Multi-Domain Dialogue Systems
Interest in dialogue systems has been increasing drastically as smart phones are getting widely used and robotics and AI have made rapid progress. However, we are still in a stage where the majority of dialogue systems can simply provide an answer to a question. HRI-JP is researching dialogue systems which will enable more sophisticated language-based communication between humans and machines, based on advanced Natural Language Processing (NLP) techniques.
A natural dialogue consists of multiple exchanges. However, most existing systems can deal with specific topics such as restaurant search and hotel reservations. Such a conversation topic is referred to as a “dialogue domain.” Whereas single-domain dialogue systems are capable of limited conversation about a single topic,“multi-domain” dialogue systems can engage in multiple domains of conversation depending on the input from the user.
HRI Intelligence Platform Based on Multiple Experts (HRiME) is a research framework for developing multi-domain dialogue systems. HRiME is composed of several modules for understanding and generating language and enables building various types of systems by adding knowledge bases depending on dialogue domains and replacing some modules by ones employing state-of-the-art algorithms.
There are two kinds of a dialogue system; one is task-oriented dialogue whose aim is to execute tasks such as making reservations or performing searches, and the other is a system whose purpose is to continue a conversation. HRI-JP is engaged in research into not only both types of systems but also systems integrating both. For instance, an interview is a dialogue with a defined purpose. However casual conversations during an interview gives users a better impression of the system. HRI-JP is investigating such kind of interview dialogue systems.
Building a dialogue system requires creating a knowledge base containing dialogue scenarios and dictionaries. It is necessary to simplify the knowledge base construction process. We have been testing different methods to extract knowledge form large datasets as well as using self-learning, where the system asks the user about the definition of a given word and adds it to the knowledge base.
The topic in conversation to help human will changes every moment like restaurant, weather, or schedule. The contents for each topic is known as the Dialogue Domain.
DeepNNNER: A Named Entity Recognition System
Named Entity Recognition (NER) is an important technology for understanding human language by identifying expressions that refer to specific entities in the world, such as the names of people, places, and organization. NER has many applications, including the detection of important entities in conversations.
HRI-JP has developed an NER system DeepNNNER, that applies a special neural network known as a bidirectional long short-term network (BLSTM) that can use an unlimited amount of context and learn long distance dependencies in data, resulting in high performance named entity detection.
DeepNNNER also uses two innovations to detect named entities in words it has never seen before. 1) it applies convolutional neural networks (CNNs) to characters in words, learning important spelling patterns. 2) a flexible matching algorithm applies existing knowledge by looking up word sequences in a dictionary of known entities.
In 2016, DeepNNNER achieved the world’s highest performance level in an evaluation conducted on both English newspaper and multi-genre Web text data sets. DeepNNNER is also used in HRI-JP’s dialogue system.
A hybrid model of bidirectional long short-term memory (BLSTM) networks and convolutional neural networks (CNN) that automatically learns both character- and word-level features (Chiu and Nichols (2016))