Since our founding in 2009, CHOWA GIKEN Corporation(CGC) has been developing and utilizing general-purpose AI engines in cooperation with research institutions. By customizing these engines to meet our clients' needs, we are able to build specialized AI solutions that achieve outstanding results. At CHOWA GIKEN Corporation(CGC), we provide our knowledge of AI and our AI engines as a service to support the continuous growth and transformation of your company.
This engine classifies the sentences into appropriate categories based on their content. It extracts features based on N-Gram、Word2Vec、Doc2Vec、Deep learning, etc. and classifies sentences properly.
The engine analyzes the emotions such as positive, negative, nutoral, etc.
The engine understands the key points of a sentence and generates summary sentences. Evaluates the importance of words based on morphological analysis, syntactic analysis and LexRank, and produces summarized sentences.
The engine extracts feature words which represent a sentence with using TFIDF and AIC ratings of word importance.
The engine detects the position of various objects in an image and recognizes the types(classed) of objects such as objects and faces, etc. By combining the YOLO algorithm, R-CNN, SVM and other classifiers, it recognizes objects at high speed.
The engine predicts the future, such as sales and visitor numbers, based on past data. It performs feature extraction and investigation of prior distributions of data to fit client's data, and predicts the data using Bayesian estimation and random forests.
The engine identifies hobbies and thoughts based on past user behavior and recommends events, coupon information, etc. It makes recommendations using memory-based methods that learn feature vectors of each user through collaborative filtering, and model-based methods that build predictive models using algorithms such as Bayesian networks.
The engine searches for the best combination and creates efficient shift proposals and other tasks in spite of constraints. It creates efficient shifts and schedules through optimization algorithms, including meta-heuristics such as genetic algorithms.
We are developing messaging support AI for messaging services between persons to enable smoother conversations by measuring each other's interests and relationship progress based on the history of previous conversations. In an empirical experiment, the use of AI improved the probability of starting a relationship by 47.5%.
【Background Technology】Topic Analysis / Logistic Regression / Deep Learning
While there is a demand for automation of call center to reduce the number of manned operations, this has been impossible to achieve with push-button automation due to the need for identification. Voice recognition + natural language processing + text-to-speech synthesis automate natural voice inquiries. With these technologies, even people with low IT literacy can make the same inquiries as those made by conventional operators. Translated with www.DeepL.com/Translator (free version)
【Background Technology】Speech recognition / Natural language processing (eigen extraction, document proofreading) / Conversation management / Speech synthesis
The system automatically reads the information from the image of the receipt and sorts it into structured data, using OCR technology to deduce the nature of the transaction and the nature of the account. It eliminates the need for manual inputting of information, and can be expected to save labor by improving efficiency and reducing input errors. We are considering implementing an identification engine that can flexibly respond to reduced tax rates. Translated with www.DeepL.com/Translator (free version)
【Background Technology】Bag of Words / Word2Vec / SCDV (Sparse Composite Document Vectors)
Based on natural language processing, we developed a searchable knowledge system that includes similar expressions, rather than simple keyword matching. The search results are displayed as a summary document, improving readability and client business efficiency.
【Background Technology】TF-IDF / Word2Vec / Doc2Vec / Approximate Nearest Neighbor（HNSW、IVFADC） / Summarization （LexRank、Feature based）
Many tabular documents created in system development involve problems such as different categories to be assigned by different people, and a large workload in assigning categories. By applying our classification engine, which is based on TF-IDF, fasttext, SCDV, SVM, random forest and other technologies, we developed an AI that automatically assigns categories to documents with the same level of accuracy as human work. Translated with www.DeepL.com/Translator (free version)
【Background Technology】TF-IDF / Word2Vec / Doc2Vec / fasttext / SCDV / BERT / SVM / Random Forest
We have developed a system that uses chatbots to investigate failures based on data collected from CRM and issue management systems that were not being used effectively. Chatbots conduct interviews on behalf of humans to find out the details of failures, and then inquire about the cases, and investigate the failures from a large number of cases. This has made it possible to save on labor for operators and reduce the time it takes to respond to inquiries.
【Background Technology】TF-IDF / Word2Vec / Doc2Vec / SCDV / Conversation Management
Using Semantec's segmentation technology, we have developed a Bone Area Image Recognition AI for high precision robotic fresh food processing. While conventional image recognition has not been able to achieve a practical recognition accuracy, we applied deep learning to achieve a sufficiently high accuracy recognition in three months of research and development. Finally, we created an AI module to be incorporated into the robot.
【Background Technology】CNN / A hybrid of semantic segmentation and old-fashioned image processing
In order to automate the process of appraisal and purchase of branded goods, we researched and developed an image recognition AI that uses CNN (Convolutional Neural Network) to identify the model number of the designer's branded goods. We developed a light-weight image recognition engine that can be implemented in a mobile app, and provided it to a real service.
【Background Technology】CNN / Hierarchical Learning Methods / Object Recognition / Text Extraction / Inter-image Similarity Extraction
Using CNN (Convolutional Neural Network), we have researched and developed an image recognition AI to exclude defective scallops that are contained in a certain percentage of scallops. Since the target organism is a living organism, there are some issues including fluctuations in features such as size and shape, and lack of stability in the subject's posture when the image is taken, but high accuracy was achieved by pre-processing the image and devising a learning method.
【Background Technology】CNN / Semantic Segmentation / Fault Detection
In order to automate the classification of dry cell batteries, we developed an image recognition AI using CNN (Convolutional Neural Network). Since the image recognition was targeted for manual sorting at industrial waste disposal sites, it was sometimes difficult to recognize images due to blurred characters and blurred vision of the objects, however, we improved the accuracy by improving the training data and showed the possibility of automation.
【Background Technology】CNN / Object Detection / Character Recognition
The engine detects cracks in the wall surface and measures the length and width, if possible. As a prelude to a manual onsite inspection, we can take video footage and cut out the images.
【Background Technology】CNN / Semantic Segmentation
This is an image generator that converts a photo of a human face into an Ukiyoe image. Although conventional algorithms cannot extract the features of human faces from Ukiyoe, our approach method of learning human faces and Ukiyoe enables us to convert a photo of a human face into an Ukiyoe image with features.
【Background Technology】Generative Framework（GAN、CycleGAN etc.）
We have developed an AI engine that makes recommendations for recipes based on "mood" and "situation". By using a neural network to learn the results of 40,000 food preference surveys, we have built an algorithm that recommends recipes that match the user's current mood and situation.
【Background Technology】Deep Learning
The engine predicts future demand based on factors such as weather, day of the week, time of day, etc., and then calculates and presents a price that maximizes revenue.
It uses neural networks to predict the amount of demand. It also learns and utilizes the relationship between how much demand there is and what price can be set to generate revenue in advance.
We have developed an AI to predict sales, number of customers purchased, etc. by day and time of day for use in various procurement and employee shift management. It uses Deep Learning and ensemble learning. The combination effect of the methods has resulted in 95% accuracy in forecasting and more precise planning. Combined with order quantities to estimate shortage and waste volumes.
【Background Technology】Deep Learning / Gradient Boosting / Random Forest / Bayesian estimation
This is a fault detection system that uses plant sensor systems. An obstacle that could not be correctly detected until several hours after the equipment stops has been developed, but now an AI has been developed that can detect the obstacle during or immediately after startup and shutdown using machine learning. As a result, the time required for fault detection has been significantly reduced.
【Background Technology】Anomaly detection / change point detection
In the public transportation industry, where there are strict legal constraints on driving time, distance traveled, and other factors, creating shifts takes a lot of manpower. Meta-heuristics has enabled us to semi-automate the creation of shifts that meet a number of constraints.
【Background Technology】Solving the problem of optimizing work assignments that satisfy shift creation rules with a genetic algorithm (GA)
We have developed an AI that automatically generates meal menus for special customers, such as the elderly. By learning internal variables from feedback from dietitians as well as quantitative metrics such as nutritional value and food cost, the system is able to optimize the menu, including qualitative metrics such as customer satisfaction.
【Background Technology】Combination Optimization
Until then, garbage collection route planning has been done manually, which is very costly and difficult to pursue with precision. We have developed an AI that automatically creates an optimal route plan from map data and garbage collection volume forecasts. Simulations have shown that the use of AI has the potential to reduce labor costs, fuel costs and CO2 emissions by about 20%.
【Background Technology】Uses the Guided Local Search (GLS) method to find the lowest cost path while avoiding the local minimum solution
The route of an AGV (Automated Guided Vehicle), which moves through a warehouse to pick up cargo, is a difficult problem that requires an explosion of calculations depending on the amount of cargo to be shipped. To solve this problem, we developed an AI that formulates the restrictions on the shape of the cargo on the pallet and uses local search methods to find a solution that can be used in actual AGV operations.
【Background Technology】Mathematical Optimization / VRP / Metaheuristics
When the amount of energy flowing into a generator using natural energy exceeds a threshold value, the generator is shut down to prevent it from malfunctioning, resulting in power generation loss. Using deep reinforcement learning, we have developed an AI that automatically learns appropriate control to reduce power generation losses in a simulation.
【Background Technology】Simulations / Deep Q-network（DQN） / Rainbow
（In Japanese alphabetical order）