package consisting of natural language understanding, text processing, dialog flow, sentiment analysis etc. in comparison to niche vendors.
- Concerns such as Vendor Locking for NLP systems can be de-emphasized as there is nothing exclusive or patented on the algorithm between various vendors on the implementation of ‘Deep Learning Neural Networks’ for language handling and processing. The only differentiator is the Lexicon (Amount of training data that is available with the vendor and the vocabulary configured in their systems in various languages). The Lexicon comes in handy when there is a large context from which your company is trying to provide answers to their customers, but in the case of ‘Intelligent Response’ and IVR’s, Lexicon has a very limited role as the context of problem space is narrow. Questions that will been asked solved via IVR are through a pre-configured set. The Lexicon plays a stronger role when implanting something like Voice Search as a functionality in our channels.
For the past 5 five years, the NLP engine and its functionality has remained relatively the same since AI technology has moved into general enterprise usage.
- Distinction between NLP vendors when using in Speech vs Chat is necessary as needs are different in cases like IVR vs Intelligent Response (Chat)
- Lexicon requirements need to be evaluated as the language used for IVR or Chat is very specific. The channel requirements for Voice Assistant &Voice Search need to be verified to more specific resolution.
Reference Architecture Stack with Capabilities:-
