In this blog post, we give a brief overview of which problem INODE solves, which approaches INODE uses and highlight the major challenges. We also provide links for further reading.

1 Which Problem does INODE Solve?

Enterprises and scientific organizations often keep vast amounts of data in large databases. Accessing these databases requires deep knowledge in database query languages such as SQL. However, end-users who do not have sufficient technical know-how, are basically excluded from querying these information sources effectively.

The goal of INODE is to enable people to explore different types of data through natural language as well as visualizations. On the one hand, INODE empowers users to interact with data stored in databases as well as in text documents similar to a dialog with a human. On the other hand, INODE proactively guides the user during data exploration and visually explains complex patterns in data.

Assume that you are a scientist who wants to study lung cancer. Further assume that you want to find out which specific biomarkers are indicators for a certain type of lung cancer. How would you tackle this problem?

One approach would be to sit down and read about what other researchers have published. This is a feasible approach but it might take quite a while to find the relevant information. Another approach would be to query the latest bioinformatics databases that contain curated information. However, these databases are typically very complex and you might not have enough technical knowledge to query these databases effectively. A better solution might be to have a tool or service that assists you in exploring the data and that guides you in the right direction.

The following figures demonstrate INODE’s vision of intelligent data exploration to tackle the above-mentioned use case. Figure 1 shows a natural language query with user assistance. INODE parses the query, provides hints for autocompletion and disambiguates terms. After the user has explored data in natural language, she can visually analyze the results. Figure 2 clusters cancer types retrieved by the natural language query. The user can explicitly choose, for instance, the distance metric for the cancer types and which biomarker to further explore.

2 Which Approaches does INODE Use?

INODE uses different types of technology for data exploration:

• Natural language interfaces to databases:
  Traditional databases require end-users to query data with specific query languages such as SQL or SPARQL. However, end-users who do not speak these query languages proficiently, are basically excluded from exploring data efficiently. Hence, INODE will use technology such as SODA (Blunschi et al. 2012), Bio-SODA (Sima et al. 2019) and Logos (Kokkalis et al. 2012) to build novel natural language interfaces for databases to explore data in a more human-like dialog. INODE will offer a novel user experience that blends natural language queries, exploration advice (recommendations) and natural language explanations to implement a conversational experience between the user and the system. The system converses with the user, not only returning query results but offering explanations and suggestions, enabling a more natural bilateral interaction with the user. For a detailed discussion on recent trends in natural language interfaces to databases we refer to an ACM SIGMOD blog post “The Rise of Natural Language Interfaces to Databases” (Stockinger, 2019) as well as to a survey paper “A Comparative Survey of Recent Natural Language Interfaces for Databases” (Affolter et al. 2019).

• Knowledge graphs for data integration and querying:
  Knowledge graphs are a widely used approach to model, integrate and enrich data. These knowledge graphs can also be used to semantically query data across data sets. In INODE, we will use the ontology-based data access technology OnTop (Calvanese et al. 2017) to integrate data from relational databases as well as from graph databases. Moreover, for data that is stored in unstructured text documents, INODE will leverage Noima (Papadakis et al. 2019) to extract the major entities and relationships out of these documents and integrate that information into a knowledge graph to enable queries across databases and text documents.

• Multi-modal data exploration:
  Making databases more accessible with the help of natural language interfaces is just one side of the coin. On the other side, interpretation and analysis of the data has to get easier, too. To foster that, we will build tools and services for data exploration that enable combining natural language queries with visual data analysis to proactively guide the user to find relevant information. In the lung cancer example above, the database schema includes literally thousands of classes. To guide the user in choosing the right classes, we will build on existing approaches to visualize the relevant parts of the schema for easier understanding (May et al. 2012). Moreover, we will extend COVIZ (Omidvar-Tehrani et al. 2018) that enables by-example exploration on medical datasets. We will also leverage VCaaS for visual exploration.

3 What are the Main Challenges?

Traditionally, the database community tackled problems related to analyzing structured data with SQL or SPARQL as the major query languages. On the other hand, the information retrieval or natural language processing communities tackled problems related to unstructured, text data with keyword search or natural language interfaces to data. INODE will break down these barriers between the communities and build services that enable exploring both structured and unstructured data in natural language. However,  because natural language is ambiguous, it is hard for a machine to be able to understand the intended meaning of the users. Hence, we will build novel services that enable combining natural language queries with visual exploration. The major challenge here is how to design visual interfaces that enable exploring large amounts of information (typically in the form of knowledge graphs) without overloading the users with too many details that distract from identifying the core insights.

References

Affolter, K., Stockinger, K., & Bernstein, A. (2019). A comparative survey of recent natural language interfaces for databases. The VLDB Journal, 28(5), 793-819.

Blunschi, L., Jossen, C., Kossmann, D., Mori, M., & Stockinger, K. (2012). Soda: Generating sql for business users. Proceedings of the VLDB Endowment, 5(10), 932-943.

Calvanese, D., Cogrel, B., Komla-Ebri, S., Kontchakov, R., Lanti, D., Rezk, M., … & Xiao, G. (2017). Ontop: Answering SPARQL queries over relational databases. Semantic Web, 8(3), 471-487.

Kokkalis, A., Vagenas, P., Zervakis, A., Simitsis, A., Koutrika, G., & Ioannidis, Y. (2012). Logos: a system for translating queries into narratives. In Proceedings of the 2012 ACM SIGMOD International Conference on Management of Data (pp. 673-676).

May, T., Steiger, M., Davey, J., & Kohlhammer, J. (2012, June). Using signposts for navigation in large graphs. In Computer Graphics Forum (Vol. 31, No. 3pt2, pp. 985-994). Oxford, UK: Blackwell Publishing Ltd.

Omidvar-Tehrani, B., Amer-Yahia, S., & Lakshmanan, L. V. (2018). Cohort representation and exploration. In 2018 IEEE 5th International Conference on Data Science and Advanced Analytics (DSAA) (pp. 169-178). IEEE.

Papadakis, N., Litke, A., Doulamis, A., Protopapadakis, E., & Doulamis, N. (2019). Multimedia Analysis on User-Generated Content for Safety-Oriented Applications. In Social Media Strategy in Policing (pp. 161-175). Springer, Cham.

Sima, A. C., Mendes de Farias, T., Zbinden, E., Anisimova, M., Gil, M., Stockinger, H., … & Dessimoz, C. (2019). Enabling semantic queries across federated bioinformatics databases. Database, 2019.

Stockinger, K. The Rise of Natural Language Interfaces to Databases. SIGMOD Blog, June 2019, https://wp.sigmod.org/?p=2897