Query Answering Using Inverted Indexes

Transcription

1 Query Answering Using Inverted Indexes

2 Inverted Indexes Query Brutus AND Calpurnia J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 2

3 Document-at-a-time Evaluation The conceptually simplest query answering method Query J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 3

4 Algorithm Find posting lists Can be implemented efficiently by keeping the top-k list at anytime J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 4

5 Term-at-a-time Evaluation J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 5

6 Algorithm Compute scores on one term Can be implemented efficiently by keeping the top-k list at anytime J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 6

7 Comparison Memory usage The document-at-a-time only needs to maintain a priority queue R of a limited number of results The term-at-a-time needs to store the current scores for all documents Disk access The document-at-a-time needs more disk seeking and buffers for seeking since multiple lists are read in a synchronized way The term-at-a-time reads through each inverted list from start to end requiring minimal disk seeking and buffer J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 7

8 List Skipping Consider an inverted list of n bytes, if we add skip pointers after each c bytes, and the pointers are k bytes long each Reading the whole list: Θ(n) bytes Jumping through the list using the skip pointers: Θ(kn/c) = Θ(n) No asymptotic gain When c is large and k is small, it may gain in practice J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 8

9 Big Skips If c gets too large, the average performance drops Consider finding p postings in a list of n bytes There are n/c total intervals in the list Need to read kn/c bytes in skip pointers Need to read data in p intervals on average, assume that the postings we want are about halfway between two skip pointers read additional pc/2 bytes The total number of bytes to read: kn/c + pc/2 When n/c p, skipping does not help Most disks require a skip of at least 100,000 postings to gain in speedup Skipping is useful in reducing the amount of time spent on decoding compressed data and processing cached data J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 9

10 Computing Cosine Score J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 10

11 Efficient Scoring For a query q = w1 w2 The unit vector v (q) has only two nonzero components If query terms are not weighted, the nonzero components are equal to 2 / 2 = Generally, for any two documents d 1 and d 2 V ( q) v( d1) > V ( q) v( d2) if and only if v( q) v( d1) > v( q) v( d2) V ( q) v( d) is the weighted sum over all terms in query q, of the weights of those terms in d J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 11

12 Efficient Scoring Algorithm Using a heap, selecting top k answers can be done with 2J comparisons where J is the number of answers of nonzero scores J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 12

13 Approximate Top-K Retrieval Retrieve K documents that are likely to be among the K highest scoring documents Goal: lower down the query answering cost Cosine measure is also an approximation of information need Major cost: computing cosine similarities between the query and a large number of documents Approximation strategies Find a set A of documents that are contenders, where K < A «N, such that A is likely to have many documents with scores near those of the top K Return the top-k documents in A J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 13

14 Index Elimination For a multi-term query q, we only need to consider documents containing at least one of the query terms Only consider documents containing terms whose IDF exceeds a preset threshold Only check those discriminative words Benefit: the postings lists of low-idf terms are generally long (many are stop words) Only consider documents that contain many of the query terms J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 14

15 Champion Lists For each term t in the dictionary, precompute the top-r documents of the highest weights for t, where r is a preset parameter Set different r for different terms larger for rare terms and smaller for frequent terms Given a query q, let A be the union of the champion lists for each of the terms comprising q Compute cosine similarity only between q and those documents in A J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 15

16 Static Quality Scores and Ordering Different documents have different importance Example: how good are reviews on a web page? Modeled by a quality measure g(d) [0, 1] V ( q) V ( d) TotalScore ( q, d) = g( d) + V ( q) V ( d) Sort documents in posting lists in g(d) descending order Suppose g(1) = 0.25, g(2) = 0.5, and g(3) = 1 J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 16

17 Using Quality Score Ordering For a well-chosen value r, maintain for each term t a global champion list of the top-r documents with the highest value of g(d)+tfidf(t, d) At query time, only compute TotalScore for documents in the union of those global champion lists Maintain for each term t two posting lists High list: m documents with the highest TF values for t Low list: the other documents containing t Use high list only if at least K answers can be generated J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 17

18 Tiered Indexes Generalization of champion lists J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 18

19 Clustering and NN Search Clustering Pick N documents as leaders at random from the collection For each document that is not a leader (called a follower), compute its nearest leader Each cluster has N = N N followers Alternatively, a follower can be assigned to b 1 leaders Query answering as nearest neighbor search For a query q, find the leader L (or b 2 leaders) that is closest to q computing cosine similarities between q and N leaders The candidate set A contains the closest leader and the followers J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 19

20 Example J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 20

21 Putting All Together J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 21

22 Summary and To-Do-List Query evaluation Document-at-a-time versus term-at-a-time List skipping Efficient scoring Approximate top-k retrieval Index elimination, champion lists, quality score and ranking, clustering and nearest neighbor search Read Section 5.7 J. Pei: Information Retrieval and Web Search -- Query Answering Using Inverted Indexes 22