Data Enrichment Services
Structured Data > Entity Matching | Relevancy and Ranking
Natural Language Processing > Conversational AI Response Validation | Intent Extraction | Named Entity Recognition | Sentiment Analysis
Computer Vision > Image Annotation | Video Annotation | Semantic Segmentation
Quality of the data collected and prepared for machine learning will clearly affect the output of the machine learning models. However, the data needs to be further enriched with features being identified for the AI/ML model. Also, multimedia data like images, audio and video need to be annotated with metadata like bounding boxes, keypoints, etc, before being used for training an AI / ML model. At the core of an AI/ML project, the following three key phases determine the quality of the data and the performance of the model.
1. Preparing and enriching the training data
2. Validating the model performance in various deployment scenarios
3. Handling exceptions that the model has low confidence on All the above are not typical responsibilities of a Data science team and are better handled by a partner like Nextwealth. We are geared to provide Data enrichment services of highest quality for various different data formats.
Performance of entity matching algorithms is key to businesses and have a direct role is product pricing in eCommerce. It takes manual effort from experienced domain experts to evaluate the outcome of entity matching algorithms. We have extensive retail and ecommerce domain knowledge to conduct this evaluation with a high level of accuracy.
Relevancy and Ranking
Number of visits to their site is a very important measure of customer engagement in ecommerce business. However once a person visits the site, the discoverability of the product he/she is looking for becomes very important. This is a direct outcome of the performance of the internal search engine of the ecommerce site. It is important to evaluate the relevancy of the search output and rank the search results based on the order of relevance.
Natural Language Processing
Human Language Technologies today are hugely impacting the way we use computers by allowing us to interact using natural language. The algorithms need to find patterns in the language and make inferences. This can only be done if they receive annotated training data that identify and indicate different elements of the language.
Conversational AI Response Validation
Though Conversational AI solutions, also known as chatbots, have been around for a long time, it is very important to use humans to validate the response. Validation of the response needs a good understanding of the domain and specific needs of customers. At Nextwealth, we have delivered Conversational AI validation for various vertical solutions like Banking, Insurance, Healthcare and many more.
The key to the success of a Conversational AI solution is to identify the intent of the customer and extract relevant information from the intent. The Intent needs to be classified as to whether they are casual, business specific or others. Quite often our customers need the Intents to be mapped to specific sections of the FAQ from which the answers will be provided.
Named Entity Recognition
Recognizing Named Entities and understanding the type of the named entity like a place, name or product, is a very important need for a successful implementation of an NLP solution.
Identifying the Sentiment of a customer based on the conversation and taking appropriate action (like handing it to a human) is key to customer satisfaction in Conversational AI.
Traditional Computer vision algorithms are about detecting patterns, matching and understanding them to identify objects and take action. But most of the Computer Vision algorithms have moved to Convolutional Neural Network based Deep learning algorithms, which means that there is an increased dependence on Labelled training data. At Nextwealth we understand the challenges involved in handling large amount of data to provide quality annotations at high levels of productivity.
Most common applications of supervised Deep learning applications are Image classification, object detection, action detection and recognition. We use various tools to provide labelling and annotation services to address the training data needs of the above applications.
• Bounding Boxes – Object detection and localization
• Keypoint annotation – Facial recognition, Action detection and recognition
• Polyline annotation – Lane labelling
• Polygon annotation – Object detection and localization
Label Annotation on video segments involves the additional task of tracking the marked objects across the frames. This involves handling large volumes of data and managing a complex workflow with different levels of QA. We work with specialized Video annotation tools to provide these services. These annotations are extensively used to prepare training data for Autonomous drive prediction projects, where objects like cars, pedestrians, traffic lights, traffic signs and others like cyclists are marked and tracked.
The autonomous drive projects also involve annotating with 3D bounding boxes to provide a view of the depth of the object. Many projects use LiDAR or RADAR point clouds for training the models and will need annotations in the point cloud.
Semantic segmentation involves associating each pixel on an image to an object. For the annotation team, it involves marking the objects at a pixel level accuracy on the image and labelling them. Besides Autonomous driving, Semantic segmentation is used in Medical image analysis.