Generation of Synthetic Query Auto Completion Logs

Conference paper
Part of the Lecture Notes in Computer Science book series (LNCS, volume 12035)


Privacy concerns can prohibit research access to large-scale commercial query logs. Here we focus on generation of a synthetic log from a publicly available dataset, suitable for evaluation of query auto completion (QAC) systems. The synthetic log contains plausible string sequences reflecting how users enter their queries in a QAC interface. Properties that would influence experimental outcomes are compared between a synthetic log and a real QAC log through a set of side-by-side experiments, and confirm the applicability of the generated log for benchmarking the performance of QAC methods.

1 Introduction

Query auto completion (QAC) systems offer a list of completions while users enter queries in a search interface. Users can either submit one of the completions as their query, or advance their partial query by selecting a completion and then continuing to type [33]. A detailed QAC log capturing the sequence of partial queries, along with the completions presented and the user interactions with them, is required in order to evaluate a QAC system [37, 38]. However, concerns about the privacy of query logs and regulatory requirements such as GDPR mean that there is a need for alternative ways of obtaining logs for academic purposes.

Here we explore a framework for generating synthetic QAC logs, extending the work of Krishnan et al. [33], who suggest converting a QAC log to an abstracted format (an abstract QAC log) that records only the length of each partial query and the lengths of words used, minimizing privacy concerns but removing the possibility of performing evaluations on actual strings. Synthetic QAC log generation seeks to produce a list of plausible synthetic partial query sequences by mapping the word lengths from the abstract QAC log to strings from a publicly available dataset. An example of the proposed process is shown in Fig. 1. On the left are partial queries typed by a user. The abstracting process converts these to the digit sequences shown in the middle column, describing the strings but not their characters; and then the corresponding synthesized strings are shown at the right. Note that it is neither necessary nor sufficient for the synthetic log to contain semantically valid phrases. Comparison between the original and synthetic logs across a range of properties show that the synthetic log can be used to evaluate QAC system performance. Moreover, the synthetic log eliminates the privacy concerns associated with the original.
Fig. 1.

A sample QAC log entry, the corresponding abstract QAC log entry, and a synthetic QAC log entry generated from the string “john of gaunt”.

2 Background

A QAC system retrieves a candidate set matching the partial query \({\small \mathsf{{P}}}\), drawing from a target string collection, with strings in the target collection having an associated score. Query Auto Completion systems typically match \({\small \mathsf{{P}}}\) against past queries from a log; or, in the absence of logs, they can also be synthesized [14, 40]. Methods of ranking the candidates include static popularity [9], search context [9, 29], forecast popularity [15], personalized ranking parameters [15, 29, 42], and diversity [16]. It is also possible to choose an initial candidate set based on popularity and then apply a second ranking criteria to obtain the final strings [15, 16]. QAC implementation strategies vary based on how the partial query \({\small \mathsf{{P}}}\) is matched against the target strings [32]. A common approach is to use a trie [3, 4, 25, 27] to retrieve candidates that have \({\small \mathsf{{P}}}\) as a prefix; or inverted index-based approaches [10, 11, 23] that offer completions independent of the ordering of the words in the partial query. The functionality of a QAC system can be extended beyond character level matches by including contextual cues [11] or synonyms [12, 28]. Error-tolerant QAC approaches [17, 28, 36, 47] allow up to a fixed number of character mismatches to account for possible typing errors.

User interactions are a key factor in implementation and evaluation of QAC systems [26, 33, 38] and have been captured using a wide range of models [31, 32, 33, 37, 38, 43, 44]. In particular, users are not limited to entering single characters, and can alter the partial query by selecting a completion or deleting characters already entered. Until now, the test collections used to evaluate traditional search systems have been anonymized commercial logs [1, 19, 34, 48] or synthesized logs [5, 30, 45, 49]. QAC system evaluations have typically been performed over large publicly available string collections [10, 14, 23], with strings taken sequentially from left-to-right to generate partial query sequences [10]. However this approach does not account for the full range of possible interactions [32, 33]. In this work, we explore an approach to generation of synthetic partial query sequences that addresses this gap.

Terminology. A partial query \({\small \mathsf{{P}}}\) is the string currently displayed in the search box. An interaction updates \({\small \mathsf{{P}}}\) and results in loggable changes. A new conversation starts when the user begins a query, and continues until either explicitly terminated by the user or as a result of a session timeout. The last partial query from a conversation is referred to as \({{\small \mathsf{{FinalP}}}}\). A QAC log records a set of conversations as a sequence of partial queries. A surrogate log is a dictionary of strings that can be used as substitutes for final partial queries, with each such string having an associated score reflecting its popularity. For a string \({\small \mathsf{{T}}}\), the ordered tuple \(\langle {\left| {{{{\small \mathsf{{w}}}}_{1}}}\right| ,\left| {{{{\small \mathsf{{w}}}}_{2}}}\right| {\dots }\left| {{{{\small \mathsf{{w}}}}_{k}}}\right| }\rangle \) representing the lengths of its words \({{{\small \mathsf{{w}}}}_{1}}, {{{\small \mathsf{{w}}}}_{2}}, \dots {{{\small \mathsf{{w}}}}_{k}}\) in \({\small \mathsf{{T}}}\) is referred to as its signature. For each partial query \({\small \mathsf{{P}}}\), an abstract QAC log records only its signature and its length \(\left| {{\small \mathsf{{P}}}}\right| \), including whitespace. The signature of each \({{\small \mathsf{{FinalP}}}}\) in the abstract QAC log is referred to as \({\small \mathsf{{seedsig}}}\). The signature of a string in the surrogate log is its \({\small \mathsf{{targsig}}}\).

Problem Definition. For each conversation in the abstract QAC log, find a target string in the surrogate log with the same signature as the final partial query \({{\small \mathsf{{FinalP}}}}\). Then, starting from the first interaction in the conversation, apply the word lengths from the abstract QAC log to the target string in order to obtain a plausible partial query sequence. For example, consider the last conversation in Fig. 2, with its signature sequence \(\langle {1}\rangle , {\langle {2}\rangle }, {\langle {6}\rangle }, {\langle {6,3}\rangle }\). The \({\small \mathsf{{seedsig}}}\) for this conversation is \(\langle {6, 3}\rangle \). The string “coffee mug” in the surrogate log has the same signature and hence might be selected as a target string. Mapping the signature sequence from the conversation, we get the synthetic partial query sequence “c”, “co”, “cof”, “coffee”, and then “coffee mug”.
Fig. 2.

The synthetic pattern generation process. Signatures of \({{\small \mathsf{{FinalP}}}}\) from the abstract QAC log are used to form the \({\small \mathsf{{seedsig}}}\) list, and the signatures from surrogate log form \({\small \mathsf{{targsig}}}\) list. A match between the signature \(\langle {6,3}\rangle \) in the \({\small \mathsf{{seedsig}}}\) list and in the \({\small \mathsf{{targsig}}}\) list might correspond to the target string “coffee mug”.

3 Generation Process

Depending on the distribution of word lengths in the surrogate log, for every final partial query in the abstract QAC log, there may not be a string having the same signature. Moreover, strings are not entered by the users in the word order of the target collection. For instance, a user looking for the Wikipedia main page might enter the queries “wikipedia”, “main page wikipedia”, or “wiki”. We narrow down the possible ways of matching a \({\small \mathsf{{seedsig}}}\) with the list of signatures in \({\small \mathsf{{targsig}}}\) list to the following hierarchical modes:
  1. 1.

    Exact. The \({\small \mathsf{{targsig}}}\) is equal to the \({\small \mathsf{{seedsig}}}\). For example, the \({\small \mathsf{{seedsig}}}\) \(\langle {3, 3, 2}\rangle \) only matches with the target signature \(\langle {3, 3, 2}\rangle \). There might be zero, or multiple strings in the surrogate log that match.

  2. 2.

    Prefix. The \({\small \mathsf{{seedsig}}}\) is a prefix of the \({\small \mathsf{{targsig}}}\). For example, \({\small \mathsf{{seedsig}}}\) \(\langle {3, 3, 2}\rangle \) matches \(\langle {3, 3, 2, 4}\rangle \) and \(\langle {3, 3, 2, 4, 7}\rangle \), but not \(\langle {9, 3, 3, 2}\rangle \).

  3. 3.

    Match by Drop (MbD). The \({\small \mathsf{{seedsig}}}\) is an ordered subset of the \({\small \mathsf{{targsig}}}\), For example, \(\langle {3, 3, 2}\rangle \) matches \(\langle {3, 4, 3, 2}\rangle \) and \(\langle {4, 3, 1, 3, 6, 2}\rangle \), but not \(\langle {3, 4, 2, 3}\rangle \).

  4. 4.

    Bag of Numbers (BoN). Relaxing the ordering requirement, a BoN match occurs if the sets of word lengths in \({\small \mathsf{{seedsig}}}\) are a subset of those in \({\small \mathsf{{targsig}}}\). For example, \({\small \mathsf{{seedsig}}}\) \(\langle {3, 3, 2}\rangle \) matches \(\langle {2, 3, 3, 4}\rangle \) and \(\langle {5, 2, 3, 6, 3}\rangle \), but not \(\langle {3, 4, 2, 6}\rangle \). In a BoN match the target string’s words are reordered to match the seed signature ordering.

Locating Target Strings. The first step locates, for a given \({\small \mathsf{{seedsig}}}\), a set of matching signatures in the \({\small \mathsf{{targsig}}}\) list. The set of target candidates is maintained in lexicographically sorted order, so that the exact-match signatures for a given \({\small \mathsf{{seedsig}}}\) can be found via two binary searches, establishing a range \([{{{\small \mathsf{{r}}}}_{beg}},{{{\small \mathsf{{r}}}}_{end}})\). A similar process can be used to find the prefix-match range, which is a larger contiguous block in the lexicographically sorted array of target candidates.

If a target signature \({{{\small \mathsf{{S}}}}_{t}}\) matches with \({\small \mathsf{{seedsig}}}\) using MbD, any \({\small \mathsf{{targsig}}}\) having \({{{\small \mathsf{{S}}}}_{t}}\) as a prefix will also have an MbD match with \({\small \mathsf{{seedsig}}}\). For example, under MbD the \({\small \mathsf{{seedsig}}}\) \(\langle {3, 5}\rangle \) matches \(\langle {3,2,5}\rangle \). Then signatures \(\langle {3,2,5,4}\rangle \) and \(\langle {3,2,5,9,16}\rangle \) will also be a match because they have \(\langle {3,2,5}\rangle \) as a prefix. Using this property, once a matching signature \({{{\small \mathsf{{S}}}}_{t}}\) for MbD is found, we can add the prefix match range for \({{{\small \mathsf{{S}}}}_{t}}\) to the MbD range for the current \({\small \mathsf{{seedsig}}}\). In contrast to exact and prefix match ranges, MbD ranges may not form a continuous range over \({\small \mathsf{{SigList}}}\).

For BoN matching, the tokens in the signature are sorted to form a canonical representation. For example, \(\langle {3, 2, 4}\rangle \) becomes \(\langle {2,3,4}\rangle \). This list is sorted into lexicographical order. A BoN match between the canonical representations of a \({\small \mathsf{{seedsig}}}\) and a target signature \({{{{\small \mathsf{{S}}}}_{tc}}}\) can be verified by a linear scan over the tokens. The matching ranges for BoN in the modified \({\small \mathsf{{SigList}}}\) are calculated by finding prefix match ranges for each \({{{\small \mathsf{{S}}}}_{tc}}\) that matches with \({\small \mathsf{{seedsig}}}\) using BoN.

Handling Deletions. To include deletions, we assume that something different was initially typed (the replacement word) and was converted to the word from the target string after the deletions. In particular, suppose that the current conversation contains the deletion of a word \({{{\small \mathsf{{w}}}}_{k}}\) starting from the ith interaction and ending with the jth interaction. Then, for every interaction preceding the jth interaction, \({{{\small \mathsf{{w}}}}_{k}}\) is substituted by a replacement word constructed via a set of deletion heuristics. If character sequences are deleted and re-entered several times during the conversation, several replacement words will be required. The last replacement word should be close to the original word, and the penultimate replacement prior to it should similarly be a modification of the last replacement word. Table 1 gives an example of the partial query sequence generated by re-writing target strings when the conversation contains multiple deletions.
Table 1.

Deriving synthetic partial query sequence for a conversation containing multiple word deletions. The fourth column shows three target strings used.


\(\left| {{\small \mathsf{{P}}}}\right| \)


Modified target string

Synthetic \({\small \mathsf{{P}}}\)



\(\langle {5,6}\rangle \)

black mirror

black mirror



\(\langle {5,5}\rangle \)

black mirror

black mirro



\(\langle {5}\rangle \)

black mirror




\(\langle {5}\rangle \)

black hole




\(\langle {5,1}\rangle \)

black hole

black h



\(\langle {5,2}\rangle \)

black hole

black ho



\(\langle {5,1}\rangle \)

black hole

black h



\(\langle {5,2}\rangle \)

black hawk down

black ha



\(\langle {5,4,4}\rangle \)

black hawk down

black hawk down

Users delete characters either due to typing errors [6, 20] made within that word, or to switch to an entirely different word. Among various typing errors that can occur in any character level entry systems [22], the following error categories that are predominantly discussed in past research [6, 20, 21, 46] are included in the generation process.

  • Deletion. A deletion error occurs when the user initially missed out one of the characters in the word, for example, in the correction “acount\(\rightarrow \)account”. Deletion errors are more frequent at character repetitions, and occur more commonly at the beginning of a word [6]. To simulate a deletion error, if there is a deletion sequence from ith interaction to the jth interaction in which more than two characters of \({{{\small \mathsf{{w}}}}_{k}}\) are removed, then \({{{\small \mathsf{{w}}}}_{k}}\) from the \(j+1\)th interaction (after the deletions) is examined to see if the last two characters in \({{{\small \mathsf{{w}}}}_{k}}\) are repeating. The error is then simulated by deleting one of the repeating characters of \({{{\small \mathsf{{w}}}}_{k}}\) in the interactions prior to \(j+1\).

  • Insertion. These arise when an extra letter is initially typed, for example, the correction “sherriff\(\rightarrow \)sheriff”. Our experimental results show that only a low fraction of insertion errors are observed in a QAC log and for that reason, they are not included in the synthetic QAC log generation process.

  • Substitution. These arise when the user enters one of the neighboring keys instead of the intended key, for example, the correction “disturv\(\rightarrow \)disturb”. Single character deletions are assumed to be substitution errors. The substituted character is found via a probability distribution for mistyped keys around the current key.

  • Transposition. This is when the user swaps two characters in a word either not knowing the correct spelling or because of mis-ordered keystrokes, for example, the correction “wierd\(\rightarrow \)weird”. Deletions of length two are considered to be transposition errors and the last two characters of the word are swapped to get the replacement word.

In a QAC interface, typing errors may not be the only reason why users delete characters. Sometimes users replace certain parts of a well-formed partial query, for instance to get a different set of completions. In such cases, referred as multichar deletions, the corrected word will differ from the original word in two or more character positions. To allow for such cases, the deleted part of the word is replaced with another string starting with the word’s remaining prefix, but differing in the next character. For example, consider the word “hawk” in Table 1. A deletion chain removed all of the characters except the “h”; and hence (working backwards), “hawk” was replaced by an alternate word that starts with “h” but not “ha”, such as “hole”.
Fig. 3.

Framework for generating a synthetic QAC log. See the text for details.

Generation Framework. The overall framework for generation of a synthetic QAC log is illustrated in Fig. 3. The seed signatures are stored with a mapping from signatures to the list of \({\small \mathsf{{cid}}}\)s having the same signature. Signatures from the surrogate log are precomputed and stored in lexicographically sorted order (TS1). Additionally, the strings in the surrogate log are re-ordered based on TS1 ordering to obtain the permutation SL1. The permutations TS2 stores a sorted list of within sorted signatures to support BoN matching, and SL2 is the corresponding permutation of the surrogate log. To generate a synthetic QAC log entry, a signature is selected from \({\small \mathsf{{seedsig}}}\mapsto {\small \mathsf{{cid}}}\) list and a list of target signatures retrieved based on the four modes. Using the mappings SL1 and SL2, strings from the surrogate log corresponding to the target signatures are then obtained.

The strings are converted to target strings by aligning the word lengths with the \({\small \mathsf{{seedsig}}}\). A target string can be generated from multiple originals. If, say, \({\small \mathsf{{seedsig}}}\) is \(\langle {9,4}\rangle \), then “wikipedia main page” gives the target string “wikipedia page” using MbD; and another string “personal web page wikipedia” gives the same target using BoN mode. If the same target string is produced by more than one string from the surrogate log, then the one with lower score is discarded.

The signature sequences obtained from the file providing the \({\small \mathsf{{cid}}}\mapsto \) signature sequence mapping are then applied to the list of target strings, to obtain a synthetic partial query sequence for each conversation. Partial query sequences are generated for each \({\small \mathsf{{cid}}}\) from the current \({\small \mathsf{{seedsig}}}\mapsto {\small \mathsf{{cid}}}\) list until either the \({\small \mathsf{{cid}}}\) list or the target strings are exhausted. Finally, after the generation process is completed, the synthetic QAC log is re-ordered so that the conversations in the synthetic QAC log have one-to-one correspondence with the abstract QAC log.

4 Experiments

Datasets Used. The Wikipedia clickstream dataset,1 generated from Wikipedia request logs containing tuples of the form (\({referrer }\), \({resource }\), \({frequency }\)), is used to generate the surrogate log. The number of requests for an article (\({resource }\)) is the \({frequency }\). Data dumps from January to March 2019 were aggregated by updating the \({frequency }\) of each \({resource }\) by the mean \({frequency }\) over the dumps. We refer to the resulting dataset as Wiki-Clickstream and the synthetic QAC log generated from Wiki-Clickstream as Wiki-Synth.

A QAC log was formed by randomly sampling the logs from Bing2 QAC system over a period of one week from 13 August 2018. This Bing-QAC-2018 records the partial queries entered by the users and a unique \({\small \mathsf{{cid}}}\) for each conversation. The abstracted version of Bing-QAC-2018, referred to as Bing-Abs-QAC-2018, is generated by recording the signature and total length of each partial query along with a unique \({\small \mathsf{{cid}}}\) for each conversation. Note that Wiki-Synth was generated from Bing-Abs-QAC-2018 only (with no use of Bing-QAC-2018), and that the subsequent comparisons between Wiki-Synth and Bing-QAC-2018 were performed on secure Microsoft servers and in accordance with Microsoft privacy requirements.

Preprocessing. Conversations in Bing-Abs-QAC-2018 where an intermediate signature contained more words than the signature of the final partial query, or contained words that were longer than the corresponding words in the final partial query, were removed. The strings from Bing-QAC-2018 were transliterated from UTF-8 to ASCII encoding using Unidecode3 Python package. Conversations containing non-converting characters in any partial query were discarded. The strings from Wiki-Clickstream were transliterated in the same way and non-converting strings were removed. The conversion from UTF-8 to ASCII resulted in the loss of 0.26 million conversations from Bing-QAC-2018 and 0.02 million strings from Wiki-Clickstream. After the preprocessing, there were 1.44 million conversations in Bing-Abs-QAC-2018 and 5.11 million strings in Wiki-Clickstream. In Bing-Abs-QAC-2018, 7.53% of signatures were unique, of which 2.26% had no matches in Wiki-Clickstream. A total of 0.23 million conversations from Bing-Abs-QAC-2018 were not included in Wiki-Synth because there were not enough matching strings in Wiki-Clickstream to map their signature sequence.

Sampling of Target Strings. The frequency of search queries has been found to exhibit a power law distribution [8] with the probability distribution \(p(x) \propto x^{-\alpha }\). Therefore, while generating a synthlog, the target strings need to be sampled to obtain a power law distribution based on their frequencies. Strings in the Wiki-Clickstream log have a score that tends to follow a power law distribution with \(\alpha = {3.09}\). Using this observation we sample target strings based on a weighted probability distribution over their scores, so that the probability of a string \({\small \mathsf{{T}}}\) given a seed signature \({\small \mathsf{{seedsig}}}\) is
$$\begin{aligned} {Prob }({\small \mathsf{{T}}}\mid {\small \mathsf{{seedsig}}}) = \frac{{score }({\small \mathsf{{T}}})}{\sum _{i=1}^{n} {score }({{{\small \mathsf{{T}}}}_{i}})} \end{aligned}$$
where \({\small \mathsf{{T}}}_1, {\small \mathsf{{T}}}_2, \dots , {\small \mathsf{{T}}}_n\) are the set of target strings retrieved for \({\small \mathsf{{seedsig}}}\). The frequency distribution of the resulting target strings used as the \({{\small \mathsf{{FinalP}}}}\) in Wiki-Synth is analyzed below, where we compare partial query frequencies.

Language Model for Finding Replacement Words. For deletion types other than multichar, the replacement word will be a modification of the original word based on the error category. For multichar deletions, the heuristics discussed above are used to find replacement words. The best replacement word based on contextual information among the candidate words that passes the heuristics is selected using a 4-gram language model (LM) trained on the title strings from Wiki-Clickstream. The language model is generated using KenLM [24],4 which is based on modified Kneser-Ney smoothing and provides fast model construction and querying. For example, an LM-based replacement for the target string “live queen”, is “live together” while a random replacement gives “live teufelshorner”; for the target string “web server”, a random replacement yields “web castelvetere” and an LM-based replacement gives “web content”. A similar character bigram model trained on the Microsoft spelling-correction dataset5 is used to find the most likely next mistyped character, given the previous character, to simulate substitution errors.

Comparison of Synthetic QAC Logs with QAC Logs. A synthetic QAC log should have some properties similar to the QAC log, and these can be used to validate the generation process. A comparison of these properties is given in Table 2. Other properties are desirable if a synthetic QAC log is to be a substitute in experiments for a QAC log. While comparing these properties, the partial queries from both the logs are treated as the test queries that will be queried against a collection of strings acting as the test collection. Extending our assumption that a user’s goal was to enter \({{\small \mathsf{{FinalP}}}}\), we claim that these strings could get eventually indexed by the QAC system and subsequently be queried. Thus, in our experiments, the list of \({{\small \mathsf{{FinalP}}}}\) strings are considered as a representative sample of a larger hypothetical test collection.
Table 2.

Basic statistics of Bing-QAC-2018 and Wiki-Synth.




Number of conversations (millions)



Number of interactions (millions)



Percentage of unique partial queries



Percentage of unique final partial queries



Lengths of partial queries – mean (characters)



Lengths of partial queries – (characters)



Table 3.

Goodness of power law fit against other distributions.

Frequency distribution







Bing-QAC-2018, \({\small \mathsf{{P}}}\)





Wiki-Synth, \({\small \mathsf{{P}}}\)





Bing-QAC-2018, \({{\small \mathsf{{FinalP}}}}\)





Wiki-Synth, \({{\small \mathsf{{FinalP}}}}\)





Heap’s Law. Heap’s law gives an estimate of the number of unique terms V in a collection as a function of the number of terms N. The relationship is given by \( V = kN^\beta \) and the typical values of the parameters are \(30 \le k \le 100 \), with \(\beta \approx 0.50 \) for English text [39] and \(\beta \approx 0.60\) for web documents [7]. The values for Heap’s law parameters estimated from Bing-QAC-2018 are \(k={14.50}\), \(\beta ={0.67}\) and the parameters from Wiki-Synth are \(k={32.61}\), \(\beta ={0.57}\). The difference in growth rate can be explained by the nature of the logs. The growth rate of V is expected to be higher in a collection containing numbers and spelling errors [39]. Each of the \({{\small \mathsf{{FinalP}}}}\) strings from Wiki-Synth is a Wikipedia title that comes from a curated collection whereas \({{\small \mathsf{{FinalP}}}}\) strings in Bing-QAC-2018 are strings entered by users. Considering, all partial queries, the parameters estimated are \(k={5.54}\), \(\beta ={0.71}\) for Bing-QAC-2018 and \(k={20.94}\), \(\beta ={0.60}\) for Wiki-Synth.

Frequency of Partial Query Strings. The likelihood of a query being repeated over time as a result of the power-law distribution affects the performance of various query processing strategies [13, 35]. The frequency distributions of partial queries and final queries from Wiki-Synth and Bing-QAC-2018 are given in Fig. 4 (left) with dashed lines showing the power-law fits for the corresponding distributions. The distributions give similar exponents except for the distribution of \({{\small \mathsf{{FinalP}}}}\) from Wiki-Synth, which has \(\alpha = {3.02}\), indicating a steeper decay in the probability of higher frequency partial query strings. The goodness of power-law fit against exponential and log-normal fits is compared using likelihood ratio R [2, 18] along with the significance level p, as reported in Table 3. Except for the frequency distribution of \({{\small \mathsf{{FinalP}}}}\) from Bing-QAC-2018, which is better explained with a log-normal distribution (\(R ={-2.42}, {0.04}\)), the other three distributions suit a power-law distribution, perhaps as a result of the sampling process followed. For some datasets, it is not surprising to get a better log-normal fit [18].

N-Gram Frequency. The distribution of terms in a collection can be modeled using a decay law which estimates the collection frequency of the ith most common term as \({{{\small \mathsf{{F}}}}_{i}} = Ci^k\), where \(k=-1\) and C is constant. Similarity in term distribution with a real log is considered to be a desirable property of a synthetic log [45]. Figure 4 (right) shows the bigram frequency distribution from partial queries and final partial queries. We find close correspondence between the bigram frequency distributions (Kolmogorov-Smirnov statistic, \(D={0.05}, {0.63}\) for \({{\small \mathsf{{FinalP}}}}\) and \(D={0.06}, {0.04}\) for all partial queries). Unigram frequencies follow a similar trend. While the bulk of the distribution can be explained with the decay law, the tail of the distributions coming from rare n-grams show deviations.
Fig. 4.

Power law fit (left) and bigram frequency distribution (right).

Empirical Entropy. The empirical entropy of the \({{\small \mathsf{{FinalP}}}}\) strings gives a measure of their compressibility. The kth-order empirical entropy of a string \({\small \mathsf{{T}}}[1\dots n]\) over an alphabet set \(\Sigma \) is given by
$$\begin{aligned} {{{\small \mathsf{{H}}}}_{k}}({\small \mathsf{{T}}}) = \frac{1}{n}\sum _{w \in \Sigma ^k} \left| {{\small \mathsf{{T}}}_w}\right| \cdot {{{\small \mathsf{{H}}}}_{0}}({\small \mathsf{{T}}}_w) \, , \text { with } {{{\small \mathsf{{H}}}}_{0}}({\small \mathsf{{T}}}) = \frac{1}{n}\sum _{c \in \Sigma }n_c\cdot \log \frac{n}{n_c} \, , \end{aligned}$$
where \({{{\small \mathsf{{T}}}}_{w}}\) is formed by collecting the characters that immediately follow context w in \({\small \mathsf{{T}}}\), and where \(n_c\) is the frequency of character c in \({\small \mathsf{{T}}}\). A lower bound on the number of bits required to encode \({\small \mathsf{{T}}}\) is given by \(n{{{\small \mathsf{{H}}}}_{k}}({\small \mathsf{{T}}})\) [41]. For a list of strings \({\small \mathsf{{T}}}_1, {\small \mathsf{{T}}}_2, \dots , {\small \mathsf{{T}}}_l\), it can be observed that \(\sum _{i} n_i {{{\small \mathsf{{H}}}}_{k}}({{{\small \mathsf{{T}}}}_{i}}) \le n_c {{{\small \mathsf{{H}}}}_{k}}({{{\small \mathsf{{T}}}}_{c}})\) where \({{{\small \mathsf{{T}}}}_{c}}\) is the concatenation of strings \({{{\small \mathsf{{T}}}}_{i}}\). This can be shown by extending the proof given by Navarro [41, Chapter 2], applying Jensen’s inequality to \(\sum _{i} n_i {{{\small \mathsf{{H}}}}_{k}}({{{\small \mathsf{{T}}}}_{i}})\). Therefore, we consider \(n_c {{{\small \mathsf{{H}}}}_{k}}({{{\small \mathsf{{T}}}}_{c}})\) as the worst case lower bound for the space required to encode the \({{\small \mathsf{{FinalP}}}}\) strings from Bing-QAC-2018 and Wiki-Synth. The values of \({{{\small \mathsf{{H}}}}_{k}}({{{\small \mathsf{{T}}}}_{c}})\) for \(k=0 \dots 4\) computed from Bing-QAC-2018 and Wiki-Synth are given in Table 4. The values for \({{{\small \mathsf{{H}}}}_{k}}\) tend to be similar between the two logs. Slightly higher values of \({{{\small \mathsf{{H}}}}_{k}}\) and increased \({{{\small \mathsf{{C}}}}_{t}}/n\) for Bing-QAC-2018 can be explained by the higher degree of randomness expected from web search queries, compared to the results obtained for Heap’s law coefficients.
Table 4.

Values of \({{{\small \mathsf{{H}}}}_{k}}\) and number of contexts \({{{\small \mathsf{{C}}}}_{t}}\) per length n (\({}\times 10^{-3}\)) of the concatenated string formed by the \({{\small \mathsf{{FinalP}}}}\) strings of Bing-QAC-2018 and Wiki-Synth.







\({{{\small \mathsf{{H}}}}_{0}}\)

\({{{\small \mathsf{{C}}}}_{t}}/n\)

\({{{\small \mathsf{{H}}}}_{1}}\)

\({{{\small \mathsf{{C}}}}_{t}}/n\)

\({{{\small \mathsf{{H}}}}_{2}}\)

\({{{\small \mathsf{{C}}}}_{t}}/n\)

\({{{\small \mathsf{{H}}}}_{3}}\)

\({{{\small \mathsf{{C}}}}_{t}}/n\)

\({{{\small \mathsf{{H}}}}_{4}}\)

\({{{\small \mathsf{{C}}}}_{t}}/n\)























Table 5.

Example for pre-correction and post-correction strings. In the table “pal” is taken as the pre-correction string and “pla” is taken as the post-correction string. This correction pair will be classified as a transposition error.

Identification of Typing Errors. Similarities in typing patterns between the synthetic QAC log and the original QAC log are verified by analyzing the typing errors present in both logs. Typing errors are identified by extracting and comparing a set of pre-correction and post-correction strings from the conversations, extending the method proposed by Baba and Suzuki [6]. When the kth word \({{{\small \mathsf{{w}}}}_{k}}\) in \({\small \mathsf{{P}}}\) gets deleted from the ith interaction, \({{{\small \mathsf{{w}}}}_{k}}\) from the \(i-1\)th interaction is taken as the pre-correction string. To find the post-correction string, the edit distance between the pre-correction string and the words \({{{\small \mathsf{{w}}}}_{k}}\) from the interactions after the end of current deletion sequence are calculated until \(\left| {{{{\small \mathsf{{w}}}}_{k}}}\right| \) from interaction j is less than that from interaction \(j+1\). From this list, the word having minimum edit distance with the pre-correction string is taken as the post correction string. Table 5 gives an example for how pre-correction and post-correction strings are calculated from a deletion sequence. The edit distance between two strings is calculated as a Damerau-Levenshtein distance, so transposition of two characters is given a cost of 1. The pre-correction and post-correction pairs are then compared to classify the possible typing errors as one of the types discussed above. The results obtained by analyzing the typing errors from Bing-QAC-2018 and the synthetic QAC log are in Fig. 5. The fraction of transposition errors was found to be higher in Wiki-Synth compared to Bing-QAC-2018 while the latter contains more deletion errors than Wiki-Synth. As a result of the method of simulating substitution errors, they tend to occur more towards the end of words in Wiki-Synth. The remainder of the error categories are distributed across similar word positions in both the logs, suggesting similarities in typing patterns.
Fig. 5.

Comparison between frequency (left) and position (right) of typing errors.

5 Conclusions

We have explored a method for generating a synthetic QAC log from an abstract QAC log, by mapping the word lengths of the abstract QAC log to those of a publicly available string collection, and applying a range of corrective techniques. Synthetic QAC formation can also be posed as a language generation problem relying on various models of QAC systems [31, 33].

We have demonstrated that the synthetic log generated from a pre-existing string collection encompasses many of the properties found in the original QAC log from which the abstract QAC log was derived. In particular, an analysis of typing errors found in actual QAC logs is reported, along with a description of how they were introduced into the synthetic log. As a result, there is a close correspondence between the real QAC log and the synthetic QAC log across a range of properties, each of which might influence the computational cost of providing the completion strings. That is, experiments using the synthetic QAC log can be expected to provide close approximations to behaviors that would be observed on a real QAC log. A particular example is efficiency, which we plan to examine in future work, comparing the performance of a range of QAC implementations using both synthetic and real QAC logs.


  1. 1.
  2. 2., accessed 29th October, 2019.

  3. 3., accessed 29th October, 2019.

  4. 4., accessed 29th October, 2019.

  5. 5.



This work was supported by the Microsoft Research Centre for Social Natural User Interfaces (SocialNUI) at The University of Melbourne. We are grateful to Peter Bailey (Microsoft Australia) for his support of that project, and for additionally facilitating the work that is reported here.


  1. 1.
    Adar, E.: User 4xxxxx9: anonymizing query logs. In: Proceedings of the WWW Query Log Analysis Workshop (2007).
  2. 2.
    Alstott, J., Bullmore, E., Plenz, D.: Powerlaw: a Python package for analysis of heavy-tailed distributions. PLoS One 9(1), 1–11 (2014)CrossRefGoogle Scholar
  3. 3.
    Askitis, N., Sinha, R.: HAT-trie: a cache-conscious trie-based data structure for strings. In: Proceedings of the Australasian Conference on Computer Science, pp. 97–105 (2007)Google Scholar
  4. 4.
    Askitis, N., Zobel, J.: Redesigning the string hash table, burst trie, and BST to exploit cache. ACM J. Exp. Algorithmics 15, 1–7 (2010)MathSciNetGoogle Scholar
  5. 5.
    Azzopardi, L., de Rijke, M., Balog, K.: Building simulated queries for known-item topics: an analysis using six European languages. In: Proceedings of the SIGIR, pp. 455–462 (2007)Google Scholar
  6. 6.
    Baba, Y., Suzuki, H.: How are spelling errors generated and corrected? A study of corrected and uncorrected spelling errors using keystroke logs. In: Proceedings of the ACL, pp. 373–377 (2012)Google Scholar
  7. 7.
    Baeza-Yates, R., Saint-Jean, F.: A three level search engine index based in query log distribution. In: Nascimento, M.A., de Moura, E.S., Oliveira, A.L. (eds.) SPIRE 2003. LNCS, vol. 2857, pp. 56–65. Springer, Heidelberg (2003). Scholar
  8. 8.
    Baeza-Yates, R., Tiberi, A.: Extracting semantic relations from query logs. In: Proceedings of the KDD, pp. 76–85 (2007)Google Scholar
  9. 9.
    Bar-Yossef, Z., Kraus, N.: Context-sensitive query auto-completion. In: Proceedings of the WWW, pp. 107–116 (2011)Google Scholar
  10. 10.
    Bast, H., Weber, I.: Type less, find more: fast autocompletion search with a succinct index. In: Proceedings of the SIGIR, pp. 364–371 (2006)Google Scholar
  11. 11.
    Bast, H., Weber, I.: The CompleteSearch engine: interactive, efficient, and towards IR & DB integration. In: Proceedings of the CIDR, pp. 88–95 (2007)Google Scholar
  12. 12.
    Bast, H., Majumdar, D., Weber, I.: Efficient interactive query expansion with complete search. In: Proceedings of the CIKM, pp. 857–860 (2007)Google Scholar
  13. 13.
    Beitzel, S.M., Jensen, E.C., Chowdhury, A., Grossman, D., Frieder, O.: Hourly analysis of a very large topically categorized web query log. In: Proceedings of the SIGIR, pp. 321–328 (2004)Google Scholar
  14. 14.
    Bhatia, S., Majumdar, D., Mitra, P.: Query suggestions in the absence of query logs. In: Proceedings of the SIGIR, pp. 795–804 (2011)Google Scholar
  15. 15.
    Cai, F., Liang, S., de Rijke, M.: Time-sensitive personalized query auto-completion. In: Proceedings of the CIKM, pp. 1599–1608 (2014)Google Scholar
  16. 16.
    Cai, F., Reinanda, R., de Rijke, M.: Diversifying query auto-completion. ACM Trans. Inf. Syst. 34(4), 25:1–25:33 (2016)CrossRefGoogle Scholar
  17. 17.
    Chaudhuri, S., Kaushik, R.: Extending autocompletion to tolerate errors. In: Proceedings of the SIGMOD, pp. 707–718 (2009)Google Scholar
  18. 18.
    Clauset, A., Shalizi, C.R., Newman, M.E.J.: Power-law distributions in empirical data. SIAM Rev. 51(4), 661–703 (2009)MathSciNetCrossRefGoogle Scholar
  19. 19.
    Cooper, A.: A survey of query log privacy-enhancing techniques from a policy perspective. ACM Trans. Web 2(4), 19:1–19:27 (2008)CrossRefGoogle Scholar
  20. 20.
    Damerau, F.J.: A technique for computer detection and correction of spelling errors. Commun. ACM 7(3), 171–176 (1964)CrossRefGoogle Scholar
  21. 21.
    Dhakal, V., Feit, A.M., Kristensson, P.O., Oulasvirta, A.: Observations on typing from 136 million keystrokes. In: Proceedings of the CHI, pp. 646:1–646:12 (2018)Google Scholar
  22. 22.
    Gentner, D.R., Grudin, J.T., Larochelle, S., Norman, D.A., Rumelhart, D.E.: A glossary of terms including a classification of typing errors. In: Cooper, W.E. (ed.) Cognitive Aspects of Skilled Typewriting, pp. 39–43. Springer, New York (1983). Scholar
  23. 23.
    Hawking, D., Billerbeck, B.: Efficient in-memory, list-based text inversion. In: Proceedings of the Australasian Document Computing Symposium, pp. 5.1–5.8 (2017)Google Scholar
  24. 24.
    Heafield, K.: KenLM: faster and smaller language model queries. In: Proceedings of the Workshop on Statistical Machine Translation, pp. 187–197 (2011)Google Scholar
  25. 25.
    Heinz, S., Zobel, J., Williams, H.: Burst tries: a fast, efficient data structure for string keys. ACM Trans. Inf. Syst. 20(2), 192–223 (2002)CrossRefGoogle Scholar
  26. 26.
    Hofmann, K., Mitra, B., Radlinski, F., Shokouhi, M.: An eye-tracking study of user interactions with query auto completion. In: Proceedings of the CIKM, pp. 549–558 (2014)Google Scholar
  27. 27.
    Hsu, B.-J.P., Ottaviano, G.: Space-efficient data structures for top-\(k\) completion. In: Proceedings of the WWW, pp. 583–594 (2013)Google Scholar
  28. 28.
    Ji, S., Li, G., Li, C., Feng, J.: Efficient interactive fuzzy keyword search. In: Proceedings of the WWW, pp. 371–380 (2009)Google Scholar
  29. 29.
    Jiang, J., Ke, Y., Chien, P., Cheng, P.: Learning user reformulation behavior for query auto-completion. In: Proceedings of the SIGIR, pp. 445–454 (2014)Google Scholar
  30. 30.
    Jordan, C., Watters, C., Gao, Q.: Using controlled query generation to evaluate blind relevance feedback algorithms. In: Proceedings of the JCDL, pp. 286–295 (2006)Google Scholar
  31. 31.
    Kharitonov, E., Macdonald, C., Serdyukov, P., Ounis, I.: User model-based metrics for offline query suggestion evaluation. In: Proceedings of the SIGIR, pp. 633–642 (2013)Google Scholar
  32. 32.
    Krishnan, U., Moffat, A., Zobel, J.: A taxonomy of query auto completion modes. In: Proceedings of the Australasian Document Computing Symposium, pp. 6:1–6:8 (2017)Google Scholar
  33. 33.
    Krishnan, U., Billerbeck, B., Moffat, A., Zobel, J.: Abstraction of query auto completion logs for anonymity-preserving analysis. Inf. Retrieval J. 22(5), 499–524 (2019). CrossRefGoogle Scholar
  34. 34.
    Kumar, R. Novak, J., Pang, B. Tomkins, A.: On anonymizing query logs via token-based hashing. In: Proceedings of the WWW, pp. 629–638 (2007)Google Scholar
  35. 35.
    Lempel, R., Moran, S.: Predictive caching and prefetching of query results in search engines. In: Proceedings of the WWW, pp. 19–28 (2003)Google Scholar
  36. 36.
    Li, G., Ji, S., Li, C., Feng, J.: Efficient fuzzy full-text type-ahead search. VLDB J. 20(4), 617–640 (2011). Scholar
  37. 37.
    Li, L., Deng, H., Dong, A., Chang, Y., Zha, H., Baeza-Yates, R.: Analyzing user’s sequential behavior in query auto-completion via Markov processes. In: Proceedings of the SIGIR, pp. 123–132 (2015)Google Scholar
  38. 38.
    Li, Y., Dong, A., Wang, H., Deng, H., Chang, Y., Zhai, C.: A two-dimensional click model for query auto-completion. In: Proceedings of the SIGIR, pp. 455–464 (2014)Google Scholar
  39. 39.
    Manning, C., Raghavan, P., Schütze, H.: Introduction to Information Retrieval. Cambridge University Press, Cambridge (2008)CrossRefGoogle Scholar
  40. 40.
    Maxwell, D., Bailey, P., Hawking, D.: Large-scale generative query autocompletion. In: Proceedings of the Australasian Document Computing Symposium, pp. 9:1–9:8 (2017)Google Scholar
  41. 41.
    Navarro, G.: Compact Data Structures: A Practical Approach. Cambridge University Press, Cambridge (2016)CrossRefGoogle Scholar
  42. 42.
    Shokouhi, M.: Learning to personalize query auto-completion. In: Proceedings of the SIGIR, pp. 103–112 (2013)Google Scholar
  43. 43.
    Smith, C.L., Gwizdka, J., Feild, H.: Exploring the use of query auto completion: search behavior and query entry profiles. In: Proceedings of the CHIIR, pp. 101–110 (2016)Google Scholar
  44. 44.
    Smith, C.L., Gwizdka, J., Feild, H.: The use of query auto-completion over the course of search sessions with multifaceted information needs. Inf. Process. Manag. 53(5), 1139–1155 (2017)CrossRefGoogle Scholar
  45. 45.
    Webber, W., Moffat, A.: In search of reliable retrieval experiments. In: Proceedings of the Australasian Document Computing Symposium, pp. 26–33 (2005)Google Scholar
  46. 46.
    Wobbrock, J.O., Myers, B.A.: Analyzing the input stream for character- level errors in unconstrained text entry evaluations. ACM Trans. Comput.-Hum. Interact. 13(4), 458–489 (2006)CrossRefGoogle Scholar
  47. 47.
    Xiao, C., Qin, J., Wang, W., Ishikawa, Y., Tsuda, K., Sadakane, K.: Efficient error-tolerant query autocompletion. Proc. VLDB 6(6), 373–384 (2013)CrossRefGoogle Scholar
  48. 48.
    Xiong, L., Agichtein, E.: Towards privacy-preserving query log publishing. In: Proceedings of the WWW Query Log Analysis Workshop (2007)Google Scholar
  49. 49.
    Zobel, J., Moffat, A., Ramamohanarao, K.: Inverted files versus signature files for text indexing. ACM Trans. Database Syst. 23(4), 453–490 (1998)CrossRefGoogle Scholar

Copyright information

© Springer Nature Switzerland AG 2020

Authors and Affiliations

  1. 1.The University of MelbourneMelbourneAustralia
  2. 2.Microsoft Australia and RMIT UniversityMelbourneAustralia

Personalised recommendations