What is it about?
Open-domain question answering on count queries of the form "how many .." often have varied distribution of numbers across texts. We aggregate evidence across text sources to provide a more comprehensive answer to the user. We found that providing explanations from multiple texts supporting the system predictions leads to better user comprehension. The user knows whether something is working or not and why is it that way.
Featured Image
Photo by Fedor on Unsplash
Why is it important?
Count queries like "How many songs did John Lennon write for the Beatles?" are difficult to answer using only a single source of text. While one snippet talks about the total songs composed by the Beatles as a band, another snippet might talk about the number of songs written by the Lennon-McCartney duo. In search engine result pages (SERP), the user is shown a top result which might contain partial or wrong result, leaving her confused if contradicting numbers are highlighted in subsequent search snippets.
Perspectives
This work brings into focus user comprehension of open-domain question answering through evidence aggregated from multiple texts. While high confidence results on popular entities or topics guarantee a well-formed answer either from a knowledge base or a single text source like Wikipedia for more complex queries or queries on long-tailed entities it is essential that users are informed of the variability in the text sources from which the answers are derived.
Shrestha Ghosh
Max Planck Institute for Informatics
Read the Original
This page is a summary of: Answering Count Queries with Explanatory Evidence, July 2022, ACM (Association for Computing Machinery),
DOI: 10.1145/3477495.3531870.
You can read the full text:
Contributors
The following have contributed to this page
