Today we’re going to discuss how the dialogue classification is structured and why it’s useful for business.

In the B2C sectors such as automotive industry, real estate, banking and medicine companies spend funds to attract leads, usually in the form of calls or manager requests. Marketing expenses share in the annual budget reaches several percent and the efficiency of spending depends on two key factors: the quality of leads provided by the marketing agency and the call center performance. 

How to analyze the call quality?

Listening to call records is the easiest and most efficient way. However, this method is unappropriated if you have hundreds and thousands of calls per day.

In this case, technologies able to process any volume of data are here to help.

On our platform, for instance, an incoming call is routed to an operator. Then the call is recorded and divided into two channels: separately the client's, and the operator's voice. After that, the dialogue is transcribed and transmitted in the text form for further analysis. 

How are the dialogues processed?

In short, based on the received text array, individual terms and their combinations are selected. The robot uses them to draw conclusions about the conversation topics. The term selection process is performed in several steps: 

  1. Compiling the dictionary of terms and word combinations – highlighting biterms, pairs of words used in one message but not necessarily standing next to each other, and trigrams, sequences of words.
  2. Word-sense disambiguation – identifying all possible meanings of the word in the context.
  3. Searching for common words – text analysis in a screen where certain words are traditionally placed together and can be analyzed regardless of their order.

An example of a native dictionary for the car business.

The completed dictionary is loaded into a model, which builds patterns and starts to identify conversation meanings.

The feature is that the dictionary is built for a certain subject. Thus, you can create a finance sector model and customize it for any bank considering its characteristics instead of building a project from scratch.

How to use a classifier in practice?

Let's consider the real estate industry. High property prices result in extremely high cost per lead attraction, from tens to hundreds of dollars. According to one real estate developer, the company spends about $820 000 per month to a marketing agency to generate calls and requests.

An example of a real estate classifier.

Understanding of how efficient budgets are spent, both in terms of the lead quality and the professionalism of call centre operators handling calls, is crucially important. 

Use case №1: lead quality

In order to earn more, marketing agencies often cheat metrics by initiating fake calls. Same people are calling from different numbers about different properties. Sometimes they can even visit apartments but never make a deal.

To identify such cases, you need to compare the numbers from which the calls come, as well as check the caller, or more precisely, – how many calls he made and what objects he was interested in. If these objects belong to the same category, for instance, the area and number of rooms match, then the request is considered relevant. In turn, if the questions are about apartments from different categories, and the CRM system data indicates that these customers have not yet bought anything, then the caller is defined as a spammer.

Several thousands of calls were analyzed using voice biometrics. Records that the system identified as similar were re-checked by experts. As a result, 4,5% of calls turned out to be suspicious, which totalled $37 000 per month.

Use case №2: operator professionalism

Not only marketing agencies try to simplify their work. Call center operators are eager to deal with customers ready to buy an apartment right now. Other calls are marked as untargeted and forgotten in the future. As a result, the company loses hundreds of potential deals.

In the second case, we analyze ‘irrelevant’ calls and try to identify target ones from them. These include conversations where apartment purchase markers were mentioned such as names of objects, budget, number and space of rooms. The willingness to take further action, such as coming to a viewing, calling back or taking out a mortgage, was also identified separately.

As a result, 72% of these records were identified as target ones. The detailed stats on target calls were as follows:

  • 47% – ready for further actions
  • 29% – considered a mortgage
  • 25% – declared their budget
  • 73% – declared the desired number of rooms

Thus, dialogue classifier allows controlling and improving the process of lead generation, evaluating calls by the specified criteria. Moreover, the technology allows to find calls in the CRM system using keywords and set up the triggering action.