Of all the fields in the chatbot-crazed world, customer service is one of the prime targets for automation. virtual customer agents (customer service-focused bots, or VCAs) are intelligent systems able to understand what users ask via chat and to provide them with adequate answers. In the context of this article, when we talk about VCAs we mean systems that are able to understand natural language and texting and do not just operate in a rules-based multiple-choice environment. In short, these VCAs compete directly with humans to resolve customer service issues.
The current reality of chatbots nicely counterbalances all the hype that AI is getting and offers guidance as to where development is needed. Here are the key learnings we’ve gleaned from deploying VCAs that autonomously answer questions and having attended major customer service automation and chatbot summits:
1. Good VCAs need customers’ chat log history
Ideally, you can train the VCA with thousands of questions (complete with misspellings, grammatical errors, and pidgin dialects) from actual users of the product/service. But the reality is that most companies do not have existing chat history data readily available for training. In that case, the options are to artificially generate thousands of different questions or to deal with the reality of not having much input data and hope to gather it when the VCA goes live. Neither solution is ideal, and even if companies have a chat log history, it is generally unlabeled. This means the questions in the chat logs are not paired with intents. Fully manual pairing of thousands of questions to intents is time-consuming. A solution that we have developed involves semi-autonomous question-intent pairing tools that considerably decrease the human effort needed to label data. Such an approach makes working with the customer data more efficient and reduces the labeling bottleneck.
2. There is no one-size-fits-all algorithm for understanding user intent
With all the advances in machine and deep learning, most algorithms rely on largely pattern-based approaches to extract intent from a large corpus of previously seen chat history. Users’ questions to banks differ from questions asked to telecom companies — and there is no off-the-shelf algorithm to fit both cases. An optimal solution is to use a host of different algorithms (SVMs (support vector machines), Naive Bayes, LSTMs (long short-term memory), and feedforward neural networks) to match user questions to specific intents. An ensemble of predictors yields a confidence score for each intent, and you can then take the best match. Such an approach provides users with more accurate answers.
3. Moving beyond current machine learning approaches
Extraction of meaning — or more specifically, semantic relations between words in free text — is a complex task. The complexity is mostly due to the rich web of relations between the conceptual entities the words represent.
For example, a simple sentence like “my older brother rides the bike” contains a lot of semantic richness, as the hidden baggage is not evident from the tokenized surface representation (e.g. “my brother is a human,” “the bike is not a living entity,” “my brother and I likely have the same mother/father,” “I am younger than my brother,” and “the bike cannot ride my brother”).
Shared collectively, this knowledge makes communication with others possible. Without it, there is no consistent interpretation and no mutual understanding. When reading a piece of text, you’re not just looking at the symbols but actually mapping them to your own conceptual representation of the world. It is this mapping that makes the text meaningful. A sentence will be considered nonsensical if mismatches are found during the mapping.
Since the computers manufactured today do not include a model of the world as part of their operating system, they are also largely clueless when fed unstructured data, such as free text. The way a computer sees it, a sentence is just a sequence of symbols with no apparent relations other than ordering in the sentence. As the problems related to financial services can be rather specific, you have to augment the typical pipeline of NLP and machine learning with semantic enrichment of inputs. You must devise semantic ontologies that are helpful for the identification of users’ problems in the financial and telecom sectors. The underlying idea of semantic ontologies is to encode commonalities between concepts (e.g. “cats” and “dogs” are both “pets”) as additional information yielding a denser representation of tokens. Another step forward is architecture capable of semantic tagging of both known and unknown tokens, based on the context.
4. Humans are still needed in customer service
VCAs must understand the bulk of cases in which users ask a question in natural language. The VCA should be able to understand the problem and help the user resolve the problem without involving human support. For narrow and only rules-based VCAs, the resolve rates can be higher, but in our experience people are impatient when dealing with customer service. Instead of reading instant articles and suggested topics, they wish to express their problem as a specific question, and they expect a relevant answer. Understanding free text is a tough problem, and current autonomous resolve rates that hover around 10-20 percent reflect that. Even so, when considering that larger companies need hundreds of people to solve highly repetitive issues for their customers, automating even that percentage can save a lot of working hours and allow humans to focus on the more creative and demanding aspects of their work.
Indrek Vainu is the CEO and co-founder of AlphaBlues, a company automating enterprise customer service chat with artificial intelligence.
How startups are scaling communication: The pandemic is making startups take a close look at ramping up their communication solutions. Learn how