Interactive Machine Learning for Compositional Models of Natural Language
Type of activity
Commission of European Communities
Funding entity code
Natural language understanding (NLU) is a branch of artificial intelligence (AI) that uses computer software to understand input made in the form of sentences in text or speech format. Think of Siri answering a question about what the traffic is like this morning or Alexa asked about the weather in your city. In other words, NLU digests a human text, translates it into computer language and produces an output in human language. NLU applications have unique information needs and require large collections of annotated data to achieve good results. The EU-funded INTERACT project will develop new interactive learning algorithms (ILA), motivated by applications in NLU. It will merge representation learning and active learning of compositional latent-state models (CLSMs) since natural language is rich, complex and compositional. INTERACT will develop new Interactive Learning Algorithms (ILA), motivated by applications in Natural Language Understanding (NLU). The main assumptions behind supervised approaches are unrealistic because most NLU applications have unique information needs, and large collections of annotated data are necessary to achieve good performance. INTERACT follows a collaborative machine learning paradigm that breaks the distinction between annotation and training. We focus on compositional latent-state models (CLSMs) because natural language is rich, complex and compositional. To reduce the amount of human feedback necessary for learning CLSMs we must eliminate annotation redundancy. We argue that to achieve this in the context of CLSMs we must combine: (1) An optimal human feedback strategy, with (2) inducing a latent structure of parts in the compositional domain. Annotation effort will be minimized because the method will only request representative feedback from each latent class. INTERACT marries representation learning (i.e. of parts) and active learning for CLSMs.
Our approach goes beyond classical active learning where the ILA asks labels for samples chosen from a pool of unlabeled data. We empower the ILA with the ability to ask for labels for any complete or partial structure in the domain, i.e. the ILA will be able to generate samples.
We work under the framework of spectral learning of weighted automata and grammars and use ideas from query learning. A key idea is reducing the problem of interactive learning of CLSMs to a form of interactive low-rank matrix completion. Our concrete goals are: (1) Develop ILAs for CLSMs based on spectral learning techniques; and (2) Investigate optimal strategies to leverage human feedback, taking into account what is optimal for the ILA and what is easy for the teacher.
We will experiment with NLU tasks of increasing complexity, from sequence and tree classification to parsing problems where the outputs are trees.