Much of the world’s knowledge has been digitized in the forms of databases, knowledge bases, or free texts. Natural language interface (NLI) helps utilizing such knowledge by transforming natural language questions into queries on the underlying data store. It is becoming an essential part of modern business intelligence and knowledge management applications.
Effective NLIs rely on robust natural language processing (NLP) components to correctly parse the question. These tasks can be accomplished using a pipeline of rule engines, statistical classifiers and taggers. Recent research has also shown promising progress in end-to-end question parsing, which transform questions directly into structured queries.
Wide adoption of NLIs is dependent upon their ability to adapt to new domains. A lot of NLI use cases root in the needs to understand business or technical data and are therefore highly domain-specific. Much research has been focusing on rapidly bootstrapping NLIs for new domains, including in-domain training data creation, semi-supervised learning, active learning, and transfer learning.
Dialog management allows NLI systems to provide a natural interaction experience: the system would be able to respond based on previous questions and answers and would confirm possible ambiguities with the user. Context-aware NLP models and confidence estimation models are leading to more interactive multi-turn dialog in NLI systems.
- A question parsing system that is generalizable to different domains and use cases.
- Domain adaptation and data bootstrapping models for question parsing that is able to build NLI models for new domains without significant annotation effort.
- Dialog strategies for multi-turn NLI interactions that is aware of the dialog context and can perform disambiguation through interaction.
Related Research Topics
- Models for text classification and sequence tagging
- Multi-task intent and slot detection
- End-to-end NL-SQL/SPARQL parsing
- Training data bootstrapping for question parsing
- Weakly-supervised semantic parsing
- Context-aware intent and slot detection
- Confidence estimation in classification and sequence tagging models
- Dialog management