"Words" to "sensors"

Customer feedback and customer reactions, such as complaints and gratitude for our products and services, employee behavior, and impressions of our products and stores, are valuable data. It's often said that your voice is a treasure trove, but what about the companies around you?How many companies are making effective use of it? FRONTEO has received inquiries from a considerable number of companies, saying, "There are too many records, but I can't grasp the contents" and "How should I sort them out?"It's easy to imagine that no matter how many "voices" you collect, if you don't classify them correctly and respond correctly, they won't help at all.If the treasure mountain remains a mountain for a long time, it can be used as a "sensor" for product development, improvement of store operations, response to crises, etc., only when the necessary voices reach the necessary departments with just a mass of data. I can do it.

So how can you use KIBIT to turn your "voice" into a "treasure mountain"? (Figure 1)
Figure 1. Issues that KIBIT can solve

As you can see in Figure 1, the most obvious advantage is the high speed processing power.A financial company's call center receives 1 inquiries per day and 7,000 inquiries per month, and one EC site receives 1 to 20 inquiries per day and 1 per month. There are as many posts and posts on SNS.When dealing with complaints at such a company, it is practically impossible to check all the inquiries, opinions and impressions received, and the contents of writing and posting on SNS, and to deal with them because it takes too much effort. is.

However, with "KIBIT", high-speed analysis is possible, so it is possible to analyze and sort thousands of data in a few hours, depending on the amount of sentences per case.In the past, you could only do a "sample check" to check within the range that people can see, but by using KIBIT, you can perform a "all-case check" with improved comprehensiveness, and you can rely on it by chance. You can reduce the uncertainty of leaks and misses.In addition, the records can be sorted in the order in which they are most likely to represent what you want to find, allowing you to check efficiently (Fig. 1).This is because when you let KIBIT learn, you can not only teach "what you want to find" but also "what you do not need to find" at the same time, so that you can reveal the characteristics of the sentence you are looking for and improve the accuracy. , It is a big feature.

Figure 2. KIBIT is fast and can be ordered by scoring.

We receive various "customer feedback" such as inquiries, complaints, gratitude, opinions, and requests.How can we classify them as negative (complaints) or positives (voices of gratitude) and deliver them to the appropriate department?The first thing that comes to mind as the easiest method is to sort by "keyword search", isn't it?If you enter the word you want to find in the data-converted "customer's voice", the sentence containing that word will be hit.However, as shown in (Fig. 3) above, there are cases where the same word is used for both gratitude and complaints.In this case, in the end, it cannot be classified correctly unless a person reviews it.
Figure 3. Even if you use the same word, the content is exactly the opposite

Also, in keyword search, if you try to find words that lead to various complaints and increase the number of keywords, the number will become enormous.And does the word "sweet" have a good meaning?Does that mean bad?There will be both ways to use it.Fluctuations in the selection and usage of these words, ambiguity (used in multiple meanings), and incorrect expressions are important when performing "natural language processing," which processes natural sentences that humans normally use on a computer. It is a key point.Also, be aware of cognitive bias that affects human judgment.When a person records what he or she thinks is a complaint and puts a flag (mark for sorting) on ​​the call to the call center, the judgment as to whether it is a complaint is blurred or wrong. There is a thing (wrong flag), and if you leave it as it is, you may run the risk of not getting the desired result.

For example, suppose an e-commerce site sells out an item you wanted to buy while you made a mistake, and the call center receives a complaint call.The operator may decide that "this is a customer operation error and will not be reported as a complaint".However, for the person in charge of the EC site, it may be that there is a problem with the UI (user interface) of the site that the operation error occurs, and information on why the operation error that makes a complaint occurred may be important. ..In a large call center, the more people involved in the work, the more different people will think about it, so no matter how many manuals you make, the information you want may be overlooked. (Fig. 4)
Figure 4. Leakage and mistakes occur due to subjective and judgmental blurring

Analysis by AI that performs natural language processing can also handle voice data by converting it to text data.Voice data is easier to record than text input, but it has major challenges.It is an erroneous conversion when converting voice data to text data using voice recognition software.The accuracy of speech recognition is improving day by day, but there are various factors that affect the accuracy of speech recognition, such as the gender, age, dialect and habits of the speaker, physical characteristics, recording and communication environment, etc. The current recognition rate of speech recognition is said to be around 90%.So, how much misconversion would occur if the voice was actually converted to text?

Figure 5 shows a financial transaction that takes about two minutes.If all misconversions had to be visually corrected by humans before analysis, the work would take an enormous amount of time, which is not realistic. In the case of KIBIT, it is possible to analyze the erroneous conversion as it is.In one experiment, I compared the case where the erroneous conversion of speech recognition was corrected and the case where the analysis was performed with the erroneous conversion without correction.Then, "The accuracy does not decrease even if the analysis includes conversion errors.I got the result.

KIBIT does not search for words as keywords or compare the meanings of registered words like a dictionary to produce analysis results, but grasps the characteristics from the composition of the entire sentence and discriminates similarities.It can be said that the results of this experiment show the strength of KIBIT, which will be described later.
Figure 5. Example of speech recognition in financial transactions