NLU | Skelter Labs
NLU(Natural Language Understanding)
Deep Learning-based Natural
Language Understanding
Skelter Labs’ natural language understanding (NLU) technology uses machine learning to understand the structure and meaning of text, and extracts and classifies the information it contains to detect the context of a conversation. Through this, answers to user questions can be found, or the content can be categorized according to the item and used for creating a virtual assistant.
NLU(Natural Language Understanding)
Deep Learning-based Natural Language Understanding
Skelter Labs’ natural language understanding (NLU) technology uses machine learning to understand the structure and meaning of text, and extracts and classifies the information it contains to detect the context of a conversation. Through this, answers to user questions can be found, or the content can be categorized according to the item and used for creating a virtual assistant.

Machine Reading Comprehension
MRC is a question-answering (QA) technology, where AI analyzes problems independently and finds optimized answers. It analyzes the document used as text, deeply grasps the meaning of the words in the question and extracts a phrase that could be the answer with high precision and recall.
In particular, when fine-tuning is performed on a general-purpose pretrained language model for Korean, it utilizes the latest algorithms, including BERT and its own algorithm, to create a language model with strong performance and accuracy. As a result, Skelter Labs’ MRC engine ranked first in both fields—KorQuAD 1.0 and 2.0—proving that the engine’s reading comprehension has surpassed human reading comprehension and delivers superior performance relative to its competitors.
NLU
Machine Reading Comprehension
MRC is a question-answering (QA) technology, where AI analyzes problems independently and finds optimized answers. It analyzes the document used as text, deeply grasps the meaning of the words in the question and extracts a phrase that could be the answer with high precision and recall.
In particular, when fine-tuning is performed on a general-purpose pretrained language model for Korean, it utilizes the latest algorithms, including BERT and its own algorithm, to create a language model with strong performance and accuracy. As a result, Skelter Labs’ MRC engine ranked first in both fields—KorQuAD 1.0 and 2.0—proving that the engine’s reading comprehension has surpassed human reading comprehension and delivers superior performance relative to its competitors.
NLU
Intent Classification
Intent classification is one of the most essential technologies for managing bot conversations. Skelter Labs’ intent classification engine uses the context before and after words and sentences to grasp the intent of the customer’s speech and respond smoothly.
Skelter Labs’ outstanding intent recognition rate is attributed to its proprietary model that applies a rule-based method and a deep learning-based method in a hybrid form. The rule-based method recognizes intents based on sentence patterns learned from a large volume of natural language sentences. On the other hand, the deep learning-based method models a pattern of mutations that are difficult for the rule-based engine to process, allowing rapid learning with a small quantity of sample sentences. These two methods have been applied in a hybrid form for optimization. SkERT, a proprietary model of Skelter Labs based on the latest models such as SpanBERT, has recently been developed and applied. As a result, we have achieved remarkable growth in the F1 Score, which indicates the weighted average of precision and recall, a significant demonstration of accuracy.

Intent Classification
Intent classification is one of the most essential technologies for managing bot conversations. Skelter Labs’ intent classification engine uses the context before and after words and sentences to grasp the intent of the customer’s speech and respond smoothly.
Skelter Labs’ outstanding intent recognition rate is attributed to its proprietary model that applies a rule-based method and a deep learning-based method in a hybrid form. The rule-based method recognizes intents based on sentence patterns learned from a large volume of natural language sentences. On the other hand, the deep learning-based method models a pattern of mutations that are difficult for the rule-based engine to process, allowing rapid learning with a small quantity of sample sentences. These two methods have been applied in a hybrid form for optimization. SkERT, a proprietary model of Skelter Labs based on the latest models such as SpanBERT, has recently been developed and applied. As a result, we have achieved remarkable growth in the F1 Score, which indicates the weighted average of precision and recall, a significant demonstration of accuracy.

Slot Filling
Slot Filling is a technology that collects and saves the needed information – Entity, which is the word with meaning – from a customer’s verbal output. Entity is classified and defined according to the predefined data by a chatbot creator, and it is sorted as names, locations, organizations, date and time. For example, if a customer asks about the nearest restaurant, AI needs to find out the location and preferred food to search for the appropriate information. For this, AI accurately recognizes entities from the customer’s verbal output to get the specific information and save it to the slot.
Skelter Labs’ entity recognition and slot filling added machine reading comprehension (MRC) technology to the generally used rule-based method. The MRC-based recognizer relieves the burden of securing training data by even conducting slot filling on non-learning words.
Slot Filling
Slot Filling is a technology that collects and saves the needed information – Entity, which is the word with meaning – from a customer’s verbal output. Entity is classified and defined according to the predefined data by a chatbot creator, and it is sorted as names, locations, organizations, date and time. For example, if a customer asks about the nearest restaurant, AI needs to find out the location and preferred food to search for the appropriate information. For this, AI accurately recognizes entities from the customer’s verbal output to get the specific information and save it to the slot.
Skelter Labs’ entity recognition and slot filling added machine reading comprehension (MRC) technology to the generally used rule-based method. The MRC-based recognizer relieves the burden of securing training data by even conducting slot filling on non-learning words.
Dialog Modeling
Existing dialog managers were designed to respond only to a generally defined conversation scenario, and were limited in that it was not possible to keep a conversation going for long. However, Skelter Labs’ dialog modeling has been developed to operate all nodes and create multiple dialogs within a single chatbot.
Especially by adopting Schema-guided Dialog State Tracking method, our dialog modeling supports learning-based response build which overcomes the boundary of the rule-based method. This method not only ensures high accuracy for cross-domain learning but also flexibly reacts to the cases that the chatbot creator has not predicted. You can greatly reduce the non-response rate of voicebots during phone calls or kiosk bot in public facilities by using our unique Dialog Manager.
Dialog Modeling
Existing dialog managers were designed to respond only to a generally defined conversation scenario, and were limited in that it was not possible to keep a conversation going for long. However, Skelter Labs’ dialog modeling has been developed to operate all nodes and create multiple dialogs within a single chatbot.
Especially by adopting Schema-guided Dialog State Tracking method, our dialog modeling supports learning-based response build which overcomes the boundary of the rule-based method. This method not only ensures high accuracy for cross-domain learning but also flexibly reacts to the cases that the chatbot creator has not predicted. You can greatly reduce the non-response rate of voicebots during phone calls or kiosk bot in public facilities by using our unique Dialog Manager.