Alistair Moffat

Inverted files for text search engines

The technology underlying text search engines has advanced dramatically in the past decade. The development of a family of new index representations has led to a wide range of innovations in index storage, index construction, and query evaluation. While some of these developments have been consolidated in textbooks, many specific techniques are not widely known or the textbook descriptions are out of date. In this tutorial, we introduce the key techniques in the area, describing both a core implementation and how the core can be enhanced through a range of extensions. We conclude with a comprehensive bibliography of text indexing literature.

Implementing the PPM data compression scheme

The prediction by partial matching (PPM) data compression algorithm developed by J. Cleary and I. Witten (1984) is capable of very high compression rates, encoding English text in as little as 2.2 b/character. It is shown that the estimates made by Cleary and Witten of the resources required to implement the scheme can be revised to allow for a tractable and useful implementation. In particular, a variant is described that encodes and decodes at over 4 kB/s on a small workstation and operates within a few hundred kilobytes of data space, but still obtains compression of about 2.4 b/character for English text. >

Self-indexing inverted files for fast text retrieval

Query-processing costs on large text databases are dominated by the need to retrieve and scan the inverted list of each query term. Retrieval time for inverted lists can be greatly reduced by the use of compression, but this adds to the CPU time required. Here we show that the CPU component of query response time for conjunctive Boolean queries and for informal ranked queries can be similarly reduced, at little cost in terms of storage, by the inclusion of an internal index in each compressed inverted list. This method has been applied in a retrieval system for a collection of nearly two million short documents. Our experimental results show that the self-indexing strategy adds less than 20% to the size of the compressed inverted file, which itself occupies less than 10% of the indexed text, yet can reduce processing time for Boolean queries of 5-10 terms to under one fifth of the previous cost. Similarly, ranked queries of 40-50 terms can be evaluated in as little as 25% of the previous time, with little or no loss of retrieval effectiveness.

Exploring the similarity space

Ranked queries are used to locate relevant documents in text databases. In a ranked query a list of terms is specified, then the documents that most closely match the query are returned---in decreasing order of similarity---as answers. Crucial to the efficacy of ranked querying is the use of a similarity heuristic, a mechanism that assigns a numeric score indicating how closely a document and the query match. In this note we explore and categorise a range of similarity heuristics described in the literature. We have implemented all of these measures in a structured way, and have carried out retrieval experiments with a substantial subset of these measures.Our purpose with this work is threefold: first, in enumerating the various measures in an orthogonal framework we make it straightforward for other researchers to describe and discuss similarity measures; second, by experimenting with a wide range of the measures, we hope to observe which features yield good retrieval behaviour in a variety of retrieval environments; and third, by describing our results so far, to gather feedback on the issues we have uncovered. We demonstrate that it is surprisingly difficult to identify which techniques work best, and comment on the experimental methodology required to support any claims as to the superiority of one method over another.

Rank-biased precision for measurement of retrieval effectiveness

A range of methods for measuring the effectiveness of information retrieval systems has been proposed. These are typically intended to provide a quantitative single-value summary of a document ranking relative to a query. However, many of these measures have failings. For example, recall is not well founded as a measure of satisfaction, since the user of an actual system cannot judge recall. Average precision is derived from recall, and suffers from the same problem. In addition, average precision lacks key stability properties that are needed for robust experiments. In this article, we introduce a new effectiveness metric, rank-biased precision, that avoids these problems. Rank-biased pre-cision is derived from a simple model of user behavior, is robust if answer rankings are extended to greater depths, and allows accurate quantification of experimental uncertainty, even when only partial relevance judgments are available.

Inverted files versus signature files for text indexing

Two well-known indexing methods are inverted files and signature files. We have undertaken a detailed comparison of these two approaches in the context of text indexing, paying particular attention to query evaluation speed and space requirements. We have examined their relative performance using both experimentation and a refined approach to modeling of signature files, and demonstrate that inverted files are distinctly superior to signature files. Not only can inverted files be used to evaluate typical queries in less time than can signature files, but inverted files require less space and provide greater functionality. Our results also show that a synthetic text database can provide a realistic indication of the behavior of an actual text database. The tools used to generate the synthetic database have been made publicly available

Offline dictionary-based compression

Dictionary-based modelling is the mechanism used in many practical compression schemes. We use the full message (or a large block of it) to infer a complete dictionary in advance, and include an explicit representation of the dictionary as part of the compressed message. Intuitively, the advantage of this offline approach is that with the benefit of having access to all of the message, it should be possible to optimize the choice of phrases so as to maximize compression performance. Indeed, we demonstrate that very good compression can be attained by an offline method without compromising the fast decoding that is a distinguishing characteristic of dictionary-based techniques. Several nontrivial sources of overhead, in terms of both computation resources required to perform the compression, and bits generated into the compressed message, have to be carefully managed as part of the offline process. To meet this challenge, we have developed a novel phrase derivation method and a compact dictionary encoding. In combination these two techniques produce the compression scheme RE-PAIR, which is highly efficient, particularly in decompression.Dictionary-based modeling is a mechanism used in many practical compression schemes. In most implementations of dictionary-based compression the encoder operates on-line, incrementally inferring its dictionary of available phrases from previous parts of the message. An alternative approach is to use the full message to infer a complete dictionary in advance, and include an explicit representation of the dictionary as part of the compressed message. In this investigation, we develop a compression scheme that is a combination of a simple but powerful phrase derivation method and a compact dictionary encoding. The scheme is highly efficient, particularly in decompression, and has characteristics that make it a favorable choice when compressed data is to be searched directly. We describe data structures and algorithms that allow our mechanism to operate in linear time and space.

Inverted Index Compression Using Word-Aligned Binary Codes

We examine index representation techniques for document-based inverted files, and present a mechanism for compressing them using word-aligned binary codes. The new approach allows extremely fast decoding of inverted lists during query processing, while providing compression rates better than other high-throughput representations. Results are given for several large text collections in support of these claims, both for compression effectiveness and query efficiency.

Vector-space ranking with effective early termination

Considerable research effort has been invested in improving the effectiveness of information retrieval systems. Techniques such as relevance feedback, thesaural expansion, and pivoting all provide better quality responses to queries when tested in standard evaluation frameworks. But such enhancements can add to the cost of evaluating queries. In this paper we consider the pragmatic issue of how to improve the cost-effectiveness of searching. We describe a new inverted file structure using quantized weights that provides superior retrieval effectiveness compared to conventional inverted file structures when early termination heuristics are employed. That is, we are able to reach similar effectiveness levels with less computational cost, and so provide a better cost/performance compromise than previous inverted file organisations.

Word-based text compression

The development of efficient algorithms to support arithmetic coding has meant that powerful models of text can now be used for data compression. Here the implementation of models based on recognizing and recording words is considered. Move‐to‐the‐front and several variable‐order Markov models have been tested with a number of different data structures, and first the decisions that went into the implementations are discussed and then experimental results are given that show English text being represented in under 2‐2 bits per character. Moreover the programs run at speeds comparable to other compression techniques, and are suited for practical use.