Systems and methods for retrieving tabular data from textual sources5950196Abstract Tables form an important kind of data element in text retrieval. Often, the gist of an entire news article or other exposition can be concisely captured in tabular form. Information other than the key words in a digital document can be exploited to provide the users with more flexible and powerful query capabilities. More specifically, the structural information in a document is exploited to identify tables and their component fields and let the users query based on these fields. Component fields can include table lines, caption lines, row headings, column headings, or other table components. Empirical results have demonstrated that heuristic method based table extraction and component tagging can be performed effectively and efficiently. Moreover, experiments in retrieval using the system of the present invention strongly indicate that such structural decomposition can facilitate better representation of user's information needs and hence more effective retrieval of tables. Claims What is claimed is: Description FIELD OF THE INVENTION
TABLE 1
______________________________________
Table extracted from the Wall Street Journal Database
by the table extractor
Such overtime throughout the past year helped Ford boost 1988-model car
and truck production 19% from 19-7 levels, the company said.
Auto Production
______________________________________
Fourth-quarter U.S. auto production
PERCENT
1988 1987 CHANGE
______________________________________
GM 964,653 843,908 +14.3%
Ford 456,954 461,150 -0.9
Chrysler 294,998 283,062 +4.2
Honda 86,330 80,227 +7.6
Mazda 59,464 3,800 --
Nissan 24,604 15,289 +60.9
Toyota 19,600 0 --
Nummi 16,156 38,901 -58.5
Diamond Star
2,316 0 --
TOTAL 1,925,075 1,726,337 +11.6%
______________________________________
First-quarter U.S. auto production
PERCENT
1989 1988 CHANGE
______________________________________
GM 925,000 780,2123
+18.7%
Ford 475,000 479,263 -0.9
Chrysler 271,000 261,433 +3.7
Honda 99,360 92,156 +7.8
Mazda 61,510 16,800 +266.1
Nissan 29,522 36,324 -18.7
Toyota 25,000 -- --
Nummi 48,000 41,479 +15.7
Diamond Star
10,000 -- --
TOTAL 1,945,392 1,707,673 +13.9%
______________________________________
NOTE:
Estimated from industry and company sources. Vehicles are counted by
manufacturer;
the Mazda total includes cars built for Mazda and Ford;
DiamondStar includes Chrysler and Mitsubishi cars and
Nummi includes Chevrolet and Toyota cars.
Although the column and row headings, and the relevant captions, are extracted by an appropriately tuned table extractor 20 along with the table entries, they are not yet identified as caption data. Caption data and heading data is identified by the component tagger 30. The component tagger 30 accepts as input an extracted text table of the sort shown in Table 1 above. It is understood that an input table may be extracted, or the data may simply be provided in tabular form. Regardless, these tables will be referred to as "extracted tables." The component tagger applies a number of heuristic techniques to determine if a line contained in an extracted table is caption data or is a table entry. Various techniques must be used to filter, separate, and sometimes undo the interactions between previously applied techniques, due to the numerous formats in which tables occur, and various idiosyncrasies which are present, in textual documents. The component tagger 30 employs mainly syntactic heuristics, as opposed to semantic heuristics, to identify caption data and table entries. In some embodiments, column headings are also treated as, and tagged as, caption data. The component tagger 30 may use any mechanism to tag lines appearing in the extracted table as either table entries or caption data; in one embodiment, the component tagger 30 tags table lines using <TABLE><LINE FROM EXTRACTED TABLE></TABLE> tags and tags caption using <CAPTION><LINE FROM EXTRACTED TABLE></CAPTION> tags. The component tagger 30 may also tag other table components such as row headings, column headings, or other table components. Some examples of useful heuristics follow. The following examples may be better understood by referencing Table 2, below. Table 2 depicts an extracted data table in which each of the characters in the extracted table is replaced by a "*" and the blanks appear as extracted from the textual document.
TABLE 2
______________________________________
Extracted table with entries replaced by "*"
______________________________________
********** *** ********
************** ********* ** ***************
********** *********
************* **************
********** *** ********* ***
************
** ******** **
****** ** ****** **
*** **** *****
** ******* **
______________________________________
Gap Structure Heuristic This heuristics uses the fact that large, empty spaces, i.e., big gaps, in the middle of an extracted table line indicates that the line is a potential table entry. Large gaps at either the beginning or end of a line are not dispositive because such gaps may only indicate that the table has been centered in the text document from which is was extracted. Referring to Table 2 above, this heuristic would identify the last five lines of the table as possible table entries. Table 2 illustrates the possible limitation of this heuristic, since the column headings appearing in the fourth line of Table 2 would tagged as a table entry, instead of caption data. This example also illustrates that heuristics may need to be applied seriatim, and possibly even in some particular order, to correctly tag all lines. Alignment Heuristic This heuristic utilizes the characteristic that if the gaps present in the middle of a line, as well as the characters, of two or more lines are aligned, those lines are likely to be table entries. Again, gaps at the beginning or end of a line may only indicate formatting present in the document from which the table was extracted. Referring to Table 2, this heuristic would indicate that the last three lines of the table are table entries. Pattern Regularity Heuristic This heuristic exploits the fact that context information, or regularity, can be used to identify table lines and captions. For example, in the pattern shown in Table 3 below, the third and fifth lines are more likely to be table entries rather than caption data. The "captions" occurring between table lines in such patterns are often table lines that have a different gap structure than the other table lines.
TABLE 3
______________________________________
Table pattern
______________________________________
Caption
Table-Line
Caption
Table-Line
Caption
Table-Line
______________________________________
Table 4, shown below, is an extracted data table exhibiting this pattern.
TABLE 4
__________________________________________________________________________
An erroneously tagged table (by the earlier heuristics)
that is later rectified by the Pattern Regularity Heuristic
__________________________________________________________________________
<CAPTION> While 52% of consumers surveyed in March by Leo J. Shapiro
&.backslash.
Associates said Sears was the best place to buy power tools, only
1%.backslash.
said it was the best place to buy a woman's dress blouse. </CAPTION>
<CAPTION>
New Sears prices vs. the competition </CAPTION>
<CAPTION>
SEARS
PENNEY
WAL-MART
WARD
TARGET
</CAPTION>
<TABLE>
Levi 501 jeans men </TABLE>
<TABLE>
$ 19.96
$ 19.99
N.A. N.A.
N.A.
</TABLE>
<CAPTION>
Vanderbilt one-ounce eau de toilette spray </CAPTION>
<TABLE>
15.96
15.50
$ 13.84
$ 15.50
$ 13.95
</TABLE>
<CAPTION>
Fruit of the Loom Men's underwear (three-pack) </CAPTION>
<TABLE>
3.96
N.A.
3.96 N.A.
3.99
</TABLE>
<CAPTION>
Prestone antifreeze </CAPTION>
<TABLE>
7.62
N.A.
7.47 7.99
7.99
</TABLE>
<CAPTION>
Sealy mattress queen size </CAPTION>
<TABLE>
549.95
557.75
N.A. 549.99
N.A.
</TABLE>
<CAPTION>
Lego basic building set </CAPTION>
<TABLE>
9.99
N.A.
9.57 9.99
9.99
</TABLE>
<CAPTION>
Etch a Sketch </CAPTION>
<TABLE>
8.99
N.A.
6.97 8.99
8.99
</TABLE>
__________________________________________________________________________
Another example of the way in which the Pattern Regularity Heuristic may be applied is illustrated by Table 5 below. Table 5 depicts a data table having an intermittent caption. Intermittent captions, i.e., caption data which is separated by more than one table entry, should not be corrected by the Pattern Regularity Heuristic, since those lines are more likely to be interspersed caption data and not mistagged table entries.
TABLE 5
______________________________________
An extracted data table with intermittent caption
______________________________________
<CAPTION> Market Share in Europe </CAPTION>
<CAPTION> (Figures are percentage for 1987) </CAPTION>
<CAPTION> Central-office telephone-switching lines </CAPTION>
<TABLE> Alcatel 40.0% </TABLE>
<TABLE> Plessey/GEC 16.7 </TABLE>
<TABLE> Ericsson 13.6 </TABLE>
<TABLE> Siemens 10.2 </TABLE>
<TABLE> Italtel 8.4 </TABLE>
<TABLE> Others 11.1 </TABLE>
<TABLE> TOTAL SHIPMENTS: 9.6 million lines
<CAPTION> Semiconductors </CAPTION>
<TABLE> Phillips 14.6 <TABLE>
<TABLE> SGS-Thomson 8.6 <TABLE>
<TABLE> Texas Instruments
7.7 <TABLE>
<TABLE> Motorola 7.5 <TABLE>
<TABLE> Siemens 7.5 <TABLE>
<TABLE> National Semiconductor
5.4 <TABLE>
<TABLE> Intel 4.5 <TABLE>
<TABLE> NEC 3.9 <TABLE>
<TABLE> ITT 3.8 <TABLE>
<TABLE> Advanced Micro Devices
3.7 <TABLE>
<TABLE> Plessey 3.7 <TABLE>
<TABLE> TOTAL MARKET: $6.36 billion
</CAPTION>
Source: Dataquest Inc.
</CAPTION>
______________________________________
Differential Column Count Heuristic This heuristic exploits the fact that lines appearing at the beginning of an extracted table which have fewer columns than the average number of columns present in the table are potential caption data. For example, if each line in a table has 4 columns, a line at the beginning of the table having only 3 columns, is a potential caption and should be marked as caption data. An illustration of an extracted table on which this heuristic could be advantageously used is shown below in Table 6.
TABLE 6
______________________________________
The second and third lines are tagged captions
by the Differential Column Count Heuristic
______________________________________
</CAPTION>
The percentage change is since year-end. </CAPTION>
</CAPTION> % This </CAPTION>
</CAPTION> Nov. 15 Nov. 14
Year </CAPTION>
<TABLE> U.S. 248.1 247.5 +8/1 <TABLE>
<TABLE> Britain 548.1 545.4 +6.3 <TABLE>
<TABLE> Canada 350.7 349.4 0.0 <TABLE>
<TABLE> Japan 1350.8 1340.1
+32.1 <TABLE>
<TABLE> World index
478.9 476.1 17.4 <TABLE>
______________________________________
Differential Gap Structure Heuristic The Differential Gap Structure Heuristic tags as caption data any lines at the beginning or end of a table having a gap alignment that does not match the average gap alignment for the rest of the lines in the extracted data table. Table 7 is an example of a table which benefits from use of this heuristic.
TABLE 7
__________________________________________________________________________
The second line of this extracted table is tagged as caption data by the
Differential Gap Structure Heuristic
__________________________________________________________________________
<CAPTION>
<CAPTION>
NAME (In billions)
SHARE CARDS*
RATE
</CAPTION>
<TABLE>
Citicorp $25.0 16.9% V/MC 19.80
</TABLE>
<TABLE>
Chase Manhattan
9.0 6.1 V/MC 19.80
</TABLE>
<TABLE>
Sears 8.5 5.8 D 19.80
</TABLE>
<TABLE>
First Chicago
6.6 4.5 V/MC 19.80
</TABLE>
<TABLE>
BankAmerica
6.0 4.1 V/MC 19.80
</TABLE>
<TABLE>
MHC Financial
5.3 3.6 V/MC 19.80
</TABLE>
<TABLE>
American Express
5.2 3.5 0 16.25
</TABLE>
<TABLE>
Bank of New York
3.8 2.6 V/MC 16.98
</TABLE>
<TABLE>
Manufacturers
2.9 2.0 V/MC 19.80
</TABLE>
Hanover
<TABLE>
Wells Fargo
2.5 1.7 V/MC 19.80
</TABLE>
<TABLE>
Total for top 10
$74.8 </TABLE>
<CAPTION>
(or about 51% of industry vs. about 47% in 1987)
</CAPTION>
<CAPTION>
*V-Visa; MC = MasterCard; D = Discover; 0 = Optima
</CAPTION>
<CAPTION>
Sources: Individual issuers; RAM Research Bankcard
</CAPTION>
__________________________________________________________________________
If a line immediately following a caption has a format that is different from that of the average table format, that line is subjected to further tests to determine if it should be tagged as caption data or tagged as a table entry. A first test to which the line may be subjected is determining if the line is mostly numerical or mostly textual. This may be done in any of a variety of ways, the simplest being to count the number of text characters versus the number of numeric characters appearing in each line. Lines having mostly numerical data should be tagged as table entries, while lines having mostly textual characters should be tagged as caption data. If the first test is indeterminate, or as a second test to increase confidence, the lines may be examined to determine if the textual characters which are present are in upper-case or mixed case. Column headings, and therefore caption data, typically appear as upper case textual entries while table entries are typically lower case or mixed case textual entries. In addition, other heuristics may be used, either as a primary method of determining whether a line is caption data or a table entry, as a secondary method to be used in conjunction with a primary method, or as a tie breaking method should two or more of the main heuristics listed above give contradictory results. For example, centered lines that display a gap alignment characteristic which is different from the average gap alignment characteristics for the remainder of the table may be tagged as caption data. Second, the length of a line may be compared to the average line length for the other lines of the extracted data table. If the line's length is very different from the average length for lines in the extracted data table, then that line may be tagged as caption data. Lines may be examined to determine if they include words like "Table" or "Figure." Lines which include such words may be tagged as caption data. For tables in which the first line of a multiline caption has been identified, each following line may be tagged as caption data until a period is encountered. If the period occurs in the middle of a line, continue tagging lines as caption data until the next blank line in the extracted table is encountered. Although the component tagging heuristics have been described in a general order, the order and number of heuristics to apply to a table may vary depending on characteristics of the corpora from which the table has been extracted. For example, pattern regularity heuristics need not be applied at all for extracted tables which do not exhibit interspersed caption and tables. Further, even for tables which do exhibit such behavior, pattern regularity heuristics must be applied after another heuristic that tentatively tags caption and table lines. Once a table has been extracted and tagged, it is field indexed to facilitate its inclusion in a structured database environment. Any field indexing program may be used to index the table on the basis of any one or more table components. For example, a suitable indexing product which may be used to create the index based on table components is the indexing module for InQuery, manufactured by Soverign Hill Software, Inc. of Dedham, Mass. Once a table has been field indexed, a user may query it using any sort of information retrieval engine, such as a Boolean retrieval engine, a natural language retrieval engine, or a probabilistic retrieval engine. Table documents may be field indexed at least partially by caption and table fields, giving the user the flexibility to query based on those fields. This provides greater flexibility because occurrence of words in a caption versus their occurrence as a table entry generally carries a different import to the user. For example, a query like "health insurance united states" could represent a variety of user requirements. It could mean that the user really wants to know about health insurance in general with some association to the United States or she could be wanting to know specifically about health insurance in the United States. In the former case "health insurance" should occur in the caption and get more weight whereas in the latter case the entire set of words in the query should get equal weight. Another example is a query like "Stock prices of oil companies". A table that has all the words in the caption field is likely to be more relevant than a table with "stock prices" in captions and "oil companies" as table entry or "oil companies" in captions and "stock prices" as table entry. A table with "stock" and "companies" in captions and "prices" and "oil" as table entries is highly likely to be irrelevant. Thus, weighting the query based on fields can give the user the flexibility and power to manipulate and express information retrieval needs more precisely. EXPERIMENTS The System The table extraction process is subject to a precision versus recall trade-off. Precision (getting only the table lines with no extraneous information) may have to be sacrificed in order to improve recall (getting all table in the text along with some non-table lines) or vice versa. The preprocessing module 12, generally, and the table extractor 20 in particular, can be tuned by adjusting the parameters alluded to above, i.e., K and M. If the parameters are highly constrained, precision improves because spurious lines are not included, but recall drops because it is more difficult for the heuristics described above to properly separate table entries from caption data. A high precision version of the system and a high recall version of the system were used for the experiments described below. As an example of the trade-off between precision and recall, the high precision version of the system excluded from Table 4 many lines, such as "Vanderbilt one-ounce eau de toilette spray" and "Fruit of the Loom men's underwear (three-pack)." The high recall version of the system, on the other hand, extracted all table entry lines along with one extraneous line "While 52% of consumers surveyed in March by Leo J. Shapiro & Associates said sears was the best place to-buy power tools, only 1% said it was the best place to buy a woman's dress blouse." Tuning a system for higher recall may ultimately prove, for many corpora, to be more advantageous, since it is highly improbable that a query term will match a randomly extracted word in an extraneous line. Hence, even if higher recall brings in extraneous lines, it may not affect the retrieval of the relevant tables. In addition, using proper weighting functions should eliminate the significance of randomly extracted sentence words versus relevant words. Results of Extraction and Tagging The preprocessing module 12 described above was run for table extraction and component tagging on Wall Street Journal database spanning the years 1987-1992. Table 8 shows the statistics for this process. The final structured document database was built using the collated set of 6509 tables
TABLE 8
______________________________________
Table Extraction & Component Tagging Statistics
Wall-Clock Table Number of
Database Size Time db Size
Tables
______________________________________
WSJ87 131 MB 17 min 3.7 MB
1857
WSJ88 109 MB 15 min 3.4 MB
1701
WSJ89 38 MB 4 min .3 MB
222
WSJ90 73 MB 10 min 2.4 MB
930
WSJ91 145 MB 20 min 4.3 MB
1417
WSJ92 34 MB 5 min 1.1 MB
382
TOTAL 530 MB 71 min 15 MB
6509
______________________________________
Tables 9 and 10 show the accuracy statistics for a set of 100 documents, each having 50 tables in them. The table extractor 20 did not exclude any table in its entirety. The errors were few and spread out. In the case of the component tagger 30, the errors were highly localized. The following two tables illustrate that most of the tables were correctly tagged for all their components.
TABLE 9
______________________________________
Accuracy results for the Table Extractor
______________________________________
Total number of lines in 100 docs
6205 --
Total number of table and caption lines
1005 --
Total number of lines extracted by table extractor
1041 --
Total number of table and caption lines missed
18 1.8%
total number of extraneous lines extracted
54 5.4%
______________________________________
Retrieval Results The system and method of the present invention was evaluated using two types of queries--natural language query and structured query (with field retrieval on caption and table fields). Query Set I contained 25 natural language queries. The relevance judgment set for each query was obtained by looking at the top 100 tables retrieved for that query. Query Set II contained 25 structured queries derived by manually reformulating the natural language queries from Set I. Table 11 shows the precision statistics generated for each Query Set. The precision measure used to assess the performance of the Query Sets is defined as: ##EQU1## It can be seen from Table 11 that the structured queries perform significantly better than the natural language queries. The average precision values improved by 20.9%. More importantly, it can be seen that the structured queries do even better in the top 20 tables. This strongly indicates that they are very good for interactive querying where the users are usually interested in the top 15 to 20 documents.
TABLE 11
______________________________________
Precision Statistics
Precision (% change) - 25 queries
Tables Natural Language Query
Structured Query
______________________________________
5 72.0 84.0 (+16.7)
10 58.8 71.6 (+21.8)
15 48.0 62.4 (+30.0)
20 42.0 51.8 (+23.3)
30 33.3 39.1 (+17.4)
100 12.8 13.6 (+6.3)
200 6.4 6.8 (+6.3)
500 2.6 2.7 (+3.8)
1000 1.3 1.4 (+7.7)
Average 52.6 63.6 (+20.9)
______________________________________
The following queries illustrate the additional searching power that structured retrieval of tabular data gives to the user. For example, the following natural language query: Natural Language Query: Stock prices of oil companies can be reformulated as: Reformulated Query:
______________________________________
#wsum (6.0 (2.0 #field (CAPTION oil))
(2.0 #field (CAPTION stock))
(2.0 #field (CAPTION company))
(2.0 #field (CAPTION prices))).
______________________________________
The reformulated query should be interpreted as follows: (2.0 CAPTION oil) means that the occurrence of "oil" in the CAPTION field gets a weight of 2.0 on a scale of 6.0. When the queries were used to retrieve data from a database, the average precision value for the natural language query was 49.5% and for the structured query was 57.7%. The structured query performed 16.4% better than the natural language query. With the natural language query, many irrelevant tables related to oil matters other than stock prices and tables concerning stock prices of companies other than oil companies were retrieved. The structured query has the power to specify, that "oil" and "stocks" should get more weight if they occur in caption and that "company" and "price" are not as important. As a second example, the following natural language query: Natural Language Query: Export of computers was reformulated as: Reformulated Query:
______________________________________
#wsum (5.0 (2.0 #field (CAPTION exports))
(2.0 #field (CAPTION computers))
(1.0 #field (TABLE computers)).
______________________________________
The average precision value for the natural language query was 80.6% and for the structured query was 64.5%. The natural language query performed 20.1% better than the structured query. The natural language query performs better because the reformulated query gives more weight to computers. So anything related to computers such as computer parts, computer prices, and computer companies were all retrieved and there were a large number of computer related tables in the database. This query illustrates an important caveat about the structured queries. Although structured queries allow a user to more precisely express their information requirements, they are susceptible to bad reformulations leading to poor retrieval results. Also note that a structured query, at the very least, can be made to perform the same as the natural language query by giving equal weight to all the words occurring in all the fields. For example, the following structured query is the same as the natural language query from which it was derived:
______________________________________
#wsum (4.0 (1.0 #field (CAPTION exports))
(1.0 #field (CAPTION computers))
(1.0 #field (CAPTION computers))
(1.0 #field (CAPTION exports)).
______________________________________
The empirical results demonstrate that heuristic method-based table extraction and component tagging can be performed effectively and efficiently and that such structural decomposition can facilitate better representation of user's information needs and hence more effective retrieval of tables. Tagging and field indexing additional table components in addition to table and caption fields, such as column and row headings, may lead to more effective query formulation and retrieval. The weighting function could be modified in such a way that caption data is given the highest weighting, column and row headings get the next highest weight, and table entries get the least weighting. Other possibilities include slicing the table into columns and treating each column as a document. The column header and body content occurring together indicates more specificity and could be a source of multiple evidence for the corresponding table. For example, if the query is "china Exports Slippers" and we have a table with "China" and "slippers" occurring together in a column, this should get more weight than the case where "Romania" and "slippers" occur in one column and "China" occurs in another one. In this embodiment, co-occurrence in a column would indicates higher relevance and, therefore, a more desirable record to return as a result. Having described certain embodiments of the invention, it will now become apparent to one of ordinary skill in the art that other embodiments incorporating the concepts of the invention may be used. Therefore, the invention should not be limited to certain embodiments, but rather should be limited only by the spirit and scope of the following claims.
|
Same subclass Same class Consider this |
||||||||||