"Words" to "sensors"
2020/4/10
Introduction to Discovery (7/7): Questions to ask in order to determine the vendor's ability to respond to discovery
2020/4/24
As we have seen, there are various difficulties in analyzing "customer feedback".So how does FRONTEO's artificial intelligence "KIBIT" clear the difficulty and analyze it?Let's take a look at its features again.
・ Specialized in text data.Light processing, no need for large-scale computational resources
→ It can be used on a general PC, and the cost and time required for installation can be reduced.
・ Can be analyzed with a small amount of teacher data
→ No preparatory work is required on the customer side to classify hundreds to tens of thousands of data
・ Reproduce human judgment criteria with original machine learning algorithms
→ The judgment axis can be made constant, and evaluation without blurring is possible.
・ Extract and learn features by looking not only at "words" but also at "whole sentences"
→ Absorbs ambiguity, fluctuations, and misconversions of words
In this way, even qualitative data such as "words" that are difficult to make objective judgments can be used as "sensors" to discover the behaviors, emotions, and reactions that you want to find. The accumulated past complaint data plays a role as a sensor when detecting complaints from "customer's voice".The image of the analysis is shown in Figure 1.
When finding complaints from "customer feedback", the machine learning teacher data given to KIBIT is, for example, dozens of "want to find" and "important" complaint data that actually existed in the past. One data can be a short sentence of about 1 to 100 characters ("""Use words as sensors" Find value from a huge amount of "customer voice" and change businessSee Figure 3).At the same time, let KIBIT learn dozens or a little more data of "customer's voice" that is not a complaint, "you do not have to find it" and "it is not important".Then, prepare the "customer's voice" such as inquiry record, questionnaire, voice record, etc. to be analyzed, and you are ready to go.
KIBIT learns the connection between words and sentences from the given teacher data, determines whether it is similar to the sentence of the complaint or not to the general sentence that is not the complaint, and performs scoring (scoring). I will continue.
At the same time as looking at the keywords that appear in the complaint when learning, we also look at the words around the keywords, so when judging new dataEven if you don't have the same keyword, you can score based on the similarity of the surrounding words.. The unique algorithm used in KIBIT assigns high weights to words that are determined to be "want to find" and "important", and low weights to words that "do not need to be found" and "not important".In this way, in the group of data to be analyzed, it is possible to sort by the difference in points between those with high correlations and similarities and those with low scores.
Even if you don't check all the reports and customer feedback that come up every day, you can reduce the work load and man-hours by checking in descending order of score, and you can quickly extract the information that users need. I can do it.This series of processes is the reason why KIBIT is resistant to ambiguity and fluctuation of words and can be analyzed as if reading the context.
Figure 1. KIBIT pre-learning and analysis image
How effective (in actual business) can you achieve by using KIBIT to analyze "customer feedback"?Let's take a look at the case studies, proof-of-concept experiments, and PoC (proof of concept) cases so far. (Figure 2)
Figure 2. Various cases using KIBIT
In addition to the cases we have seen so far, "customer voices" are not limited to those sent to stores and call centers, but also communication using chatbots, writing reviews on social media through SNS, smartphone apps, etc. Communication data that is useful to companies, such as communication using voice recognition such as smart speakers, is steadily increasing.From these data, for example, by correctly grasping complaints, it can be used as a "sensor" that can detect problems with our products and services at an early stage.
This time, we have given an example of "customer's voice", but in the same way, "natural language" sentences that express people's feelings, impressions, feelings, and events are, for example, daily business reports, surveys, personnel interviews, etc. It appears in many daily tasks such as various records of nursing and long-term care.By analyzing qualitative data, which was previously difficult to handle, with artificial intelligence KIBIT, companies can use it to discover various opportunities and issues.Please think about a mechanism that allows you to discover risks and opportunities by mastering "natural language x AI".
Customer feedback and reactions include the following:
・ Inbound call (power reception) from the customer at the call center / contact center, outbound call (call) to the customer
・ Questionnaire at the store
・ Voices and emails from customers to store clerks and employees
・ Review on EC site etc.
・ Inquiries and writing to corporate websites
・ Writing on SNS etc.
Quantitative data is what can be expressed as "numerical values", and qualitative data is what cannot be expressed.
In the case of quantitative data, the state can be judged by whether it is above or below a certain value, and changes can be seen objectively, but qualitative data is difficult to handle because the judgment and interpretation change depending on the subjectivity of the person. I will.
FRONTEO's AI engine KIBIT can score words that are representative of qualitative data on a certain scale and evaluate them quantitatively.
