[ https://issues.apache.org/jira/browse/LUCENE-8015?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16277214#comment-16277214 ]
Adrien Grand commented on LUCENE-8015: -------------------------------------- I have been looking into the following model G failure. {noformat} 7.0E-45 = score(DFRSimilarity, doc=0, freq=0.99999994), computed from: 1.4E-45 = boost 3.09640771E16 = NormalizationH1, computed from: 0.99999994 = tf 1.61490253E9 = avgFieldLength 112.0 = len 9.2892231E16 = BasicModelG, computed from: 12.0 = numberOfDocuments 1.0 = totalTermFreq 4.8443234E-17 = AfterEffectB, computed from: 3.09640771E16 = tfn 1.0 = totalTermFreq 1.0 = docFreq 5.6E-45 = score(DFRSimilarity, doc=0, freq=1.0), computed from: 1.4E-45 = boost 3.09640792E16 = NormalizationH1, computed from: 1.0 = tf 1.61490253E9 = avgFieldLength 112.0 = len 9.289224E16 = BasicModelG, computed from: 12.0 = numberOfDocuments 1.0 = totalTermFreq 4.844323E-17 = AfterEffectB, computed from: 3.09640792E16 = tfn 1.0 = totalTermFreq 1.0 = docFreq DFR GB1 field="field",maxDoc=46519,docCount=12,sumTotalTermFreq=19378830951,sumDocFreq=19378830951 term="term",docFreq=1,totalTermFreq=1 norm=59 (doc length ~ 112) freq=1.0 NOTE: reproduce with: ant test -Dtestcase=TestBasicModelG -Dtests.method=testRandomScoring -Dtests.seed=3C22B051C61EEC84 -Dtests.locale=cs-CZ -Dtests.timezone=Atlantic/Madeira -Dtests.asserts=true -Dtests.file.encoding=UTF-8 {noformat} In short, the scoring formula here looks like {{(A + B * tfn) * (C / (tfn + 1))}} where A, B and C are constants. This function increases when tfn increases when B > A, which is always the case. The problem is that tfn is so large (ulp(tfn) = 4) , that {{tfn+1}} always returns {{tfn}} and {{A + B * tfn}} always returns the same as {{B * tfn}}. So when tfn gets high, the formula is effectively {{(B * tfn) * (C / tfn)}}. This is a constant, but since we compute the left and right parts independently, this might decrease when tfn increases about half the time. Even though I triggered it with BasicModelG, I suspect it affects almost all DFRSimilarity impls. > TestBasicModelIne.testRandomScoring failure > ------------------------------------------- > > Key: LUCENE-8015 > URL: https://issues.apache.org/jira/browse/LUCENE-8015 > Project: Lucene - Core > Issue Type: Task > Reporter: Adrien Grand > Attachments: LUCENE-8015_test_fangs.patch > > > reproduce with: ant test -Dtestcase=TestBasicModelIne > -Dtests.method=testRandomScoring -Dtests.seed=86E85958B1183E93 > -Dtests.slow=true -Dtests.locale=vi-VN -Dtests.timezone=Pacific/Tongatapu > -Dtests.asserts=true -Dtests.file.encoding=UTF8 -- This message was sent by Atlassian JIRA (v6.4.14#64029) --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org