We describe evaluation experiments conducted by submitting retrieval runs for the monolingual German, French, English and Persian (Farsi) information retrieval tasks of the Ad-Hoc Track of the Cross-Language Evaluation Forum (CLEF) 2008. In the ad hoc retrieval tasks, the system was given 50 natural language queries, and the goal was to find all of the relevant records or documents (with high precision) in a particular document set. We conducted diagnostic experiments with different techniques for matching word variations, comparing the performance on the robust Generalized Success@10 measure and the non-robust mean average precision measure. The measures generally agreed on the mean benefits of morphological techniques such as decompounding and stemming, but generally disagreed on the blind feedback technique, though not all of the mean differences were statistically significant. Also, for each language, we submitted a sample of the first 10000 retrieved items to investigate the frequency of relevant items at deeper ranks than the official judging depth of 60. The results suggest that, on average, the percentage of relevant items assessed was less than 55% for each of German, French and English and less than 25% for Persian.
http://www.opentext.com/ Livelink ECM - eDOCS SearchServerTM is a toolkit for developing enterprise search and retrieval applications. The SearchServer kernel is also embedded in other components of the Livelink ECM - eDOCS Suite1.
1Livelink, Open TextTM and SearchServerTM are trademarks or registered trademarks of Open Text Corporation in the United States of America, Canada, the European Union and/or other countries. This list of trademarks is not exhaustive. Other trademarks, registered trademarks, product names, company names, brands and service names mentioned herein are property of Open Text Corporation or other respective owners.
SearchServer works in Unicode internally [
This paper describes experimental work with SearchServer for the task of finding relevant documents for natural language queries in various European languages using the CLEF 2008 AdHoc Track test collections. 2 2.1
The CLEF 2008 Ad-Hoc Track document sets consisted of tagged (XML-formatted) records or documents in 4 different languages: German, French, English and Persian (also known as Farsi). For German, French and English, the records were library catalog cards (bibliographic records describing publications archived by The European Library (TEL)). For Persian, the documents were newspaper articles (Hamshahri corpus of 1996-2002). Table 1 gives the collection sizes.
The CLEF organizers created 50 natural language “topics” (numbered 451-500 for German, French and English and 551-600 for Persian) and translated them into many languages. Sometimes topics are discarded for some languages because of a lack of relevant documents (though that did not happen this year). Table 1 gives the final number of topics for each language and their average number of relevant documents (along with the lowest and highest number of relevant documents of any topic). For more information on the CLEF test collections, see the track overview paper. 2.2
Our indexing approach was mostly the same as last year [
For some experiments, some stop words were excluded from indexing (e.g. words like “the”,
“by” and “of” in English). For our Persian experiments, our stop word list was based on Savoy’s
list [
By default, the SearchServer index supports both exact matching (after some Unicode-based normalizations, such as decompositions and conversion to upper-case) and morphological matching (e.g. inflections, derivations and compounds, depending on the linguistic component used). 2.3
We experimented with the SearchServer CONTAINS predicate. Our test application specified SearchSQL to perform a boolean-OR of the query words. For example, for German topic 451 whose Title was “R¨omisches Milita¨r in Britannien” (Roman Military in Britain), a corresponding SearchSQL query would be: SELECT RELEVANCE(’2:3’) AS REL, DOCNO FROM CLEF08DE WHERE FT_TEXT CONTAINS ’Ro¨misches’|’Milita¨r’|’in’|’Britannien’ ORDER BY REL DESC;
Most aspects of the SearchServer relevance value calculation are the same as described last year
[
For the diagnostic runs listed in Tables 2, the run names consist of a language code (“DE” for German, “EN” for English, “FA” for Persian, and “FR” for French) followed by one of the following labels: ² “none”: No linguistic variations from stemming were matched. Just the surface forms were searched on (after case-normalization). ² “lexstem”: Same as “none” except that linguistic variations from stemming were matched.
The lexicon-based inflectional stemmer in SearchServer was used. For German, this stemmer
includes decompounding.
² “algstem”: Same as “lexstem” except that an algorithmic stemmer was used. For Persian, our
stemmer was ported from Savoy’s [
2.5
Traditionally, different retrieval measures have been used for “ad hoc” tasks, which seek relevant items for a topic, than for “known-item” tasks, which seek a particular known document. However, we argue that the known-item measures are not only applicable to ad hoc tasks, but that they are often preferable. For many ad hoc tasks, e.g. finding answer documents for questions, just one relevant item is needed. Also, the traditional ad hoc measures encourage retrieval of duplicate relevant documents, which does not correspond to user benefit.
The traditional known-item measures are very coarse, e.g. Success@10 is 1 or 0 for each topic,
while reciprocal rank cannot produce a value between 1.0 and 0.5. In 2005, we began investigating
a new measure, Generalized Success@10 (GS10) (introduced as “First Relevant Score” (FRS)
in [
Primary Recall Measures “Primary recall” is retrieval of the first relevant item for a topic. Primary recall measures include the following: ² Generalized Success@30 (GS30): For a topic, GS30 is 1:0241¡r where r is the rank of the first row for which a desired page is found, or zero if a desired page was not found. ² Generalized Success@10 (GS10): For a topic, GS10 is 1:081¡r where r is the rank of the first row for which a desired page is found, or zero if a desired page was not found. ² Success@n (S@n): For a topic, Success@n is 1 if a desired page is found in the first n rows, 0 otherwise. This paper lists Success@1 (S1) and Success@10 (S10) for all runs. ² Reciprocal Rank (RR): For a topic, RR is 1r where r is the rank of the first row for which a desired page is found, or zero if a desired page was not found. “Mean Reciprocal Rank” (MRR) is the mean of the reciprocal ranks over all the topics.
Interpretation of Generalized Success@n: GS30 and GS10 are estimates of the percentage of potential result list reading the system saved the user to get to the first relevant item, assuming that users are less and less likely to continue reading as they get deeper into the result list. Comparison of GS10 and Reciprocal Rank : Both GS10 and RR are 1.0 if a desired page is found at rank 1. At rank 2, GS10 is just 7 points lower (0.93), whereas RR is 50 points lower (0.50). At rank 3, GS10 is another 7 points lower (0.86), whereas RR is 17 points lower (0.33). At rank 10, GS10 is 0.50, whereas RR is 0.10. GS10 is greater than RR for ranks 2 to 52 and lower for ranks 53 and beyond.
Connection of GS10 to Success@10 : GS10 is considered a generalization of Success@10 because it rounds to 1 for r·10 and to 0 for r>10. (Similarly, GS30 is considered a generalization of Success@30 because it rounds to 1 for r·30 and to 0 for r>30.) 2.5.2
Secondary Recall Measures
“Secondary recall” is retrieval of the additional relevant items for a topic (after the first one).
Secondary recall measures place most of their weight on these additional relevant items.
² Precision@n: For a topic, “precision” is the percentage of retrieved documents which are
relevant. “Precision@n” is the precision after n documents have been retrieved. This paper
lists Precision@10 (P10) for all runs.
² Average Precision (AP): For a topic, AP is the average of the precision after each relevant
document is retrieved (using zero as the precision for relevant documents which are not
retrieved). By convention, AP is based on the first 1000 retrieved documents for the topic.
The score ranges from 0.0 (no relevants found) to 1.0 (all relevants found at the top of the
list). “Mean Average Precision” (MAP) is the mean of the average precision scores over all
of the topics (i.e. all topics are weighted equally).
² Geometric MAP (GMAP): GMAP (introduced in [
¢GS10
For example, in German topic 471 (Uhrenherstellung (Watchmaking)), the lexical stemming run substantially outscored the 4-gram run in the GenS@10 measure. The stemmer produced stems of ‘uhr’ and ‘herstellung’, with the ‘uhr’ (clock) stem getting higher weight from inverse document frequency, and a relevant record was retrieved at rank 4 from the ‘uhr’ stem matching subject terms of ‘Uhr’ and ‘Uhrmacher’. The 4-gram approach did not match either ‘Uhr’ or ‘Uhrmacher’ (e.g. the 4-gram ‘Uhre’ from the query word is not a 4-gram of ‘Uhr’ or ‘Uhrmacher’) and it put a lot of weight on the less specific ‘herstellung’ part of the query word (several 4-gram terms) and it did not retrieve a relevant record until rank 525.
One Persian topic scored much higher in both GenS@10 and MAP using 4-grams instead of the stemmer, namely topic 559 (best Fajr film). The reason though appears to be not from 4gramming finding better matches than the stemmer, but that the 4-gram mode did not use the stopword list. We see that topic 559 scored higher without stopping words (as per the “alg-all” line included in Table 5). The longest of the 3 Persian words in the topic title (we suspect the Persian word for ‘best’) was in the stopword list, perhaps inadvertently; we should investigate further. 4
For each language, we submitted 4 experimental runs in June 2008 for official assessment. In the identifiers (e.g. “otFA08tdnz”), ‘t’, ‘d’ and ‘n’ indicate that the Title, Description and Narrative field of the topic were used (respectively), and ‘e’ indicates that query expansion from blind feedback on the first 3 rows was used (weight of one-half on the original query, and one-sixth each on the 3 expanded rows). The ‘z’ code indicates that special sampling was done, as described below. From the Description and Narrative fields for most languages, instruction words such as ² “td”: Same as “t” except that the Description field of the topic was additionally used. ² “tde”: Same as “td” except that blind feedback (based on the first 3 rows of the “td” query) was used to expand the query. ² “tdn”: Same as “td” except that the Narrative field of the topic was additionally used. (This run was not submitted.) ² “tdz”: Depth-10000 sampling run based on the “td” run as described below (German, French and English only). ² “tdnz”: Depth-10000 sampling run based on the “tdn” run as described below (Persian only). Table 7 shows the impact of including the Description field on the GenS@10 and MAP measures (and for Persian, it also shows the impact of including the Narrative field). We see large impacts on individual topics in both directions. The only statistically significant mean differences were for the Persian MAP score, for which both the Description terms and Narrative terms were detrimental. ² “Precision”: Estimated precision of the depth range (R/(R+N+U+X)). ² “Wgt”: The weight of each sample point. The weight is equal to the difference in ranks between sample points, i.e. each sample point can be thought of as representing this number of rows, which is itself plus the preceding unsampled rows. ² “EstRel/Topic”: Estimated number of relevant items retrieved per topic for this depth range.
This is the Precision multiplied by the size of the depth range. Or equivalently, it is (R * Wgt) / 50.
Because each sample point is at the deep end of the range of rows it represents, the sampling should tend to underestimate precision for each depth range (assuming that precision tends to fall with depth, which appears to be the case for all 4 languages).
Table 13 shows the sums of the estimated number of relevant items per topic over all depth ranges in its first row. The official number of relevant items per topic for each language is listed in the second row. The final row of the table just divides the official number of relevant items by the estimated number in the first 10000 retrieved (e.g. for German, 32.7/59.9=55%). This number should tend to be an overestimate of the percentage of all relevant items that are judged (on average per topic) because there may be relevant items that were not matched by the query in the first 10000 rows.
However, the sampling was very coarse at the deeper ranks, e.g. for German, 1 relevant item out of 200 samples in the 6001-10000 range led to an estimate of 20 relevant items per topic in this range. If the sampling had turned up 0 or 2 relevant items, a minor difference, the estimate would have been 0 or 40 relevant items per topic in this range, leading to a substantially different sum (39.9 or 79.9 instead of 59.9). We should compute confidence intervals for these estimates, but have not yet done so. Also, there is a lot of variance across topics, which we have not yet analyzed.
These preliminary estimates of judging coverage for the CLEF 2008 collections (55% for
German, 52% for French, 53% for English, 25% for Persian) tend to be lower than the estimates we
produced for the CLEF 2007 collections last year [
For Persian, the topics appear to have been relatively broad (more relevant documents per topic on average) which led to the judging coverage being relatively shallow (based on the sampling experiment). It is not clear however whether these are factors in the unusual results we found for Persian (e.g. normally the Description and Narrative terms increase retrieval scores instead of decrease, and normally we see more impact from stemming on at least some individual topics).
The incompleteness results for German, French and English are similar to what [
Fortunately, [
algorithms. October Multilingual information retrieval resource