Adaptive Access for a Digital Library of Corporate Information

1. Introduction

2. The problem: Turning EDGAR into a digital library of corporate information

• Data overload: Each company annually files anywhere from a dozen to several hundred documents, with some documents almost a thousand pages long. Only a tiny fraction of the information in these filings will be relevant to any one query. The software engineering challenges of providing rapid, highly reliable access to this volume of data are compounded by the difficulty of finding the right indices. The standard indices for corporate data, such as the four-digit Standard Industrial Classification (SIC) codes, are notorious for errors caused by false hits and missed data.
• Incomparability of data: Numbers such as earnings reports must be evaluated within the context of other information, such as industry-specific conventions and variations in accounting policies terminology and reporting period, etc. This information is often reported in the management discussions and the footnotes that accompany financial statements, but are only indexed on specialized databases such as NAARS; complex concepts such as "securitization of accounts receivable" are usually not indexed at all.
• Missing or incorrect data: Computerized databases often contain missing fields, old data, and omit information from filings such as registration statements or 8-Ks. Coverage is also likely to be incomplete for small firms or those not actively traded. Often data entry errors creep into the records, entered as text without integrity constraint checking.

Our approach: Using adaptive techniques to capture meta-information

EDGAR filings will first come into RELATE over the Internet as a series of ASCII text documents. Bill Kornfeld, a consultant to Price Waterhouse, has implemented a pre-processor to decompose the filings before storing them in order to focus the effort of the subsequent content-based analysis. For example, each 10-K (the formal annual report required to be filed with the SEC within 90 days of the end of each fiscal year) must have certain items and schedules, and the set of financial statements that includes the income statements mentioned above will always be in Item 8. Our analysis indicates that finite state grammars (implemented in Prolog) are both sufficiently powerful and efficient to extract this information. We are also making rapid progress in the use of finite state grammars to parse tables from the raw EDGAR text, identifying row and column headers with high accuracy, despite significant variation and noise in the original data. Weak grammars may also be sufficient to find descriptive information such as listings of products and services, and meta-information about accounting policies such as inventory valuation methods or depreciation schedules.

The next phase of processing the data is more semantically-based, and will be more difficult to implement. RELATE will attempt to identify as much structure as possible in the extracted information and provide additional indexes where appropriate. So for example, the table representing an income statement may be decomposed into a net-income section and a per-share-section, while the net-income section may have subsections for operating income, adjustments, net income, etc. This decomposition is a heuristic process and depends on having detailed models of financial items and their inter-relationships. Our goal is for this stage to identify as much of the explicit structure as possible in the individual filings at storage time. Then, on an as-needed basis at query time, further analysis will derive implicit information (such as performance ratios) and perform comparisons across multiple filings. This requires problem solving to apply knowledge about the repository structure, semantics of the data, and strategies for conflict detection and resolution. We are building such reasoning capabilities using the Soar architecture for building intelligent agents (Laird et al., 1987). Soar responds to difficulties in problem solving by creating a series of subgoals, and caches the results of problem solving for each subgoal using an experience-based learning mechanism called chunking. Each section of income statements referred to above has corresponding operators within Soar, and a simple induction algorithm is used to learn applicability conditions for each operator. In RELATE, this will lead to more efficient and accurate information access over time.

A final issue we are addressing is how a user will interact with RELATE to offer guidance about how to detect and resolve inconsistencies during data access and integration. In the case of an income statement for an airline, a line item for "aircraft fuel" row may be listed between the "Operating expenses" header and the "Total operating expenses" total. RELATE could notice that all the line items between the two add up to the total, which would lead it to posit that "aircraft fuel" is a type of operating expense. RELATE could learn to look for such items in the context of income statements from airlines, so that future queries could then quickly compare fuel expenditures across airlines over time. We plan to develop techniques, based on previous research on learning from instruction in Soar, for the user to monitor RELATE's progress and direct its attention to relevant features of the situation.

The RELATE project is clearly a long-term effort, and many challenges remain. Yet we have been pleasantly surprised so far by the progress we have been able to make with relatively simple techniques (e.g. finite state grammars for text scanning). Our hope is that our luck will hold out, and relatively simple learning techniques such as Soar's chunking will afford us the same leverage in developing capabilities for adaptive access.

References