Knowledgeable NLP

The First Workshop on Integrating Structured Knowledge and Neural Networks for NLP


Deep learning methods achieve state-of-the-art results in most NLP tasks, where a contributing factor to their remarkable performance is the ability to encode knowledge in terms of language models. For example, they can capture country-capital city association using word embeddings pre-trained on unlabelled textual data. However, in some cases, the required knowledge is rather specific and thus infrequent in text. This prevents neural models to learn it with an unsupervised approach. Still, if such specific knowledge were available, e.g., in terms of structured resources, deep learning models could boost their performance.

There exist many domain-specific structured resources (e.g., databases, ontologies) of good quality - being manually composed by experts, they are often reliable, have high coverage of their domain and in-depth information on individual objects.

The goal of this workshop is to deepen researchers expertise on the use of structured knowledge for NLP and to facilitate the development of this line of research, e.g., standardizing datasets, tasks, and techniques. Indeed, the best way to integrate structured knowledge into modern neural architectures is currently unclear. Namely, we would like to answer the following questions:

  • what is the best way to embed information available into the KB in a NN?
  • How should the KBs be externally accessed by NN models?
  • Are additional regularisation terms enough for KB access or should we look to other ways?

In particular, we would like to focus on papers that introduce new techniques for integrating structured knowledge and deep learning algorithms or apply such methods to new problems. Related to the topic of our workshop is the recent special issue of the Cambridge Natural Language Engineering journal on informing neural architectures for NLP with linguistic and background knowledge indicating the overall interest of the community to this topic

Topics of interest

Topics of interest include, but are not limited to:

  • Applications of structured knowledge for NLP tasks.
  • Integration of structured knowledge for NLP by querying a DB.
  • Leveraging knowledge graphs for NLP tasks.
  • Embedding structured resources for NLP tasks.
  • Reasoning and interpretability of models using structured knowledge.
  • Assessing the importance of structured knowledge for NLP.
  • Incorporation of multiple resources for NLP.
  • Introducing new resources that can be used for NLP.
  • Datasets and tasks to accurately measure the use of knowledge approaches.

Organising committee

Oren Sar Shalom - Intuit AI

Alexander Panchenko - Skoltech

Cicero dos Santos - Amazon

Varvara Logacheva - Skoltech

Alessandro Moschitti - Amazon

Ido Dagan - Bar Ilan University