How AI Chatbot Answers

Chatbots work by linking "applications" and systems called "bots" via APIs, and returning text or voice responses to users' questions. APIs are interfaces that connect different software and services for smooth processing.

AI In a chatbot, the workflow is as follows:

1. Receive user input in a bot application
2. Speech and language analysis
3. Keyword extraction
4. AI analyzes relevant issues and understands questions based on a database
5. Logic outputs optimal solution
6. Return to bot application
7. Answer is returned in text display or voice

Unlike scenario type where answers to questions are set in advance, AI chatbot analogizes questions from keywords and derives the optimal solution by logic-based processing.

1) What is natural language processing?

AI chatbot's In the response, natural language processing is used to understand the human "spoken language. Natural language" is the language used in everyday speech, both spoken and written. Its counterpart is programming language.

Natural Language Processing (NLP) is the process of analyzing natural language. It is processed by a machine to extract the content of a question from the input language.

Natural language processing is performed by the following steps

▶ Morphological analysis: Breaking down into parts of speech (the smallest units in a sentence) such as verbs, nouns, adjectives, etc.

▶ Syntactic analysis: Classify into subject, predicate, and object

▶ Semantic analysis: Determine the relationship between each word based on syntactic analysis

▶ Context analysis: Conduct morphological and semantic analysis on the surrounding text to determine relationships

Causes of the "Shaking Problem" in Japanese

Natural language, or the language that humans use on a daily basis, is inherently rich in diversity in expression. When humans speak or write, they naturally select expressions and freely combine them to express what they want to convey. Diversity is particularly evident in Japanese, and this tendency is greatly reflected in the notation in chat rooms. This is why even the same question can be written differently by different people.

The following are some examples of shaky Japanese.

▶Differences due to the variety of character types (hiragana, katakana, and kanji): "kirei/kirei/klei," "otafuku/otafuku/otafuku," "ari/ari/ant," etc.

▶ Differences by Kanji forms: "Saito, Saito, Saito", "Gaku, Gaku", "Hen, Bye", etc.

▶ Sending kana: "Reception, Accept, Accept", "Discount, Discount", etc.

▶ Foreign words: "Printer, Printer", "Data, Data", "Server, Server", etc.

▶ Abbreviations: "Instruction manual and Instruction manual", "Vending machine and Vending machine etc.

▶ Notation shaking: "empty can, empty can, empty can", "one month, one month, one month", etc.

▶ Different notation for each user: "customer, customer, user", "document, document", etc.

How to improve the accuracy of chatbots

Points for improving accuracy include the following

▶ Constant PDCA cycle

In order to respond to the above characteristics of the Japanese language and give highly accurate answers, it is necessary to implement the PDCA cycle on a constant basis. We will set appropriate KPI's and establish a system that enables continuous verification of the effectiveness of the system. It is important to monitor the correct answer rate and check user satisfaction with the answers to their questions, while working toward improvement.

It is difficult to achieve perfection all at once in chatbot operation. We will use the correct answer rate before the chatbot is released as a reference and aim to improve it step by step.

With regard to the improvement of the AI part, there are some types that are difficult to tune and PDCA cannot be applied well. It is important to select a chatbot that is easy to handle in order to make repeated improvements while exposing issues.

▶Selecting a chatbot service with high natural language processing capability

Similar to the tuning issue, there is a big difference in processing capability among chatbot services. When selecting a service, be sure to look at examples that are similar to your business, and use their track record as the basis for your decision.

▶ Utilizing a dictionary with a large number of registered words

The number of registered words in the dictionary used by a chatbot determines the breadth of its response. The ease of adding unique dictionaries required for your company's chatbot operations is also a key factor in selecting a service. Selecting a service that is flexible and professionally maintained will allow you to tailor your chatbot operation to your company's needs.

Research on chatbots that are resistant to "shaking" is progressing

Research is underway to overcome the "shaking problem" of notation in order to improve response accuracy. Here we introduce the "Works Tokushima NLP Laboratory for Artificial Intelligence," which develops AI chatbots.

1) About Works Tokushima NLP Laboratory for Artificial Intelligence

Works has released the world's first business application with artificial intelligence, HUE, and began offering it in December 2015.

In February 2017, Works opened the "Works Tokushima Artificial Intelligence NLP Laboratory" in Tokushima Prefecture, Shikoku, Japan. The Tokushima Laboratory is conducting research and development utilizing natural language processing in order to make effective use of operation logs accumulated at HUE and to put AI functions more in line with user needs into practical use.

At the Tokushima Research Institute, R&D is being conducted with a particular focus on Japanese language processing among natural language processing. Major R&D activities include the following

Chatbots for various business scenarios

・Research and development of input-less technology and image processing technology to streamline input work

・R&D of AI functions to improve productivity by utilizing corporate data

(2) Development results at the Tokushima Research Institute

The institute has developed "SudachiDict," Japan's largest scale dictionary for Japanese language processing, which contributes to the realization of comfortable conversations. In the development of "HUE," an ERP system with artificial intelligence, the institute is focusing on usability and expanding its versatility. The Tokushima Research Institute's research and development has greatly advanced the technology in the field of business systems, which had been thinly layered, and in October 2020, SudachiDict was released on Open Data on AWS. Furthermore, in February 2021, the HUE Works Suite DX Solutions Chatbot (HUE Chatbot), an AI dialogue tool, was launched.

SudachiDict embedded in systems and services in various industries

SudachiDict is Japan's largest natural language processing dictionary developed by Tokushima Research Institute.

It contains more than 2.9 million words, maintained by experts, and has three types of separators to suit various usage scenarios. For example, it is designed so that the AI can accurately understand "Works Tokushima NLP Research Institute for Artificial Intelligence" no matter which word is used to separate the words.

SudachiDict is normalized for character types, kana, misspellings, and proper nouns, and can handle differences in notation among users and the fluctuations and ambiguity inherent in Japanese with high SudachiDict is able to handle differences in spelling among users, as well as fluctuations and ambiguities unique to the Japanese language. In addition, synonym information detailing synonymous relations for approximately 60,000 words is provided. This information can be used for various purposes, including full-text search.

M3 Group," which is highly reliable by continuously expanding and improving its vocabulary and always providing the latest dictionary, and "Recruit Holdings," which provides diverse services for a wide range of needs and domains in the medical industry and develops a wide range of businesses such as recruiting and business succession, GMO Internet, Inc.", which operates advertising and financial businesses via the Internet, and others have also adopted the Japanese language.

Selecting a chatbot that can respond to the diversity of the Japanese language

Japanese is known as one of the most expressive languages in the world. However, this diversity can sometimes be a barrier. Chatbot responses require machines to correctly understand the free expressions of humans. SudachiDict accurately handles such fluctuations in the Japanese language, enabling chatbots to engage in highly satisfying conversations. If you are looking for a chatbot with a high accuracy rate and natural conversation, you should definitely consider using it.

HUE Chatbot" is equipped with SudachiDict, developed by the Works Tokushima NLP Research Institute for Artificial Intelligence. Not only is it resistant to fluctuations and ambiguities in the Japanese language, but it is also easy to build a PDCA cycle, allowing for modifications to be made by a few people with no code. Please take a look at our introduction materials.