airline Big Island Air Business Jet Services Ltd. city Aberdeen Volcano Burbank Stuart
date Houston information before they can compute an answer to their query or to build extraction pipelines (e.g., when using [1]) which however require substantial eforts.
Hence, we advocate for a diferent route where users can extract structured data relevant to satisfy an information need from a collection of text documents without the need to program, train or specify extraction systems.
Instead, the aim is to provide a system that allows users to explore new (unseen) text collections by simply issuing a query to receive structured information from the corpus. In contrast to [2] this should not require data already in tabular form, rather the idea is to automatically identify the relevant target structure and then, again automatically, fill it from unstructured text.
system for ad-hoc structured text exploration. The main Figure 1: Aim: Query a text collection and receive an approx- idea of WannaDB is that a user specifies their informaimate structured result without manual extraction tion need by composing SQL-style queries over the text collection. For example, in Figure 1, the user issues a query to extract information about dates, airlines, and Motivation In many domains, users face the problem cities of incidents. WannaDB then takes the query and of needing to quickly extract insights from large col- evaluates it over the given document collection by autolections of textual documents. For example, imagine a matically populating the table(s) required to answer the journalist who wants to write an article about airline query with information nuggets from the documents. security that was triggered by some recent incidents of To do this, WannaDB uses a novel pipeline as shown a well-known US airline. For this reason, the journalist in Figure 2 which first extracts a superset of informamight decide to explore a collection of textual accident tion nuggets from texts (e.g., all named entities), then reports from the National Transportation Safety Board in order to answer questions like ’What incident types are the most frequent ones?’ or ’Which airlines are involved aCpopmropxuitmea&tereretusurnlt most often in incidents?’. And clearly, there are many more domains where end users want to explore textual User doYcuetm, eexnitsctionlglecatpiponrosaicnhaessitmoialanrs wfaesrhisounc.h queries over qisusueerys ApproKxBimate new text collections force users to either read through SELECT... vast amounts of text and manually extract the relevant
Extract & organize (additional) relevant information determines the information need from the query, and ifnally matches nuggets to the relevant attributes of the user’s query. As a result, WannaDB allows to answer the queries, even if the information is not explicitly stated in the corpus but has to be calculated (e.g., when the query contains aggregation functions like AVG or SUM). A main observation here is that in many cases a sample of extractions (i.e., a table with partially missing or incorrect values) is suficient to produce approximate results to answer the user’s query.
Acknowledgments
Ministry of Education and Research as part of the Project Software Campus 2.0 (TUDA), Microproject INTEXPLORE, under grant ZN 01IS17050, by the German Research Foundation as part of the Research Training Group Adaptive Preparation of Information from Heterogeneous Sources (AIPHES) under grant No. GRK 1994/1, as well as the German Federal Ministry of Education and Research and the State of Hesse through the National High-Performance Computing Program.
Architecture & Initial Evaluation WannaDB contains components to determine the information need from queries, aggregate the relevant information, and References compute the actual query result. For the extraction of possibly relevant information nuggets, WannaDB relies on [1] C. De Sa, A. Ratner, C. Ré, J. Shin, F. Wang, of-the-shelf extractors like Stanza [ 3]. The key contribu- S. Wu, C. Zhang, Deepdive: Declarative knowltion of WannaDB, however, is a new matching approach edge base construction, SIGMOD Rec. 45 (2016) that uses a novel embedding space exploration algorithm 60–67. URL: https://doi.org/10.1145/2949741.2949756. incorporating interactive user feedback: The matching doi:10.1145/2949741.2949756. process is done separately for each relevant attribute. [2] S. Zhang, K. Balog, Ad hoc table retrieval using It starts by selecting information embeddings close to semantic similarity, in: Proceedings of the 2018 the attribute embedding. Afterwards, other embeddings World Wide Web Conference, WWW ’18, Internathat might be matches are searched by applying several tional World Wide Web Conferences Steering Comselection rules based on the closeness of embeddings to mittee, Republic and Canton of Geneva, CHE, 2018, known matches. Each candidate is presented for feedback p. 1553–1562. URL: https://doi.org/10.1145/3178876. (yes/no) to the user. The algorithm balances between ex- 3186067. doi:10.1145/3178876.3186067. ploration and exploitation to select those information [3] P. Qi, Y. Zhang, Y. Zhang, J. Bolton, C. D. Mannuggets for feedback that quickly allow identifying the ning, Stanza: A python natural language processareas in the embedding space relevant for the attribute ing toolkit for many human languages, ArXiv with as little feedback as possible. This area can then be abs/2003.07082 (2020). used to populate the remaining rows in the target table. [4] B. Hättasch, J.-M. Bodensohn, C. Binnig, Aset: AdPreviously extracted information (and user feedback) can hoc structured exploration of text collections, in: 3rd be reused for follow-up queries. A detailed description International Workshop on Applied AI for Database of the matching process can be found in [4]. Systems and Applications, Copenhagen, Denmark,
Our experiments on diferent text-collections each fo- 2021. URL: https://sites.google.com/view/aidb2021/ cused around certain topics lead to promising results: home/accepted-papers. 10 − 25 quick iterations of feedback for each attribute (i.e., confirming whether an information nuggets belongs in a certain column of a table) suficed for high matching scores for both textual and numeric attributes.
The Road Ahead In the next steps, we want to enlarge the scope, i.e., support more general corpora. We also want to leverage certainties from the extraction and matching process for the computation of the approximate result and provide useful interfaces for the end users, not only through a standalone application but also e.g., in form of a Jupyter Notebook extension.