No abstract

No abstract

No abstract

Information surrounds us and keeping track of relevant details can be challenging. Although there are multiple applications to take notes, organize ideas, or set reminders, existing solutions are semantic-agnostic and rely on the user to manually search for desired information by keywords. We propose a novel method to help people store and retrieve such details with ease in Romanian language. Our conversational agent built on top of the RASA framework is capable to extract relevant information from the user's utterances, store them in a persistent knowledge graph, and ultimately, access them when requested. A set of specific intents regarding locations, timestamps, and properties were created and learned by the agent using manually built examples. In addition, an interaction logic based on a knowledge graph was added to enable the storage and retrieval of information, based on the identified semantic components from the input sentences. The performed tests showed a good accuracy for intent detection, and promising results for the sentence parser. Our conversational agent is accessible as a web application which can process text or speech inputs, and responds with a textual representation of the user's memorized facts.

Technology is becoming omnipresent in our lives due to its accessibility and ease of use. Conversational agents facilitate interactions in natural language and are frequently employed to perform repetitive tasks in a specific context. We introduce a conversational agent for Romanian built on top of the open-source RASA framework, capable to communicate in predefined microworlds. Two scenarios were considered, namely: a smart home assistant which interprets commands to IoT devices, and an interactive infopoint for our university focusing on providing guidance to students. Several enhancements were considered, including an NLP pre-processing pipeline from spaCy and a knowledge graph implemented using Grakn for conceptualizing the information accessible to the agent. Our agent can quickly classify intents and extract entities with high accuracy for a given microworld (F1-score of 97% for the first microworld and 93% for the second). A survey on 10 users showed high satisfaction in terms of the usefulness and the succinctness of the provided information.