ContentsCopyright © 1994 Harry Hahne
Contents
IntroductionIn recent years Bible-search software has become readily accessible to the average scholar. Unfortunately, using a computer to search the Bible may lend unwarranted credibility to research. Computer-assisted biblical research is subject to the same errors as traditional research methods and opens up new potential sources of error. Comparative tests of several popular Bible-search programs GRAMCORD, Bible Windows, Bible Works and TheWord)[1] show that the same grammatical searches often produce radically different results.
These variations are due to several factors:
Differences in the Underlying Biblical Texts
1. Grammatically Tagged New Testament TextsHowever, the grammatical tagging systems used for Greek and Hebrew texts vary significantly. A tagged text attaches parsing, lemmas (dictionary forms) and sometimes word definitions to each Greek and Hebrew word in the biblical text. This allows sophisticated stylistic and grammatical searches that would be impossible with the biblical text alone.
The selection of grammatical tags is based on subtle and often unstated assumptions. Although the function of a Greek word is indicated largely by its spelling (its "morphology"), at times the function of a word must be determined by its relation to the context. There is always a tension between purely morphological analysis based on word forms and a more functional analysis based on the interaction of a word with other words in the sentence. Grammatical tagging schemes range along a spectrum from formal (morphological) to functional classifications. No scheme for classifying Greek words is purely formal or purely functional, since the function of a word is determined both by its morphology and its relation to the context. However, the more a tagging system tends toward the functional end of the spectrum, the more subjective the classifications become.
For example, Bauer's lexicon classifies óu as an adverb of place However, in 5 instances (Mt 18:20; Rom 4:15; 5:20; 1 Cor 16:6; 2 Cor 3:17), the Friberg text, which is a largely functional system, classifies it as a conjunction, based on the nuances of the word in the context. A user would need to be aware of such functional classifications in order to find every occurrence of a particular part of speech.
There are four major morphologically-tagged Greek NT texts commonly used in Bible-search software:[4]
The print edition of the Friberg text has an appendix outlining the criteria used for the tags.[8] This is helpful when using BWorks, which stays very close to the original Friberg text. However, in some instances TheWord deviates from the Friberg tags without documenting the changes. BWorks and TheWord only use one of Friberg's multiple classifications of ambiguous words, without documenting which classification was chosen.
The following chart shows the "family tree" of various texts and the programs which use these texts:
2. Database Text and Classification ErrorsErrors can fall into three classes: (1) errors in the basic biblical text; (2) incorrect lemmas attached to words in the text; and (3) incorrect grammatical classifications.
Errors in the biblical texts are rare, since the texts often derive from machine-readable texts used for typesetting the print editions. However, incorrect lemmas and grammatical tags are more common. For example, in Jn 15:26, the Friberg 1 text claims `on comes from eimi. This error appears in both Bwin 2 and BWorks, but it is fixed in TheWord. BWin 3, which uses the Friberg 2 text, parses it correctly. James K. Tauber has collected a list of hundreds of errors in the CCAT tags, mostly in the lemmas. He suggests that many of the errors appear to be due to automatic parsing and lemmatization.[9]
3. Functional and Unusual ClassificationsIn GRAMCORD, many foreign words such as hosanna are classified as interjections. However, foreign proper nouns are classified as nouns and parsed by function in context. By contrast, BWin and BWorks classify hosanna as a particle.
Conjunctions and particles are particularly difficult words to classify. A beginning user might miss many occurrences of kai if he only searches for the word as a conjunction. Since kai also functions as an adverb in some cases, most programs will sometimes classify it as an adverb. However, as the following chart shows, the classification choices in individual instances vary considerably:
Program: Conjunction: Adverb: No subclass: Copulative: Correlative: GRAMCORD 8049 words 187 words 656 words BWorks 8214 words 801 words 4727 verses 733 verses Bwin 750 words (?) 750 words (?) TheWord 5126 verses 753 verses
Bwin was unable to report the total number of occurrences of kai, because it only allows 750 matches in a search. Since it is hard to predict how a program will classify the word in any given passage, the safest approach is to search for all possible classifications and manually eliminate invalid matches. The GRAMCORD manual documents how many times each word is classified as a conjunction, particle or adverb, which makes it easier to define searches that will find all occurrences of such words.
Since the Friberg text (BWorks and TheWord) attempts to classify many words by function based on discourse analysis, some classifications may be surprising to users. Friberg 1 uses the category of "substantive adjective" to refer to adjectives which are used as nouns in context. For example, agathos ("good") is classified as a substantive adjective in Mt 5:45 ("he makes the sun shine on the evil and the good). This type of classification affects 4131 occurrences of 1068 words in 3009 verses![10] While adjectives can certainly function as substantives, the term "substantive adjective" is not a part of speech used by most Greek grammars. It would be easy for a user to accidentally miss many important occurrences of adjectives unless he searches both for "adjectives" and "substantive adjectives". The Friberg 2 text eliminates the substantive adjective classification, but it introduces other surprising functional classifications. For example, in most cases Friberg 2 classifies relative pronouns as adjectives, with an adjective subtype of "relative."
Functional classifications such as those frequently used in Friberg's text are more subjective than formal classifications. Their value depends largely on the accuracy of the classifier's interpretation of the text. While they appear to be objective raw data, in fact they contain the prior conclusions of another researcher, which tends to skew the search results to fit the classifier's own viewpoint.
4. Treatment of Classification AmbiguitiesEven the strictest formal classification method must classify certain words by function in context, since the morphology of these words is inconclusive.
While in most cases the meaning is clear in the context, in some instances the grammatical classification is subject to scholarly debate. For example, the gender of ponhrou could be either neuter or masculine. In Mt 6:13 the meaning is debated: Does the Lord's Prayer ask for deliverance from "evil" (neuter) or "the evil one" (masculine)? Since BWin 2, GRAMCORD and Accordance classify ponhrou in Mt 6:13 as neuter, a search for masculine adjectives will not find the verse. By contrast, TheWord and BWorks classify the word as masculine and do not allow the word to be found in a search for masculine adjectives! Only BWin 3 acknowledges both possible parsings and allows the word to be found with either search.
Bible-search programs would be more useful if they marked such words as ambiguous and allowed searching on the multiple classifications. The print version of the Friberg text includes multiple classifications in many instances. However, at this time only BWin 3 allows searching on Friberg's multiple classifications. Although BWorks and TheWord both remove the multiple parsings in Friberg 1, the documentation does not explain the criteria used to make these choices.
GRAMCORD makes a good attempt at handling ambiguous classifications. In many cases, it tags words in multiple ways and flags the ambiguous classification in the resulting concordance. The documentation lists all ambiguous classifications which are used.[11] However, even GRAMCORD could be improved in this area. For example, it does not include the ambiguous classification of ponhrou in Mt 6:13.
Differing Capabilities of Search SoftwareWhile on the surface most Bible-search programs appear to allow similar searches, there is considerable variation in the search capabilities of programs. Many programs lack the sophistication to perform finely tuned searches. Further, some of the hidden or poorly documented assumptions in the search engines can produce surprising results.
1. Use of WildcardsWildcards are symbols that indicate that any letter or letters will be accepted at a certain point in a word. Thus a search for "apo*" will find apoluw, apodidwmi and other words which begin with "apo".
Some programs place a limit on the number of words that can match wildcards. TheWord allows a maximum of 300 words to match a wildcard and may not warn if this limit is exceeded. Logos 1.6 only allows 32 words to match a wildcard, which is too restrictive for many useful searches.
Most programs (e.g. TheWord, GRAMCORD, BWorks, Logos) assume that the search expression includes the whole word, unless wildcards are explicitly included. However BWin uses full word searches for grammatical searches and double wildcard searches for word and phrase searches. (In a double wildcard search, the search letters can be found anywhere within a word.) This inconsistent behavior in BWin can easily confuse users and result in erroneous searches.
2. Limits on the Number of MatchesBWin has an undocumented limit of 750 matches per search. Since there is no error message that warns that the maximum number of matches has been exceeded, this can lead to misleading conclusions.
3. Different Ways of Reporting StatisticsTheWord reports matches in terms of the number of verses which contain the desired construction. BWin and GRAMCORD report the number of occurrences of the desired construction. BWorks reports both the number of occurrences and verses.
4. Word Order Sensitivity.For some searches, word order is very important. For example, a search for substantival adjectives should find all occurrences of an article in agreement with an adjective only when the article appears just prior to the adjective, not after the adjective. In other cases, it is important to find all permutations of word order. For example, a search for genitive absolutes should allow either the genitive noun or the participle to appear first.
Programs differ in the importance they place on word order in search expressions.
GRAMCORD requires an exact match of the order of the search elements. However,
searches can be defined that find several combinations of word order in one pass
and distinguish them in the resulting concordance. BWorks and TheWord do not
distinguish the word order of search elements. This can result in many false
matches. For example, a search for
5. Duplication of Search Terms
Many grammatical constructions require that the same search term appear more than
once (e.g.
6. Exclude Intervening TermsFalse matches can frequently be eliminated by specifying terms that must not appear between search elements. For example, a future perfect periphrastic construction requires a future tense of eimi and the perfect participle of another verb in the same clause. Since it is highly unlikely that a finite verb will occur between these two search terms, the search can be improved if it specifies that no finite verb can intervene.
GRAMCORD and Accordance allow multiple intervening exclusion and inclusion terms. They can specify words and parts of speech that may occur between search terms as well as words and parts of speech that may not occur between search terms. Other programs do not include a true "exclude intervening term" option. At first glance it would appear that the "and not" Boolean operator which is available in BWin and TheWord would accomplish the same thing. However, the "and not" operator defines what a search term may not be, not the types of words that cannot appear between search terms. Thus this feature can produce undesired interactions between the search terms.
The following chart illustrates the effect of excluding intervening terms in a search for future perfect periphrastics:
Source: Matches: Invalid: Missing:
Nigel Turner, Syntax, p. 89[12] 6
BWorks 13[13] 9 0
GRAMCORD
Not exclude intervening 12 6 0
verbs
Exclude intervening finite 8 2 0
verbs
BWin
Not exclude intervening 9 3 0
verbs
Use "and not" indicative 0 0 6
verbs
7. Proximity of Multiple Search TermsMany grammatical constructions require that two or more words be in close proximity, though not necessarily side by side. A search program should allow restriction of search expressions to a definable number of words.
GRAMCORD allows the user to specify that up to 200 words span from beginning to end of a construction and BWin allows specifying up to 20 words. By default both programs assume that all elements in a search expression are juxtaposed. If this number is not set appropriately many valid examples of a construction will be missed.
BWorks and Logos have less flexibility than either of these programs. By default, all words must appear somewhere in the same verse. The user can specify a maximum number of verses in which to find the search terms. This is far less valuable for grammatical searches than a limit by number of words, though it can have value for discourse-level research.
8. Search Boundaries for Multiple Search TermsFor grammatical searches it is more valuable to set the search boundaries by sentences or clauses than by verses. A program based on verse boundaries would have difficulty with sentences that span several verses (e.g. Eph. 1:3-14). For discourse analysis, search boundaries should be set at the paragraph, chapter or book level. An ideal program would allow setting search boundaries by clause, sentence, a specific number of verses, paragraph, chapter, book. It would also allow the option of stopping at or ignoring various types of punctuation marks.
Accordance allows boundaries to be clause, sentence, verse, paragraph, chapter or book. TheWord allows boundaries to be verse, paragraph, chapter or book. Most other programs are more restricted. BWin does not allow specifying boundaries, though it will cross verse boundaries if the word proximity is set high enough. BWorks and Logos use verses as boundaries, but search expressions can cross verse boundaries if the proximity is set to 2 verses. GRAMCORD uses the sentence as a boundary, so it is more likely than a verse-oriented program to find all occurrences of a grammatical construction.
Programs differ in how they handle a conflict between the number of words in
proximity and the search boundary. BWin will cross verse boundaries in an effort
to compare the specified number of words. GRAMCORD will never cross a full stop
(period, semi-colon (raised dot) or question mark), regardless of the maximum
number of words allowed in the proximity. This subtle difference can produce
significantly different search results. GRAMCORD misses 12 examples of
9. Forced Agreement of Grammatical FeaturesMany grammatical constructions require either that certain grammatical features agree or not agree. For example, a genitive absolute requires a clause with a genitive noun and a genitive participle that agree in gender and number. If agreement cannot be required between search elements, many false matches must be manually removed.
Since BWorks and Logos do not allow specifying agreement of grammatical features, these programs make it difficult to find genitive absolutes without substantial manual labor. BWin allows specifying agreement, but has no way of limiting the agreement to specific search terms. Thus a search for genitive absolutes is quite simple with Bwin, but it is difficult to find constructions in which individual pairs of search terms agree. For example, it would be difficult to find the common Greek construction "article1 article2 noun2 noun1", where article1 and noun1 agree with each other and article2 and noun2 are genitive and agree with each other (e.g. ó tou tektonos úios in Mt 13:55). GRAMCORD and Accordance allow great flexibility in agreement. Any grammatical feature of selected pairs of words can be required or forbidden to agree. As a result, these programs can find very complex grammatical constructions with relatively few false matches.
10. Sensitivity to Diacritical MarksTheWord requires that Greek accents and breathing marks be entered as they appear in the biblical text. This makes entry of search expressions tedious and error prone. It also results in missed matches, where the context changes the accents.
On the other hand, required entry of breathing marks is desirable, since otherwise it is difficult to distinguish similar word pairs such as eis and éis. BWin includes breathing marks in the word pick list for grammatical searches and gives the option of including them in word searches. If they are not included in word searches, breathing marks are ignored. With GRAMCORD ambiguities can be resolved by specifying the desired part of speech. BWorks has no way to easily distinguish óu (an adverb of place) from ou (the negative particle), since it classifies both as adverbs.
11. Comparative Results in a Typical Search
A search for
Source: Matches: Invalid: Duplicate: Missing:
Nigel Turner, Syntax, 110
p. 332
GRAMCORD
Not exclude inter- 112 0 14 12
vening de
Exclude intervening de 98 0 0 12
BWorks
Context of 1 verse 104 11 0 17
Context of 2 verses 157 47 0 0
BWin
Word search mode 134 27 0 3
Grammatical search mode 97 0 0 13
TheWord
Context of 1 verse 97 0 0 13
12. Summary of Search CapabilitiesThe following table summarizes the search capabilities of several programs:
Feature: GRAMCORD: BWorks: BWin: TheWord:
Gram. Word
Search Search
Wildcards: optional optional none implicit optional
Match limit: unlimited unlimited 750 750 unlimited
Statistics
reported: occurrences occurrences occur. occur. verses
and verses
Word order
sensitivity: Yes Yes Yes No No
Allow duplicate
terms: Yes No Yes No Yes
Exclude interven-
ing terms: Yes No No No No
Proximity:
Type: Words Verses Words Words Verse,
para-
graph,
chapter,
book
Limit: 200 Unlimited 20 20 1
Boundary:
Type: Full stop Specified None None Verse,
(period, number of para-
question verses graph,
mark, chapter,
semicolon book
or weak stop
(comma, colon)
Priority: Boundary Proximity Proximity Proximity Proximity
Agreement of Any None All search None None
grammatical combination terms
features: of search
Diacritical
marks: Ignored Ignored Required Optional Required
Common User ErrorsEven the most powerful search software and carefully tagged database can produce misleading results if the user does not use the software correctly. There are several common sources of error when using Bible search software.
1. Lack of Understanding of the Program
A user may not understand the assumptions used in tagging the text or the limit-
ations of the search engine. For example, if a user does not realize that word
searches are insensitive to word order in BWin word searches, he may be surprised
to discover the number of false matches in the
Unfortunately, many important search assumptions and database tagging guidelines are not documented for most programs. For this reason, it is wise to experiment with a search that has known results before attempting searches with unknown results.
2. Lack of Understanding of the Grammatical ConstructionThe user may not understand the grammatical construction well enough to formulate a search accurately. When searching for future perfect periphrastics, it is not enough to find future forms of eimi within a certain number of words of a perfect participle. The two words must function together as a single grammatical unit in the sentence. Restricting the search to a single clause and excluding intervening finite verbs will help reduce the number of false matches.
It is wise to read about a construction in conventional grammar books before attempting a computerized search. This will enable the search expression to be formulated more precisely and provide some guidelines for the reasonableness of the search results.
3. Failure to Search for all Permutations of the ConstructionIt is easy to assume erroneously that a single search has found all examples of a construction. A thorough search must consider all valid orders of the search terms. For example, a search for genitive absolutes must look for constructions with the genitive noun first as well as constructions with the genitive participle first. GRAMCORD allows a single search to include many different search expressions and flags the resulting concordance according to the construction category. This reduces the need for multiple searches.
A thorough search must also find constructions with functionally equivalent parts of speech. For example, many constructions which call for a noun would be valid with a substantival participle or substantival adjective.
4. Failure to Manually Eliminate False MatchesEven the best software will sometimes produce false matches which must be manually eliminated. Programs such as GRAMCORD which can exclude intervening terms produce fewer false matches than programs which cannot. For example, in a search for future perfect periphrastics, GRAMCORD reports only 8 matches instead of 12 matches if the search excludes intervening finite verbs. However, Lk 1:45 and 6:40 can only be eliminated by hand, since the fact that the participle functions as a substantive is only indicated by context.
ConclusionsThese considerations lead to several suggestions for users of Bible-search software:
Go to Top
Go to Bible Software Web Page
Footnotes1. The following programs were used in researching this paper: Bible Windows 2.5 and 3.0, Bible Works 2.2.2, GRAMCORD 4.04, TheWord 3.05, Logos 1.6B. The results for Logos are usually not reported, since its grammatical search capabilities are too limited for scholarly research. For a review of these programs see Harry Hahne, "High-Tech Bible Study: PC Bible Programs with a Graphical User Interface" Computer-Assisted Research Forum, 1, no. 3 (Spring/Summer 1993): 7-17 and idem, "PC Bible-Search Software for Scholarly Research" ARC: The Journal of the Faculty of Religious Studies, McGill University 22 (1994): 109-125. Back to Text
2. BWin = Bible Windows. If no version number is specified, the statement applies to both version 2 and version 3. BWorks = Bible Works. Back to Text
3. Accordance is a sophisticated Macintosh grammatical-search program that uses the same database as the DOS-based GRAMCORD program. It is available from the GRAMCORD Institute. Back to Text
4. Texts based on a Strong's numbering system, such as Logos are not included. Although Logos uses largely formal morphological tags, the tags are based on the Textus Receptus, even when the NA26 Greek text is displayed. Thus a search for future participles with Logos finds 17 matches, while programs which base their grammatical tags on the UBS3 or NA26 text find 12 matches. Back to Text
5. Project GRAMCORD began in the late 1970's under the direction of Paul A. Miller. The grammatical tags were primarily selected by James L. Boyer, then Chairman of the Department of NT and Greek at Grace Theological Seminary. Back to Text
6. The tagged text was originally created by Barbara and Timothy Friberg to provide raw data for Timothy's linguistics dissertation on NT Greek word order. Back to Text
7. This text is one of several biblical and classical texts created by the Center for Computer Analysis of Texts (CCAT) at the University of Pennsylvania under the direction of Robert Kraft. Back to Text
8. Barbara and Timothy Friberg, eds., Analytical Greek New Testament (Grand Rapids: Baker, 1981). The second edition is in press. Back to Text
9. James K. Tauber (jtauber@tartarus.uwa.edu.au) is a student of Linguistics at the University of Western Australia. He can supply a patch file which corrects many CCAT text errors. Unfortunately, it will not work with the versions of the database used by BWin, BWorks or TheWord. Back to Text
10. According to the statistics reported by BWorks. Back to Text
11. GRAMCORD Users Guide, pp. 77-80. Back to Text
12. James Hope Moulton, A Grammar of New Testament Greek, vol. 3, Syntax, by Nigel Turner (Edinburgh: T & T Clark, 1963), p. 89. Back to Text
13. Since BWorks reports the number of matching verses, not constructions, it does not indicate that more than one matching construction sometimes occurs in the same verse. Back to Text
Go to Top
Go to Bible Software Web Page
Harry Hahne
Written November 24, 1994. HTML updated April 27, 1995