Information extraction from the web using a search engine Citation for published version (apa)
Yüklə 0,9 Mb. Pdf görüntüsü
|
|
a
by selecting the most applicable label in I g . 136 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0 5 10 15 20 25 30 35 40 45 50 precision k Precision for k-NN Artist Similarity linear sqrt baseline Figure 6.10. Precision for the sets of 1995 artists using the three ambiguity esti- mators. Musical Artists In this experiment, I 224 is again the set of all artist names in the list composed by Knees et al. [2004]. This list consists of 14 genres, each with 16 artists. To find the most appropriate genre for the artists in I 224 , the genres mentioned in the list are not all suitable for finding co-occurrences. For example, the term classical is ambiguous and Alternative Rock/Indie is an infrequent term. We there- fore manually rewrote the names of the genres into unambiguous ones (such as classical music) and added some synonyms. After collecting the numbers of co- occurrences of artists and genres, we summed up the scores of the co-occurrences for synonyms. In this way, for each artist b the number of co-occurrences with the terms Indie and Alternative Rock are added to the co-occurrences of b with the genre Alternative Rock/Indie. Although the absolute number of co-occurrences with Alternative Rock/Indie may increase using this approach, it is notable that we use a relative measure to determine the most applicable category per artist. Motivated by the results in [Schedl et al., 2005], for PCM we used the allintitle option in the artist categorization experiment. For PM we selected for the genre-artist relations the patterns in Table 6.12 from a list of patterns expressing this relation. For all three methods, we reuse the artist similarities computed in the previous experiments. In Table 6.13 the performance of the initial mappings can be found for the 6.4 Experimental Results 137 baseline using p lin using p sqrt Babylon Zoo: 1. Tool Juli Juli 2. Live Chumbawamba Chumbawamba 3. Fish Jamiroquai Jamiroquai 4. Juli Shakira Tool 5. Chumbawamba Sonic Youth Shakira 6. Play Right Said Fred Janet Jackson B12: 1. Tool Juli Juli 2. Live Carl Craig Carl Craig 3. Fish Jamiroquai Jamiroquai 4. Juli Autechre Tool 5. Play Shakira Autechre 6. Japan Speedy J. Shakira B. Springsteen: 1. Neil Young T. Petty & Heartbreakers Neil Young 2. U2 Tom Petty T. Petty & Heartbreakers 3. Bob Dylan The Afghan Wigs Tom Petty 4. Tom Petty Neil Young The Afghan Wigs 5. Tool Patti Smith Bob Dylan 6. The Afghan Wigs Robert Plant Patty Smith Tool: 1. Freefrom Mudvayne Mudvayne 2. Mudvayne Type O Negative Type O Negative 3. Racoon Hothouse Flowers Hothouse Flowers 4. Strauss Massive Attack Massive Attack 5. Type O Negative Nine Inch Nails Nine Inch Nails 6. Hothouse Flowers Dream Theater Dream Theater U2: 1. Hothouse Flowers Hothouse Flowers Hothouse Flowers 2. Radiohead Radiohead Radiohead 3. Sin´ead O’Connor Sin´ead O’Connor Sin´ead O’Connor 4. Madonna Coldplay Coldplay 5. Coldplay Bruce Springsteen Talking Heads 6. Elvis Presley Pearl Jam Bruce Springsteen Table 6.11. Examples of most related artists using the baseline method and the two alternatives p lin (3.3) and p sqrt (3.4). 138 [Genre] artists like [Artist] [Genre] artists such as [Artist] [Genre] artists for example [Artist] [Artist] and other [Genre] artists Table 6.12. Four of the patterns for the artist-genre relation. In the other patterns, artists is respectively replaced with acts , musicians and bands . 0.7 0.72 0.74 0.76 0.78 0.8 0.82 0.84 0.86 0.88 0.9 0 5 10 15 20 25 precision k artist-genre categorization dm pm pcm Figure 6.11. Precision for the categorization of the musical artists. three methods (k = 0). We were able to map all artists to a genre. Co-occurrences between genres and artists thus could be found using PCM , PM as well as DM . The latter performs best. With respect to the preliminary mapping, the method with the smallest amount of Google queries performs best. Using DM only few related artists can be found on the documents visited. In- creasing k hence does not effect the performance for the final mapping, as the lists of related artists are small (Figure 6.11). Contrary to especially PCM , large num- bers of k do not deteriorate the precision. The performance of the pattern-based method strongly improves by consider- ing related artists, the best performance is obtained for k = 8. All methods perform best for values of k between 5 and 13. The Rock n’ Roll artists proved to be the most problematic to categorize. The artists in the genres classical, blues and jazz were all correctly categorized with the best scoring settings. With the supervised music artist clustering method discussed in [Knees et al., 2004] a precision of 87% was obtained using complex machine learning techniques |