Log in     Support     Status

5 Reasons Why Advances in NLP Can Enhance a Helpdesk

by | Oct 26, 2020

One of the questions we get asked the most is “What are the advantages of Natural Language Processing (NLP)?” 

Helpdesks literally run on human language—adding organizational structure, keyword tags, workflows, and more to incoming questions and answers. As a result, it stands to reason that major advances in natural language processing can facilitate equivalent progress for the customer or employee experience at a modern helpdesk.

Recent major advances in NLP come in the form of gigantic, pre-trained neural networks like Google’s BERT and OpenAI’s GPT-3. These models are essentially a sprawling lookup table for how words should appear alongside one another. These models are based on probabilities and logic that it catalogs by reading (or “training” itself on) billions and billions of articles, websites, and other data from all kinds of contexts. These pre-trained algorithms are general-purpose tools that excel at many tasks right off the shelf but can be fine-tuned (or transfer-trained) to achieve even better performance and accuracy for specific tasks as needed. You might think of them as an entry-level employee with a general degree like mathematics or psychology, but with enough raw intelligence that they can quickly pick up new skills on the job.

Keep reading to discover five reasons why major advances in NLP can enhance a helpdesk.

1. Answer users’ questions 24/7 with an automated chatbot

Not long ago, keyword recognition and conversational pathing were so brittle that chatbots were not a viable solution. Natural Language Understanding (NLU) has become significantly better, and now powers chatbots that can hold their own in human conversations. One of the clearest benefits of automating the “front desk” for questions in this way is that it is always staffed.  24/7/365. Plus the answers come in seconds, so it’s quick and easy for users. 

Moreover, query accuracy and comprehensiveness are supercharged by taking advantage of the way that these neural networks organize and search for information. All of the billions of words and phrases they see during their training are sorted into a numeric representation. Remember the Dewey Decimal System at the library? The models effectively create their own Dewey Decimal Index so that adjacent concepts from similar documents are indexed close to each other. For example, “How do I reset my password” and “Locked out of computer” will be search neighbors, despite having no words in common. 

2. Graceful escalation to a human when necessary

Even with the best technology, we still know that cases will arise when the front desk will need to escalate to a specialist. We have built our chatbot platform according to this ethos. Depending on the complexity of the use case, we find that the chatbot can automate anywhere from 70% to 100% of inquiries. A simple application like requesting information or a product demo might be 100% automated, but a more nuanced interaction like onboarding a new employee might be 85% automated and pull a human reviewer into the loop 15% of the time.

From the very beginning, we knew that designing a graceful handoff from machine to human would be critical. We detect multiple conditions such as a direct user request, negative sentiment, or a low NLU match score, and we can escalate the conversation either to a real-time live chat with a human or to a ticketing queue for human follow-up later. 

3. Analyze user sentiment 

One of the attributes that a neural network can include in its representation of words and phrases is a prediction of the user’s sentiment. Whether happy or sad, positive or negative, etc. This is based on supervised or labeled training, where a model is fine-tuned on previously human-labeled sentences, questions, or articles. 

Sentiment analysis is useful, not only to escalate issues that might be emotionally explosive as discussed above but to track trends over time. For example, the image below shows a chatbot where significant negative sentiment was appearing, which led to the identification and resolution of an issue, which was fixed and subsequently improved user sentiment. 

4. Integrate with application data

We can greatly extend the usefulness of our Helpdesk Platform by plugging into third-party applications like Salesforce, Workday, SQL databases, Gmail, Outlook, or virtually any other custom integration through our Developer Platform. 

This allows us to perform actions and workflows across platforms to get the job done. For example, a user might ask Capacity, “What is the contact info for Alice at Acme LLC?”  In this case, our NLP is matching to the intent so it can run the correct skill—maybe a Salesforce or Hubspot contact lookup. But that skill uses an API that also calls for two additional variables: a name and a company. Our NLP, therefore, uses entity detection and slot filling algorithms to determine that Alice is the name of the person required for variable 1, and Acme LLC is the name of the Company required for variable 2. We send those along as a request through the API, get the answer back from the CRM system, and return the contact information to the user.

5. Instantly scale to multiple languages

Having the chatbot understand user inquiries in virtually any language is largely a matter of swapping out one neural network for another. 

The way that these neural networks index and store information when reading texts is remarkably universal across all human languages. It turns out that neural net training in any language will tend to index, for example, financial terms are farther away from artistic terms. Or they might cluster verbs together in one part of the dense vector index, and nouns in another.  Researchers at Facebook trained a model called XLM-R on 100+ languages and found that it performed just as good if not better than models trained on a single language.

Understanding user inquiries in multiple languages is a readily solvable issue since this involves search, best-fit matching, and has some room for error; but the remaining limitation is that your answers or responses are typically still created in one language. If you have responses in multiple languages, you can return the appropriate response that matches the user’s inquiry. However, translating on-the-fly from your primary language in a response might introduce meaningful “drift” in what you are answering, particularly if you have very specific terminology for your use case or business, so you may want to receive inquiries in multiple languages, but send all responses in English. 

Putting it all together

These are some examples of the exciting things that you can do with Capacity in the world of Helpdesks now that we have integrated the latest advances in NLP technology. A holistic platform that takes advantage of all of these technologies vs. a patchwork of point solutions makes it that much more powerful, and also facilitates reporting and analytics on each of the trends and use cases.