|Publication number||USRE41152 E1|
|Application number||US 09/952,602|
|Publication date||Feb 23, 2010|
|Filing date||Sep 14, 2001|
|Priority date||Aug 6, 1996|
|Also published as||CA2263453A1, CA2263453C, DE69737892D1, DE69737892T2, EP0970419A1, EP0970419A4, EP0970419B1, US5951623, WO1998006028A1|
|Publication number||09952602, 952602, US RE41152 E1, US RE41152E1, US-E1-RE41152, USRE41152 E1, USRE41152E1|
|Inventors||Jeffrey C. Reynar, Fred Herz, Jason Eisner, Lyle Ungar|
|Original Assignee||Pinpoint Incorporated|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (41), Non-Patent Citations (15), Referenced by (50), Classifications (27), Legal Events (2)|
|External Links: USPTO, USPTO Assignment, Espacenet|
1. Field of the Invention
The present invention relates to a method and apparatus for compressing and decompressing textual data stored in digital form in a lossless manner. In other words, the original data is reconstructed in its original form after having first undergone the compression and then the decompression processes. The data is assumed to be drawn from a particular alphabet which is specified in advance, such as the ASCII code, which consists of a 7 or 8 bit representation of a particular set of characters.
2. Description of the Prior Art
Many different types of text compression techniques are described in the prior art. The text compression techniques described herein are based on the text compression techniques developed by Lempel and Ziv, who developed two techniques for text compression which are similar but have important differences. These two methods were outlined in papers entitled “A Universal Algorithm for Sequential Compression,” IEEE Transactions on Information Theory, Vol. IT-23, No. 3, pp. 337-343, and “Compression of Individual Sequences via Variable-Rate Coding,” IEEE Transactions on Information Theory, Vol. IT-24, No. 5, pp. 530-536, and are referred to commonly as LZ77 and LZ78, respectively.
LZ77 is a text compression technique in which pointers to previously compressed material within a fixed size window are used to compress new material. The fixed size “compression window” is moved across the text data as it is being compressed to exploit the principle of locality, i.e., that data is likely to be most similar to proximal data. An example of LZ77 will now be described with respect to FIGS. 1(A) and 1(B).
In FIGS. 1(A) and 1(B), a small window size of 8 characters is assumed for illustrative purposes. As shown in FIG. 1(A), the text that has not yet been compressed is compared to the contents of the (up to) 8 character window containing the (up to) 8 characters of the compressed text immediately preceding the text which has not yet been compressed. The longest match starting at the beginning of the text which has not yet been compressed with a sequence in the 8 character window is identified. In FIG. 1(A), the longest match from the 8 character window is “BB.” A pointer to this sequence (“BB”), its length (2), and an extension (the next character after the match in the text to be compressed) are then either transmitted or compressed and stored locally, depending on the application of the data compression algorithm. However, if no matching sequence is found in the 8 character window, then a literal character is transmitted. Once the block of data pointed to by the pointer has been compressed and the information regarding its compression has been transmitted, the window is moved by the number of characters referred to by the pointer (BB) plus the extension, if any. In addition, the pointer to the region of text being compressed is updated by this number of characters. As illustrated in FIG. 1(B), this process repeats for the shifted data until the data being compressed is exhausted.
LZ78 differs from LZ77 in that the text compression is achieved by parsing the data being compressed into phrases which are entered into a compression dictionary. Pointers to these phrases or dictionary entries are then used to compress new data. Initially, the dictionary contains only an empty string (phrase of length zero). The phrase to be compressed, at each step, is the longest phrase at the start of the new data such that the prefix of this phrase is an entry in the dictionary, where the prefix is defined to be the phrase with its final character removed. The remaining character is called the extension. Thus, when the first phrase is seen, it is encoded as a reference to a dictionary entry consisting of the empty string (which is the only entry initially found in the dictionary), followed by the last and only character of the phrase. This character is then placed in the dictionary (assuming the dictionary is not full), and the process of identifying a phrase and transmitting its prefix (as a reference to the dictionary entry matching the prefix) and extension is repeated. The not yet compressed data will then be compared to both dictionary entries: the empty string and the phrase consisting of the already encountered character. If the next character in the new data does not match the already compressed character, then it, too, will be compressed as the empty string plus the character being compressed. In this way, each phrase of the data is compressed as a prefix, which is found in the dictionary and is chosen to be as long as possible, and an extension, which is the character which follows the prefix in the input data. An example of LZ78 will now be described with respect to FIG. 2.
Numerous data compression systems have been described in the prior art which utilize the concept of a compression dictionary as described by Lempel and Ziv.
For example, Giltner et al. describe in U.S. Pat. No. 4,386,416 a system for use in transmitting data over a Telex or similar network. The system described by Giltner et al. uses two dictionaries. The first is pre-filled with frequent words from the data's language, while the second dictionary is initially empty and is filled with words which are encountered in the data but which are not present in the first dictionary. When transmitting data, if a word is found in the first dictionary, an escape code and the number of the word's entry in the first dictionary are transmitted. If a word is not found in the first dictionary, it is compressed using Huffman coding and it is added to the second dictionary for later use. As a result, if the word is encountered again, it can be transmitted by sending an escape code indicating that the number of the word's entry in the second dictionary refers to the second dictionary, followed by the number of the word's entry in the second dictionary. Giltner et al. define a “word” to be either a predetermined number of characters or a sequence of characters surrounded by white space or a combination of white space and punctuation. A small, but fixed number of words common to all the types of messages handled by the Telex or similar network is provided in the first dictionary, and additional “words” are stored in the second dictionary. However, Giltner et al. do not address frequently occurring sequences of text which fall outside the limited definition of a valid word. As a result, Giltner et al. do not take advantage of the fact that the similarity of texts is greater when the comparison between them is made at the level of character sequences. Also, Giltner et al. do not address how words which occur frequently in the text can be chosen for the first dictionary whereby it is filled with valid words which are frequent within the type of text being transmitted. Giltner et al. also fail to teach how to identify the most appropriate library of text or the identification of the genre of the document to be compressed. Furthermore, Giltner et al.'s library is fixed; users cannot create their own pre-filled dictionaries as needed.
Similarly, Weng describes in U.S. Pat. No. 4,881,075 an “adaptive” data compression technique which uses two dictionaries. The first dictionary is used to perform compression or decompression while the second is being rebuilt to better reflect the local characteristics of the most recent input data. The second dictionary is then used to compress and decompress the input data while the first dictionary is being rebuilt using the most current input data. Weng repeatedly switches between dictionaries until compression is completed.
Kato et al. describe in U.S. Pat. No. 4,847,619 a modification to adaptive compression techniques in which the compression system's degree of compression is monitored and the dictionary is reset when the degree of compression drops below a threshold. The reset is not permitted to occur before the dictionary is sufficiently full in order to prevent the dictionary from resetting prematurely. This technique could be used in conjunction with a LZ compression technique, or any other adaptive technique.
In U.S. Pat. No. 5,153,591, Clark describes a modification to the Lempel-Ziv compression algorithm in which the dictionary is stored as a tree data structure. This allows large dictionaries to be stored in less space than in the original embodiment described in U.S. Pat. No. 4,464,650. In addition, it allows these dictionaries to be searched more easily and more quickly.
In U.S. Pat. No. 5,243,341, Seroussi et al. outline a Lempel-Ziv variant in which two dictionaries are used. The first dictionary is used until it is filled, then it is replaced with a standby dictionary, which is filled with those entries from the first dictionary which yield the most compression before compression continues.
Many other modifications to the original Lempel-Ziv compression techniques appear in the prior art.
For example, Welch describes in U.S. Pat. No. 4,558,302 an implementation of Lempel-Ziv in which the encoding and decoding processes require less complicated computation and, therefore, are faster than in the implementation described in U.S. Pat. No. 4,464,650—Eastman et al.
Miller et al. suggest in U.S. Pat. No. 4,814,746 several modifications to the Lempel-Ziv algorithm. The first of these modifications is to include all possible characters in the dictionary before compression actually begins. As a result, it is not necessary to transmit a flag which indicates that the following datum is a character rather than a pointer. In addition, Miller et al. also associate a time stamp with each dictionary entry in order to facilitate the removal of the least recently used entry when the dictionary becomes full. These modifications are aimed at reducing the memory requirements by limiting the dictionary to a fixed size and improving compression by allowing the dictionary to more accurately reflect the current characteristics of the data being compressed.
Storer describes in U.S. Pat. No. 4,876,541 a compression technique which does not suffer from some of the same difficulties as prior Lempel-Ziv techniques. In particular, unencoded characters never need to be transmitted, since the compression dictionary initially contains all of the characters in the alphabet, as in U.S. Pat. No. 4,814,746—Miller et al. In addition, a least recently used queue is maintained so that the dictionary can be purged of less useful entries. The encoding and decoding dictionary in Storer's system can vary in size, and there may be several active at a time. The compression ratio of each of the dictionaries is monitored, and the one which yields the best compression is used.
In U.S. Pat. No. 4,906,991, Fiala et al. describe a substitution style data compression technique which is somewhat similar to Lempel-Ziv compression. Their technique relies on searching a fixed window of characters (e.g. 4096 characters) which have already been compressed in order to determine whether the text being compressed can be encoded as a pointer to a location within the window. If the text being compressed can be encoded in this manner, a pointer to the starting location, along with the length of the overlap between the text being compressed and the location within the window is generated. If the text being compressed cannot be encoded in this manner, it is encoded as a length followed by a literal string of that length. Like LZ78 compression, this technique fails to compress data much at the beginning of documents, since the window is devoid of strings which can be pointed to in order to bring about compression.
O'Brien suggests in U.S. Pat. No. 4,988,998 modifications to the Lempel-Ziv algorithm which allow for enhanced compression of data which contains long strings of repeated characters. Since the Lempel-Ziv algorithm adds entries to the compression dictionary by appending a single character to an existing dictionary entry, it will take many occurrences of a repeated string of characters before such strings are found in the dictionary. Accordingly, O'Brien preprocesses the data using a run-length encoding technique in which the run-lengths are inserted into the text. The resulting combination of text and run-lengths for repeated characters is then compressed using the Lempel-Ziv technique.
In U.S. Pat. No. 5,049,881, Gibson et al. describe a data compression system which creates its own pointers from the sequence of characters previously processed and emphasizes maximizing the product of the data rate and the compression ratio. Thus, previously input data is used as the dictionary and is combined with a hashing algorithm to find candidates for string matches without the requirement of a string matching table.
In U.S. Pat. No. 5,058,137, Shah describes a Lempel-Ziv decoder which has memories for storing code words and data separately. Upon receipt of a code word, the decoder stores the previously received code word, applies the newly received code word to the code word memory to obtain the location of the last data element which is part of the data represented by the newly received code word, and another code word associated with the prefix. Upon completion of decoding the latest code word, the first data element of the decoded word is appended to the next previously received code word, and the combination is stored as the equivalent of a code word which is next after the highest code word already received. At least one memory is shared for use during encoding and decoding.
In U.S. Pat. No. 5,087,913, Eastman describes a Lempel-Ziv algorithm which uses a searchtree database to allow later portions of the data to be decompressed without having to decompress all preceding portions. The searchtree database is grown to a fixed, predetermined size and is allowed to grow no further. The fact that the compression searchtree database is established in advance of decompression allows decompression of portions of the data without decompressing the entire preceding portion of the data.
In U.S. Pat. No. 5,140,321, Jung details a Lempel-Ziv modification which allows for enhanced compression speed at the cost of a reduction in compression. Rather than attempting to find the optimal matching substring in the entire compressed portion of the data, the principle of locality is exploited and only the most recent compressed data in a first-in-first-out buffer is examined to find a matching sequence. A hash table is used to store the strings which have been compressed recently and to allow for fast retrieval of matching strings.
In U.S. Pat. No. 5,179,378, Ranganathan et al. describe a Lempel-Ziv implementation which uses a systolic array of processors to improve performance by forming fixed-length codewords from a variable number of data symbols.
In U.S. Pat. No. 5,262,776, Kutka describes an implementation of the Lempel-Ziv algorithm which takes advantage of a tree data structure to avoid the search step normally required by the compression process. The sequence of elements in a primary sequence is converted into elements in a reduced set of elements using escape sequences. This technique is particularly suited to compressing data representing the coefficients of a discrete cosine transform of an image.
In addition to the above modifications to the Lempel-Ziv compression technique which are described in the patent literature, others appear in technical journals.
For example, in “Linear Algorithm for Data Compression via String Matching,” Rodeh et al. describe a modification to the LZ77 technique in which the size of the window is not fixed. As a result, pointers to previous strings in the compressed portion of the data grow in length and are encoded in a variable-length code.
In “Better OPM/L Text Compression,” Bell describes a Lempel-Ziv variant referred to as LZSS in which not all compression is done using a combination of a prefix and an extension. Instead, if the cost of transmitting a pointer is higher than that of merely transmitting a character or sequence of characters, then the character or sequence of characters is transmitted. A binary search tree is used to find the longest string match, without a limit on the length of the match.
In “A Technique for High-Performance Data Compression,” Welch describes a modification to LZ78 in which only pointers to previously compressed data are used, rather than a combination of pointers and characters. A string table is formed which maps strings of input characters into fixed-length codes. For every string in the table, its prefix string is also in the table. The string table contains strings that have been encountered previously in the message being compressed. It consists of a running sample of strings in the message so that available strings reflect the statistics of the message. This technique is commonly referred to as LZW. It uses a “greedy” parsing algorithm in which the input string is examined character-serially in one pass, and the longest recognized input string is parsed off each time. The strings added to the string table are determined by this parsing.
In “Variations on a Theme by Ziv and Lempel,” Miller and Wegman describe another variant of LZ78. In their version, the dictionary is filled in advance with all strings of length 1 (that is, all of the characters in the alphabet over which compression is taking place), which helps to reduce, but does not eliminate the problem of starting with a dictionary devoid of useful entries. Also, rather than reset the compression dictionary when it becomes full, they propose to delete strings from the dictionary which were least recently used. However, the largest contribution of their version, which will be referred to as LZMW, is that extensions are never transmitted. Rather, since the dictionary begins with all strings of length 1, it is possible to encode all of the data using the initial dictionary. However, this would result in no compression. Instead, the dictionary is grown by adding entries to the dictionary which consist of the concatenation of the previous two matches.
Commonly used compression algorithms also use variations of the LZ77 and LZ78 algorithms. For example, the compression algorithm used in commercially available “zip” and freely available “gzip” is a variation of LZ77 which finds duplicated strings in the input data. In “gzip,” the second occurrence of a string is replaced by a pointer to the previous string in the form of a pair (distance, length). When a string does not occur anywhere in the previous number of bytes within a designated distance, such as 32 Kbytes, it is transmitted as a sequence of literal bytes. Literals or match lengths are compressed with one Huffman tree, and match distances are compressed with another tree. The trees are stored in a compact form at the start of each block. The blocks can have any size, except that the compressed data for one block must fit in available memory. A block is terminated when “gzip” determines that it would be useful to start another block with fresh trees. Duplicated strings are found using a hash table. All input strings of length 3 are inserted in the hash table, and a hash index is computed for the next 3 bytes. If the hash chain for this index is not empty, all strings in the chain are compared with the current input string, and the longest match is selected. The hash chains are searched, starting with the most recent strings, to favor small distances and thus take advantage of the Huffman coding. The hash chains are singly linked. There are no deletions from the hash chains; the algorithm simply discards matches that are too old. To avoid a worst-case situation, very long hash chains are arbitrarily truncated at a certain length as determined by a runtime option. As a result, “gzip” does not always find the longest possible match but generally finds a match which is long enough.
Unfortunately, despite the large number of Lempel-Ziv variants in the prior art, none adequately addresses the problem that beginning with a dictionary completely devoid of words virtually prevents small files from being compressed at all and prevents larger files from being compressed further. U.S. Pat. Nos. 4,814,746 and 4,876,541 and the work of Miller and Wegman begin to address this problem by beginning with a dictionary containing all of the characters in the character set in which the data is encoded. This solves the problem of needing to transmit escape codes to indicate that what follows is not a dictionary entry number, but a character.
However, the present inventors have found that further compression may be obtained by observing that many documents, especially those textual documents written in either natural human languages or computer programming languages, such as English or C, have a small number of words which are statistically extremely frequent and which should be part of the compression dictionary. Zipf has shown this to be true for language in a book entitled Human Behavior and the Principle of Least Effort. In fact, the frequency of words in English obeys what has come to be known as a Zipfian distribution. That is, the product of the rank of a word and its frequency is approximately constant. Thus, the second most common word will appear roughly half as many times as the most frequent word. This implies that the most common words will comprise a large fraction of the total occurrences of all words in a document. For instance, in the Wall Street Journal data collected as part of the treebank project described in an article by Marcus et al. entitled Building a Large Annotated Corpus of English: The Penn Treebank, Computational Linguistics, Vol. 19, No. 2, pp. 313-330 (1993), the first ten “words”, which are “,”, “the”, “.”, “of”, “to”, “a”, “and”, “in”, “'s”, “is” and “that” account for slightly more than 25 percent of all of the words in the corpus. The 100 most frequent words alone account for 48.1 percent of all words in the corpus, while the 500 most frequent words account for 63.6 percent. This means that the remaining 80901 words account for the remaining 37.4 percent of all words in the corpus. Thus, the average word found in the top five hundred words is approximately 275 times more frequent than the average word not found in the top five hundred words.
It is thus desired to modify the Lempel-Ziv text compression techniques described in the prior art to take advantage of this observation so as to allow further compression of large documents as well as significant compression of smaller documents.
The present invention expands upon the Lempel-Ziv text compression techniques of the prior art. In particular, the present invention provides a compression dictionary and/or a compression window which contains some of the statistically significant words which will almost certainly allow compression to be achieved earlier into a document than if the compression software were required to relearn the list of extremely common words, letter sequences and phrases in the document to be compressed. For example, when using compression techniques based on LZ78, in which entries are added to the compression dictionary by extending previous entries by a single character, the invention makes it unnecessary to take several repetitions of even extremely frequent words before they are entered into the compression dictionary. This also means that the compression dictionary used in accordance with the invention does not necessarily contain prefixes of the long, frequent words that are initially present in the dictionary, unless such prefixes are themselves frequent. Rare prefixes aid the compression process little, and their main contribution in LZ78 is to allow a longer word to enter the compression dictionary, a step which is unnecessary when using the techniques of the invention, where the longer, frequent word is in the dictionary at the outset. Omitting the rare prefixes from the dictionary leaves room for other, more useful phrases.
While most extremely frequent words are short, the present inventors recognize that knowledge of the type of text being compressed can lead to larger benefit from a pre-filled dictionary. If, for instance, scientific text is being compressed, then samples of similar texts can be used to get a profile of the type of text being compressed. In scientific texts, some long phrases may be extremely common. These phrases would be learned by the compression software and would be used when the first reference to them was made in the data being compressed. So, if, for instance, a length 10 word or phrase were common, then even the first occurrence of that word could be encoded using a single dictionary reference, whereas under conventional LZ78, such an encoding might not be possible until as late as the eleventh occurrence of the word. The dictionary will also be prevented from containing substrings of this frequent word or phrase as would be the case in LZ78 unless such substrings are independently added during the compression process because they are useful outside the context of the frequent word or phrase.
When compressing text which is in a computer programming language, the present inventors have recognized that there is potentially even more to gain at the outset of compression. The programming language C, for instance, has a small number of words which represent constructs in the language. While the user can add to the list of these frequently occurring words by creating variables and functions, all C programs will make use of a subset of this basic set of words. As a result, having these words in the dictionary prior to the start of compression will result in better compression, especially for small files.
Thus, in order to take advantage of the frequency distribution of words in human languages and other sorts of data as well, the present inventors propose modifying standard Lempel-Ziv compression algorithms to incorporate a dictionary pre-filled with statistically important words and the entire alphabet over which compression is taking place. The process of filling the dictionary with words that are significant is handled in any of a number of ways. One possibility is for a user of the compression software to create a list of frequently used words. Since this process is objectionable in most circumstances, automatic processes are preferably used in accordance with the invention. However, additional pre-filled compression dictionaries may be created by the user of the compression software as desired.
On the other hand, several pre-filled compression dictionaries may be automatically tested at the start of the compression process on a small subset of the entire document to be compressed, and the dictionary that achieves the highest compression on this subset is chosen for compressing the entire document. Of course, the user may also select a particular compression dictionary as desired. Preferably, the identity of the pre-filled dictionary being used for the compression is stored in the compressed data so that the decompression software can identify which pre-filled dictionary to use during decompression.
The same pre-filled dictionary must be available to both the comprising device at compression time, and to the decompressing device at decompression time. It is possible to transmit such dictionaries from one device to another over a communications network, as necessary, to make the appropriate dictionaries available. For example, a plurality of pre-filled data compression dictionaries may be stored on communicating servers so that any text shared among the servers may be compressed/decompressed using the techniques of the invention. Also, the pre-filled data compression dictionaries are preferably created for different genres of text data and stored hierarchically so that the optimum pre-filled data compression dictionary may be selected for the data to be transmitted. In order to save memory space, any common entries among the plurality of pre-filled data compression dictionaries need only be stored once on each server. On the other hand, a pre-filled data compression dictionary used to compress a particular document may be the data specific dictionary formed during the compression of another, related text. In this manner, the compression of a plurality of documents may be iterative and based on a single original pre-filled data compression dictionary.
As will become apparent to those skilled in the art, the techniques of the invention may be used in connection with any of the known variants of the Lempel-Ziv compression algorithms.
The above and other objects and advantages of the invention will become more apparent and will be more readily appreciated from the following detailed description of the presently preferred exemplary embodiments of the invention taken in conjunction with the accompanying drawings, of which:
FIGS. 1(A) and 1(B) together provide an example of prior art LZ77 text compression.
The present invention will be described in detail below with respect to
The extension to the Lempel-Ziv algorithms in accordance with the invention can be used in conjunction with all of the known variants of the Lempel-Ziv compression techniques described in the patent literature as well as in the text compression literature. However, in presently preferred embodiments, the present invention is used as a modification to the LZ77 or LZ78 compression techniques or as a modification to a Lempel-Ziv variant which is itself a modification to or extension of the LZ77 or LZ78 techniques. The extension to the Lempel-Ziv algorithms in accordance with the invention is preferably implemented as part of a software compression package or as a module which is used in conjunction with pre-existing software packages which would benefit from data compression, such as a commercial word processing package. A sample system configuration is shown in FIG. 4.
The present invention is preferably implemented as software containing instructions for controlling a processor of a user's computer 1. In a preferred embodiment, the executable version of the software implementing the Lempel-Ziv compression algorithm is stored on a fixed program storage device such as a hard disk 2 which is readable by a processor (CPU) 3 of computer 1 whereby the program of instructions stored thereon for implementing the Lempel-Ziv compression algorithm is executable by the processor 3 to perform the desired data compression. As shown in
In accordance with the invention, the pre-filled compression dictionaries are stored on the hard disk 2 along with the executable version of the software implementing the Lempel-Ziv compression algorithm and are loaded into the RAM 4 of CPU 3 during the execution of the Lempel-Ziv data compression algorithm. The data to be compressed may be stored as files on the hard disk prior to being compressed and then restored locally or transmitted after compression. In other words, if the data compression is being performed strictly to save local storage space, the results of the compression process are written to the hard disk 2 or other local memory. Alternatively, the input data and the pre-filled compression dictionaries may be provided on a computer readable medium such as a floppy disk 5, an optical disk 5′, a CD ROM 5″, received via modem 6, or provided directly via a network connection. On the other hand, if the data compression is being performed to allow faster data transmission across a point to point connection (e.g., via modem 6), then the CPU 2 is connected to the modem 6 or some other transmission device, and the compressed data is transmitted via that transmission device. In addition, the compressed data may be returned to the original computer readable medium (e.g., floppy disk 5, optical disk 5′, or CD ROM 5″).
Accordingly, during operation of the compression algorithm in accordance with the invention, RAM 4 typically receives the executable code, one or more pre-filled compression dictionaries, and the input data from hard disk 2, some other memory element such as a floppy disk 5, optical disk 5′, or CD ROM 5″, or via a modem connection, and the compressed data is either restored on the hard disk 2 or some other memory element such as floppy disk 5, optical disk 5′, or CD ROM 5″, or is transmitted via modem 6 to another computer for storage and/or decompression. In addition, when used with LZ78 algorithm variants, RAM 4 will preferably allocate space for a data specific dictionary of the type conventionally developed during operation of the LZ78 algorithm and its variants.
The process of pre-filling a compression dictionary with words that are significant to the Lempel-Ziv algorithm is handled in any of a number of ways in accordance with the invention. One possibility is for a user of the compression software to create a list of frequently used words. Since this process is objectionable in most circumstances, automatic processes are preferably used in accordance with the invention. As in U.S. Pat. No. 4,386,416, one possibility for such an automatic process is to restrict the pre-filled compression dictionary to what are normally referred to as words-that is, sequences of characters which are delimited by either spaces, tabs or punctuation marks. However, it is preferred that statistics be collected from a representative sample of text to determine the most frequent sequences of characters of various lengths. The number of dictionary entries of each length can be determined by a function of the amount of compression which will result from the use of a particular length entry (that is, the number of bits originally required to encode a string of the particular length less the number of bits required to transmit a reference to a dictionary entry, which will not vary with the length of the dictionary entry, but rather will be a fixed quantity determined from the size of the dictionary) and the observed frequency of that sequence within the sample of text utilized.
For example, assuming a dictionary reference when performing compression costs 12 bits (that is, the dictionary contains at most 212 or 4096 entries) and also assuming a 7 bit encoding of the ASCII character set, a sequence of 10 characters would require 70 bits of storage space without compression, while a sequence of 9 characters would require 63 bits of storage space without compression. If in the sample of text from which frequency statistics were derived a particular length 10 sequence occurred 60 times, then the savings that would result, if that sequence were in the dictionary, would be 60*70=4200 (the cost of transmitting the raw data) less 60*12 =720 (the cost of transmitting the pointers) which is 3480 bits of savings. Assuming the length 9 string was more frequent and occurred 65 times, the cost savings associated with having that length 9 sequence in the dictionary would be (65*63)−(65*12)=3315 bits of savings. Thus, the length 10 sequence would be preferred as a dictionary entry over the length 9 sequence.
The contents of the pre-filled dictionary may be efficiently selected as follows. The text of all of the documents from which the pre-filled dictionary will be learned should be loaded into a data structure which allows all of the subsequences of the text to be identified easily. Generally, a maximum length will need to be chosen for the strings in the dictionary to prevent the cost of the computation from getting too high. A length of 10 or 15 characters would probably suffice, but an appropriate maximum length may be determined experimentally, if desired. However, since there is not a particular maximum length specified in the design of the algorithm of the invention, the choice of a maximum length is left to the user of the dictionary building software. As a result, the chosen maximum length may not result in optimal compression. This is because there is a trade-off between compression and the cost of generating the pre-filled dictionary, and the user of the dictionary building software must choose a maximum length which will allow the pre-filled dictionary to be built in a reasonable amount of time on his or her computer.
A tree data structure in which each node contains a one character string and a count will be used to identify frequently occurring strings. The root of this tree contains the empty string and the total number of characters in the corpus. Each daughter node will extend the string created by concatenating all of the characters contained by nodes on the path from the root of the tree to the current node. In this way, all of the character sequences up to a particular length can be stored in a tree no deeper than the chosen maximum length. This tree could grow to be very large, with the leaves of the tree numbering the size of the alphabet raised to the power of the depth of the tree. However, because of the sparsity of data, the tree should in practice be much smaller. As will be appreciated by those skilled in the art, the data is sparse as a result of the properties of human language whereby the languages do not contain, or at least do not frequently contain, all of the possible n-grams. For example, the letter sequence “zqwxv” is extraordinarily rare and unlikely to appear in the training data if the data is in the English language.
Once the tree data structure has been filled, sequences which maximize the cost savings formula outlined above will be identified. Overlapping strings are handled since all subsequences of an identified sequence need to have their count in the tree data structure updated by subtracting the number of occurrences of the sequence chosen to enter into the dictionary. This may be illustrated by an example. Suppose a sequence “ABC” which occurs 10 times in the data is chosen as a dictionary entry. The count associated with the node for “AB” in the tree would then be decremented by 10, as would the count for “BC”, the count for “A”, the count for “B”, and the count for “C”. Even the node associated with the sequence “ABC” itself would be decremented by 10, thus eliminating it from the tree and from future consideration as a dictionary entry.
The address space for the dictionaries preferably will be shared, thus no flag indicating which dictionary is being used need explicitly be sent. The address space can be divided between the pre-filled dictionary and the conventional data specific or adaptive dictionary in any way seen fit by the builder of the compression dictionary. Also, the overall size of the combined compression dictionary can be determined by the builder of the compression dictionary. However, the size should be an even power of 2, since the dictionary entry numbers themselves will be encoded in a binary representation.
Those skilled in the art will appreciate that a large representative sample of uncompressed text may include documents that fall naturally into different genres of text, each with its own terminology and statistical properties. The documents of a given genre thus constitute a smaller text sample specific to that genre. Each such sample may be used to create a separate pre-filled compression dictionary tailored to its genre. Thus, any combination of the above techniques or other techniques for deriving a compression dictionary can be applied to several different types, or genres, of data such as English text, French text, computer programs written in C, computer programs written in Pascal, database files, images, and the like. Once the most frequent “words” for each type of in the context of text data or “bits” or “bit strings” in the context of image data or other ASCII-type data are discovered, a dictionary for each type of data can be created. This dictionary, in conjunction with an initially empty dictionary, to which new “words” will be added, will then be used to perform Lempel-Ziv compression using conventional techniques. For example, if English newspaper text were being compressed, the dictionary would be pre-filled with the most frequent English letter sequences, words and/or phrases found in a sample of newspaper articles.
In order to compress a document of a particular genre, the compressing computer first identifies the genre to which the document is most similar by automatic means (e.g., based on key words or clustering methods) and selects the pre-filled compression dictionary appropriate to that genre. Information identifying the selected pre-filled compression dictionary is then appended to the compressed data file. At decompression time, the decompressing computer examines the appended information in order to determine which pre-filled dictionary to use when decompressing the received data.
For this method to work as desired, the collection of sample documents must initially be partitioned into multiple genres. This may be done by any one of several methods. For example, a human or humans may decide the partition manually based on subjective or objective criteria. Alternatively, a computer may automatically scan each document for words or statistical orthographic patterns that indicate membership in a particular genre, such as the genres of “Spanish text” or “compiled computer programs.” In addition, if nothing is known about the collection of sample documents, the collection may be automatically partitioned into a reasonable set of genres by using data clustering methods well known in the literature.
Thus, additional pre-filled dictionaries can be created by the user of the compression software as desired. If the user routinely compresses data which has a standard vocabulary of its own, such as computer manuals or business documents, both of which make frequent use of words not commonly found in other forms of text, he or she can apply the dictionary creation process to a large body of this type of text and create a customized dictionary for use with these types of documents. Obviously, for decompression to be possible, the software performing the decompression must also have access to this dictionary. Similarly, if the data is not being compressed for archival purposes, but is being compressed for on-line transmission, the recipient of the compressed data must have access to this custom-made dictionary. Thus, the dictionary creation process should not be undertaken frequently, as it would require distribution and storage of a large number of dictionaries and would eliminate or, at least, reduce the added benefit of using a more appropriate dictionary for compression purposes.
During the compression process, the pre-filled dictionary which is likely to achieve the highest level of compression can be determined by performing compression on a small subset of the entire document to be compressed using each of the pre-filled dictionaries. In other words, the first N characters of data of the text to be compressed (N is arbitrary and may be chosen in advance or computed as a percentage of the document length, subject to a maximum so that identifying the dictionary does not consume too much time) are compressed using each of the putative pre-filled dictionaries. The compression using each pre-filled dictionary is computed, and the pre-filled dictionary which maximized compression is selected for compressing the entire text. In this way, the most appropriate dictionary can be used for the particular text to be compressed. In addition, the user can choose a dictionary manually if he or she is aware that the document being compressed is typical of the class of documents represented by a particular pre-filled dictionary. This would prevent the software from experimenting with various compression dictionaries and result in a time savings during the compression process. Additionally, manual selection may be necessary to allow the transmission of a document to another user who does not possess all of the pre-filled dictionaries the sender possesses or in the event of compression for archival purposes, so that the document may be decompressed at a later time by software not cognizant of all of the available compression dictionaries, such as software at a remote site.
In addition, if it is determined during the compression of the subset of characters in the data to be compressed that starting with a completely empty data compression dictionary would allow for the most compression, then the pre-filled data compression dictionary may be eliminated from the encoding of that particular data.
During the decompression process, the same modifications to the data-specific dictionary are made which were made to the data-specific dictionary built during compression. In this way, references to the data-specific dictionary within the compressed data may be expanded properly into the reconstituted text created by the data decompression system. The identity of the pre-filled dictionary being used for compression is preferably stored in the compressed data so that the decompression module is able to identify which pre-filled dictionary to use during decompression. Similarly, an indication of which variant of the Lempel-Ziv algorithm used to perform the data compression may also be stored with the compressed data as well as an indication of which policy was used during compression to flush the entries in the two dictionaries. In addition, an indication of how the dictionary address space is allocated to the dictionaries involved in compression may be stored within the compressed data—that is, what the starting and ending addresses are for the data specific dictionary as well as the starting and ending addresses for the one or more pre-filled dictionaries used in the compression. Also, if the compressed data is transmitted to a remote site, it may be necessary to send the compression dictionary in order to permit decompression at the remote site. However, this may defeat the purpose of compression unless the time savings is greater than the time needed to transmit the dictionary. On the other hand, even if the time savings does not exceed the cost of transmitting the dictionary, it might be wise to transmit the dictionary if other documents compressed using this dictionary are likely to be transmitted. Accordingly, when the compressed data is to be transmitted to a remote site, it is generally desirable to use only pre-filled compression dictionaries which are available at the recipient site.
However, if the sender determines that the optimal pre-filled dictionary is dictionary A, but compresses with dictionary B instead because the receiver does not have dictionary A, then the sender could also transmit a message to the receiver suggesting that the receiver obtain dictionary A for future use. Once the receiver gets enough such messages, the receiver will eventually request a copy of dictionary A from the sender or another server on which dictionary A is available. The receiver then places the copy of dictionary A in local long-term or medium-term storage.
As another option, the sender may compress the data file using pre-filled dictionary A, without consideration of whether the receiver has dictionary A. Then, whenever the receiver decompresses the file for the first time, it will need a copy of dictionary A. If dictionary A is not currently stored at the receiver, the receiver will obtain a copy via the network. The receiver then places a copy of dictionary A in long-term or medium-term storage, for future use, in addition to using it to decompress the file.
On the other hand, it may happen that one may desire to compress a large collection of documents at once, e.g., for archival purposes. Rather than compress the documents separately, it is desirable to exploit the similarities among the various documents, not just the similarities within each document or its similarity to a preselected corpus. In other words, if some of the documents are related to each other in some manner, many of them will contain similar strings which can be exploited to provide more compression.
For example, suppose that document A were compressed in accordance with the invention using a dictionary D pre-filled with common English language words. At the end of the compression of document A, an extended dictionary D′ remains that contains entries from D as well as strings that appeared in document A. Another document B may now be compressed using the dictionary D′ instead of the pre-filled dictionary D (or some compromise between dictionary D and dictionary D′) as the starting dictionary. If document B is quite similar to document A, this technique has been found to provide superior compression.
However, in the general case in which it is desired to compress a large number of documents, any compressed document B ought to specify the starting dictionary to use when decompressing. It may specify one of the stock pre-filled dictionaries, or it may name another compressed document, document A, and specify that the final dictionary resulting from the decompression of document A be used as the starting dictionary for the decompression of document B.
Clustering techniques may be used to determine which documents are dependent on the final dictionaries of other documents. Documents determined through clustering to be very similar to each other are compressed using one another's compression dictionaries. The information about which documents depend on one another is specified in the header of the document when it is compressed so that the decompression software can correctly decompress the document once all of the documents on which it depends are themselves decompressed.
Thus, in general, the compression rates for certain data can be improved by having a great many compression dictionaries available, each tailored for a highly specific type of text. However, there is a cost associated with storing, and sometimes transmitting, so many dictionaries. This cost can be minimized by recognizing that the different dictionaries code many of the same strings and hence overlap in content to a substantial degree. This characteristic of the dictionary coding can be taken advantage of so that pre-filled dictionaries for a plurality of genre-specific text samples can be created simultaneously from a single large corpus of documents. In this example, a special combined representation for the multiple dictionaries would be created which would take up less space than that necessary to store all of the different genre-specific dictionaries separately. In short, memory space is saved by storing a string which appears in multiple dictionaries only once rather than separately storing the entry in each respective dictionary.
Those skilled in the art will appreciate that, to a certain degree, the invention trades off compression efficiency with dictionary compactness. To illustrate this, a parameter can be defined which indicates the relative importance of compression efficiency and dictionary compactness. When the parameter is set to 1, maximum compression efficiency is obtained. In other words, when the parameter is 1, a specialized dictionary is independently built for each genre and the dictionaries for each genre are only combined to eliminate the redundant storage of duplicates as described above. On the other hand, if the parameter is reduced somewhat, then the genre dictionaries are not built wholly independently; they are deliberately constructed to improve their overlap. If the parameter is reduced all the way to 0, then maximum overlap is obtained by using the same dictionary for every genre. The parameter setting of 0 thus produces the smallest combined dictionary, but it does not take advantage of the genre-specific properties at all. At present, the inventors believe that the technique of the invention will work best when the many specialized genres have been grouped together two or three at a time into hierarchical “super-genres” analogous to the “Dewey Decimal” classification system for sample documents. Such hierarchical classification techniques may be accomplished, if necessary, by well known automatic clustering methods.
A preferred embodiment of the invention now will be described with respect to
The pre-filled dictionary space may be regarded as two separate dictionaries sharing the same address space or as one or more independent pre-filled dictionaries which may or may not be concatenated to form one larger pre-filled dictionary prior to the start of compression. When a sequence of characters to be compressed is encountered, the longest match (LM) within either the pre-filled dictionary or dictionaries or the data specific dictionary is found in either a look-up table or using a tree data structure. Once this match is found, then a dictionary entry number and the character following this sequence is transmitted. Then, the concatenation of the just transmitted dictionary entry and the character transmitted is added to the data specific dictionary. This process repeats until the data to be compressed is exhausted.
As with LZ78, the pre-filled dictionary has a fixed size, which determines the size of the dictionary entry number which needs to be transmitted to effect compression. Once the pre-filled dictionary is filled, no further entries to it are allowed. However, any of the modifications to dictionary handling outlined in the literature may be used. For example, a least recently used algorithm could be used to discard dictionary entries when new entries need to be added. Alternately, performance could be monitored and the data specific dictionary could be reset when the compression ratio drops below a certain threshold or deviates from the compression ratio achieved for the preceding sections of the data being compressed.
Once LM is determined in steps 66-70, the character following LM in the input data stream is identified at step 72 and stored as an extension (C) in the LZ78 data compression module 42. The dictionary entry number of LM (and which dictionary) is then transmitted/stored as desired at step 74. Similarly, the extension (C) is transmitted/stored as desired at step 74. Then, at step 76, the extension (C) is concatenated to the longest match (LM) and stored in the data specific dictionary 46 unless it is already full. Of course, prior art updating techniques may also be used to allow the latest entry to be inserted into the data specific dictionary 46 in place of, e.g., the least recently accessed entry. At step 78, the current pointer to the input data is then moved by the length of the longest match plus one (for the extension). Control then returns to step 58, and the compression process is repeated for the next sequence of input data until the data stream is exhausted and the compression process completed.
The method of
Decompression of a LZ78 compressed data sequence in accordance with the invention is performed in substantially the same way as with conventional LZ78 decompression techniques. All modifications to the compression dictionary made during compression are made during decompression as well. This limits the usefulness of the algorithm somewhat, as is the case with conventional LZ78, since decompressing a portion of data requires that the preceding portion be decompressed in its entirety. When performing decompression, the list of dictionary entry number and character pairs is processed one pair at a time until it is exhausted. Generally, each dictionary entry is looked up within the dictionary and the text which it refers to is displayed followed by the character in its dictionary entry and character pair. In this way, lossless compression/decompression is achieved.
Those skilled in the art will appreciate that the dictionary entry number is assigned differently during decompression depending on the variation of compression used. If conventional Lempel-Ziv compression is used, then the dictionary entry number assigned will be the next one in sequence (i.e., the first one will be numbered 0, the next 1, the next 2, etc.). Since conventional Lempel-Ziv does not make any provision for performing any special functions on the dictionary, then the dictionary will either be reset and the dictionary numbers will revert to 0 when the dictionary is full, or the dictionary will not be allowed to grow further. On the other hand, if a more complicated scheme is used to manage the dictionary, like “least recently used” (LRU), then the dictionary entry number assigned will be the same as the one just described until the dictionary fills up. At that point in time, the LRU algorithm will come into play, and the dictionary entry number of the least recently used dictionary entry will be assigned. The dictionary entry previously associated with that dictionary entry number will be removed from the dictionary.
As with the pre-filled dictionary 48 on the compression side, the dictionary 82 can be regarded as two separate dictionaries sharing the same common address bus 88. Also, a plurality of pre-filled dictionaries 90 may be used during decompression if a plurality of such pre-filled dictionaries were used during compression. All of the active elements in
As shown in
Of course, as will be appreciated by those skilled in the art, a pre-filled dictionary also will improve the compression performance of the LZ77 data compression variants as well. However, to understand how the pre-filled dictionary of the invention can be used with the LZ77 data compression variants, one must recall the differences between the LZ77 and LZ78 compression techniques noted above. As described with respect to
A pre-filled dictionary in accordance with the invention can be incorporated into the LZ77 technique in any of a number of ways. First, and most simply, a standard text can be created which contains many common strings, and this text can be kept prepended to the window throughout the compression of any document. It should be noted that the prepended text is not actually compressed and transmitted, but since it is in the window, the compression software may refer to it in the same way that it refers to text that was recently compressed and transmitted. Thus, at the beginning of the compression process, when no text would otherwise be available to serve as the window, the compression algorithm can still transmit pointers into the text containing common strings. This improves compression performance, especially at the beginning of compression, where the standard LZ77 technique must transmit many literal characters or other short strings due to lack of sufficient text in the window. As compression proceeds, the text containing common strings may optionally be gradually shortened to allow more room in the window for text from the document being compressed. The optimal ratio of pre-filled dictionary text to compressed text can be determined experimentally and will generally vary from one document genre to another. In LZ77, a fixed-sized window of text is generally used for such compression.
Another preferred, but more complicated, technique involves treating the set of pointers used by the compression algorithm as being divided, perhaps unevenly, into two classes. Pointers in the first class refer to positions in the window, which consists of part of the already compressed portion of the document, as in standard LZ77. Pointers in the second class refer to entries in a dictionary that lists frequently occurring strings. This hybrid method in effect combines some of the properties of LZ77 and LZ78. At each step, the system selects, encodes, and transmits a prefix of the remainder of the document by one of two methods. It may follow the LZ77 method, transmitting a pointer to a substring in the window, followed by the length of this substring, or when necessary, transmitting an escape code followed by a literal character. Alternatively, the system may follow the LZ78 method, transmitting a pointer to the longest dictionary entry that is a prefix of the text remaining to be compressed. The latter method is used at any step where such a dictionary entry exists, provided that the method achieves better compression (measured by the ratio of bits transmitted to text length) at that step than the LZ77 method does. One advantage of the hybrid method over LZ77 is that no length needs to be transmitted: each dictionary entry has a fixed length that is permanently stored (or otherwise recorded) in the dictionary. The hybrid method is also more efficient than LZ78 in that dictionary entries are not accompanied by literal characters.
At step 126, the costs of transmitting PDLM as a dictionary entry number are compared with the costs of transmitting WLM as a pointer and a length, and, optionally, a one character extension EXT, or as a literal. Specifically, it is determined at step 126 which technique saves the most compared to transmitting the data uncompressed. Generally, this may be accomplished by computing the compression ratio for each technique. Based on the results of step 126, the technique which saves the most of the PDLM and the WLM is then transmitted/stored as desired at step 128 (with or without EXT). Finally, at step 130, the current pointer to the input data is then moved by the length of the longest match or the length of the longest match plus one (if an extension EXT is used). Control then returns to step 112, and the compression is repeated for the next sequence of input data at the current pointer until the data stream is exhausted and the compression process completed.
On the other hand, if it is determined at step 138 that the entry in the compressed data stream is a WLM, then the entry at the pointer is used at step 146 to lookup the entry in the previous window of text pointed to by that entry. That value is stored in the LZ77/LZ78 decompression module as E. If desired, the decompressed data is also displayed to the user. At step 148, if the optional extension EXT was sent (as determined prior to running the algorithm), the character in the input compressed data stream following the current entry is then stored in the LZ77/LZ78 decompression module as C. If desired, that character is also displayed to the user. At step 150, the character C and the dictionary entry E are concatenated and stored in the LZ77/LZ78 data decompression module as ADD. Once again, if desired, the resulting concatenation is also displayed to the user.
At the completion of either step 144 or step 150, the current pointer to the input data stream is advanced to the beginning of the next entry in the data stream at step 152. Control then returns to step 134, and the decompression process is repeated for the next sequence of input data until the input compressed data stream is exhausted. The decompression process is then complete.
As illustrated in
Any of the standard improvements to LZ77, such as applying some form of compression to the literals, lengths, and addresses being transmitted, seeding the dictionary or window so that no literals need to be transmitted, and the like, may be equally well applied to any of the above-mentioned embodiments which are extensions of the LZ77 compression technique. Also, as described above, different dictionaries (or different texts to prepend to the window) can be used for different genres of data being transmitted. An initial code number sent at the beginning of each message would then indicate which dictionary or prepended text was being used (e.g., the one for English prose, for computer software, or for business transaction forms), or a special code would be transmitted indicating that the dictionary or prepended text has been specially derived from a previously transmitted text.
In a sample implementation of the techniques of the invention, a set of pre-filled data compression dictionaries is distributed to a number of servers that regularly exchange compressed documents amongst themselves. As described above, the dictionaries may be combined before distribution to save transmission costs. Periodically, a new set of pre-filled dictionaries is distributed to supplement the old set of pre-filled dictionaries. The new set might include dictionaries for new genres, as well as more up-to-date dictionaries for old genres which reflect terminology changes. Ideally, the new dictionaries are constructed so that they overlap substantially with the old dictionaries, whereby the combined representation for all the new and old dictionaries does not differ much from the combined representation for the old dictionaries alone. A file listing the changes between the two representations of a pre-filled dictionary is then distributed. Upon receiving this file, each of the servers modifies its combined dictionary representation in order to add the new dictionaries. The old dictionaries may now be either retained in the combined representation or deleted from it, depending on whether or not they will be needed in order to decompress old files. Such a technique will lead to substantially reduced data transmission and storage costs, particularly between two servers which often exchange files of the same genre.
For example, a simulation of the invention was performed using half of a corpus of data, where the pre-filled dictionary was learned from the other half. For comparison purposes, the data was compressed using “gzip”, a simulation of “gzip”, and a simulation of “gzip” with a pre-filled dictionary in accordance with the invention. A simulation was used since “gzip” would have to be modified to be used with a pre-filled dictionary in accordance with the invention. The simulation of “gzip” was tested without the pre-filled dictionary merely to verify that the simulation approximated the performance of “gzip” before the addition of the pre-filled dictionary. The following results were obtained:
Although numerous embodiments of the invention and numerous extensions of the inventive concept have been described above, those skilled in the art will readily appreciate that many additional modifications are possible in the exemplary embodiment without materially departing from the novel teachings and advantages of the invention.
For example, as noted above, the pre-filled dictionaries of the invention may be formed in any of a number of ways and may be used with any of a number of variations of the basic Lempel-Ziv compression technique. Moreover, compression performance may be optimized by separately monitoring or tracking the compression performance of the pre-filled data compression dictionary and the data specific compression dictionary. For example, if the pre-filled data compression dictionary is not providing better data compression than the data specific compression dictionary, then in the event that the data specific dictionary becomes full and needs to be reset, the pre-filled data compression dictionary is replaced by the data specific dictionary and the data specific dictionary is reset. On the other hand, in the event that the pre-filled data compression dictionary is more useful for compression, it is retained and the data specific compression dictionary is reset. To determine which dictionary is providing better compression than another dictionary, the savings associated with each of the dictionaries is maintained while compression is proceeding. The length of each of the strings compressed using each of the dictionaries is also kept, as is the amount of data resulting from the portion of the compression effected by each dictionary. A threshold is chosen below which the pre-filled dictionary is said to be yielding so little improvement that allowing the data-specific dictionary to grow further would be better. This is done both globally (for the entire document compressed so far) as well as locally by maintaining several different sets of these data. In effect, this process allows the entire address space afforded to the pre-filled data compression dictionary and the data specific compression dictionary to be utilized by the data specific dictionary when the pre-filled dictionary is not providing an improvement to the compression process.
Moreover, a least recently used (LRU) method of the type described by Storer in U.S. Pat. No. 4,876,541 makes it unnecessary to keep the two dictionaries separate, for the least recently used entry is always discarded, regardless of which dictionary it came from. It should be noted that, in this context, “used” includes reading or writing a code as part of a compressed data stream, but the entry of the code into the dictionary on the first appearance of its string may or may not be counted as a “use.”
In addition, multiple pre-filled data compression dictionaries may be used in the compression process at the expense of reducing the size of the data specific dictionary. In the most extreme case, the data specific dictionary could be eliminated entirely and the entire address space normally shared by the pre-filled dictionary and the data specific dictionary would be shared by the relevant pre-filled dictionaries. Of course, in this case, all character sequences, words, and phrases in the text to be compressed would need to be present in one or more pre-filled data compression dictionaries, which is quite possible when the pre-filled dictionary or dictionaries are large enough to contain all combinations of characters likely to occur in the text to be compressed. Moreover, even if all combinations of characters are not found in the pre-filled data compression dictionary, combinations not found could simply be stored uncompressed and/or the pre-filled dictionary could be updated during compression to include that character combination. On the other hand, an algorithm could be used to determine whether the entire character set is present in the pre-filled data compression dictionary created by the software described with respect to
As another modification to the preferred embodiments of the invention, the codeword addresses of the dictionary entries may themselves be encoded using a technique such as Huffman coding so that the more frequently used addresses can be represented using fewer bits. Similarly, the extension characters may be encoded using a variable length coding, such as Huffman coding, to improve performance. In such a case, it is not necessary to limit the size of the data specific dictionary, and hence unnecessary to ever flush the data specific dictionary, for the Lempel-Ziv codewords are just integers. These integers are then encoded by variable-length bit sequences. Moreover, if the distribution of these integer codewords varies through the document, so that its local entropy is lower than its global entropy, then adaptive versions of the variable-length coding scheme, such as adaptive Huffman coding or adaptive arithmetic coding, should be used.
In accordance with this technique, the size of the code being transmitted using the encoding for codewords is known and is based on the number of codewords in the dictionary. A list of recently used codewords can be maintained and at such a time as the recently used codewords form a small subset of the entire space of codewords involved in the compression of codewords scheme, compression of these codewords can be restarted. That is, a new dictionary to compress codewords can be created. This new dictionary could be constructed from recently used codewords on the assumption that recently used codewords are more likely to be used again locally, while those which have not been recently used are less likely to be used again.
One skilled in the art may also choose to store a data specific compression dictionary created during compression of a particular document as a pre-filled data compression dictionary for use in the compression of other related documents. On the other hand, the data specific dictionary may be analyzed to determine whether particular entries should be included in other pre-filled data compression dictionaries.
Accordingly, all such modifications are intended to be included within the scope of this invention as defined in the following claims.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US4386416||Jun 2, 1980||May 31, 1983||Mostek Corporation||Data compression, encryption, and in-line transmission system|
|US4464650||Aug 10, 1981||Aug 7, 1984||Sperry Corporation||Apparatus and method for compressing data signals and restoring the compressed data signals|
|US4558302||Jun 20, 1983||Dec 10, 1985||Sperry Corporation||High speed data compression and decompression apparatus and method|
|US4672679||Aug 16, 1983||Jun 9, 1987||Wang Laboratories, Inc.||Context redundancy text compression|
|US4701745||Mar 3, 1986||Oct 20, 1987||Ferranti, Plc||Data compression system|
|US4814746||Aug 11, 1986||Mar 21, 1989||International Business Machines Corporation||Data compression method|
|US4847619||Oct 19, 1987||Jul 11, 1989||Hewlett-Packard Company||Performance-based reset of data compression dictionary|
|US4876541||Oct 15, 1987||Oct 24, 1989||Data Compression Corporation||Stem for dynamically compressing and decompressing electronic data|
|US4881075||Oct 15, 1987||Nov 14, 1989||Digital Equipment Corporation||Method and apparatus for adaptive data compression|
|US4906991||Apr 29, 1988||Mar 6, 1990||Xerox Corporation||Textual substitution data compression with finite length search windows|
|US4988998||Sep 5, 1989||Jan 29, 1991||Storage Technology Corporation||Data compression system for successively applying at least two data compression methods to an input data stream|
|US5001478||Dec 28, 1989||Mar 19, 1991||International Business Machines Corporation||Method of encoding compressed data|
|US5003307||Oct 6, 1989||Mar 26, 1991||Stac, Inc.||Data compression apparatus with shift register search means|
|US5016009||Jan 13, 1989||May 14, 1991||Stac, Inc.||Data compression apparatus and method|
|US5023610||Jun 13, 1990||Jun 11, 1991||Cordell Manufacturing, Inc.||Data compression method using textual substitution|
|US5049881||Jun 18, 1990||Sep 17, 1991||Intersecting Concepts, Inc.||Apparatus and method for very high data rate-compression incorporating lossless data compression and expansion utilizing a hashing technique|
|US5058137||Jul 31, 1989||Oct 15, 1991||North American Philips Corporation||Lempel-Ziv decoder|
|US5087913||Aug 27, 1990||Feb 11, 1992||Unisys Corporation||Short-record data compression and decompression system|
|US5126739||Nov 27, 1990||Jun 30, 1992||Stac Electronics||Data compression apparatus and method|
|US5140321||Sep 4, 1991||Aug 18, 1992||Prime Computer, Inc.||Data compression/decompression method and apparatus|
|US5153591||Jul 4, 1989||Oct 6, 1992||British Telecommunications Public Limited Company||Method and apparatus for encoding, decoding and transmitting data in compressed form|
|US5155484||Sep 13, 1991||Oct 13, 1992||Salient Software, Inc.||Fast data compressor with direct lookup table indexing into history buffer|
|US5179378||Jul 30, 1991||Jan 12, 1993||University Of South Florida||Method and apparatus for the compression and decompression of data using Lempel-Ziv based techniques|
|US5243341||Jun 1, 1992||Sep 7, 1993||Hewlett Packard Company||Lempel-Ziv compression scheme with enhanced adapation|
|US5262776||Mar 21, 1990||Nov 16, 1993||Siemens Aktiengesellschaft||Process for use with lempel/gin encoding|
|US5373290||Dec 23, 1992||Dec 13, 1994||Hewlett-Packard Corporation||Apparatus and method for managing multiple dictionaries in content addressable memory based data compression|
|US5455576||Feb 7, 1994||Oct 3, 1995||Hewlett Packard Corporation||Apparatus and methods for Lempel Ziv data compression with improved management of multiple dictionaries in content addressable memory|
|US5485526||Jan 14, 1994||Jan 16, 1996||Hewlett-Packard Corporation||Memory circuit for lossless data compression/decompression dictionary storage|
|US5561421||Jul 28, 1994||Oct 1, 1996||International Business Machines Corporation||Access method data compression with system-built generic dictionaries|
|US5590317 *||May 27, 1993||Dec 31, 1996||Hitachi, Ltd.||Document information compression and retrieval system and document information registration and retrieval method|
|US5872530 *||Jan 28, 1997||Feb 16, 1999||Hitachi, Ltd.||Method of and apparatus for compressing and decompressing data and data processing apparatus and network system using the same|
|EP0286719A2||Oct 2, 1987||Oct 19, 1988||International Business Machines Corporation||Text compression and expansion method and apparatus|
|JPH0628149A||Title not available|
|JPH0764765A||Title not available|
|JPH0869370A||Title not available|
|JPH05241775A||Title not available|
|JPH05241777A||Title not available|
|JPH06161705A||Title not available|
|JPH07152533A||Title not available|
|JPH08162874A||Title not available|
|JPH08505008A||Title not available|
|1||"Adaptive Dictionary Encoders: Ziv-Lempel Coding", Section 8.3.1 LZ 77, Jan. 2, 1990, 218-227, XP-002132837.|
|2||Bell, et al. "Text compression", 1990, Englewood Cliffs, New Jersey, Prentice Hall.|
|3||Bell, et al., "Adaptive dictionary encoders: Ziv-lempel coding", Dictionary Techniques, Chap 8, 214-234.|
|4||Bell, T.C., "Better OPM/L text compression", IEEE Trans. On Comm., 1986, COM-34(12), 1176-1182.|
|5||Bentley, J.L. et al., "A locally adaptive data compression scheme", Comm of ACM, 1986, 29(4), 320-330.|
|6||Huffman, D.A., "A method for the construction of minimum-redundancy codes", Proceedings of the I.R.E., 1952, 40, 1098-1101.|
|7||Langdon Jr., G.G., "A note on the Ziv-Lempel model for compressing individual sequences", IEEE Trans. On Information Theory, 1983, 29(2), 284-287.|
|8||Marcus, et al., "Building a large annoted corpus of English: The Penn Treebank", Computational Linguistics, 1993, 19(2), 313-330.|
|9||Miller, V.S. et al., "Variations on a theme by Ziv and Lempel", Combinatorial Algorithms on Words, NATO ASI Series, Apostolico, A and Galil, Z., eds., 1985, vol. F12, 131-140.|
|10||Rodeh, et al., "Linear algorithm for data compression via string matching", J. of the Assoc. for Computing Machinery, 1981, 28(1), 16-24.|
|11||Storer, J.A. et al., "Data compression via textual substitution", J. of the Assoc. for Computing Machinery, 1982, 29, 928-951.|
|12||Welch, T.A., "A technique for high-performance data compression", IEEE, 1984, 8-19.|
|13||Zipf, G., "Human Behavior and the principle of least effort", 1949.|
|14||Ziv, J. et al., "A universial algorithm for sequential data compression", IEEE Trans. On Information Theory, 1977, IT-23(3), 337-343.|
|15||Ziv, J. et al., "Compression of individual sequences via variable-rate coding", IEEE Trans. On Information Theory, 1978, IT-24(5), 530-536.|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US7982636 *||Aug 20, 2009||Jul 19, 2011||International Business Machines Corporation||Data compression using a nested hierachy of fixed phrase length static and dynamic dictionaries|
|US8355585 *||May 12, 2009||Jan 15, 2013||Red Hat Israel, Ltd.||Data compression of images using a shared dictionary|
|US8542135||Nov 24, 2011||Sep 24, 2013||International Business Machines Corporation||Compression algorithm incorporating automatic generation of a bank of predefined huffman dictionaries|
|US8566323||Dec 29, 2009||Oct 22, 2013||Novell, Inc.||Grouping and differentiating files based on underlying grouped and differentiated files|
|US8593308||May 31, 2012||Nov 26, 2013||International Business Machines Corporation||Method of accelerating dynamic Huffman decompaction within the inflate algorithm|
|US8601368 *||Jan 13, 2009||Dec 3, 2013||Canon Kabushiki Kaisha||Processing method and device for the coding of a document of hierarchized data|
|US8610604||Nov 24, 2011||Dec 17, 2013||International Business Machines Corporation||Compression algorithm incorporating a feedback loop for dynamic selection of a predefined Huffman dictionary|
|US8610606||Nov 24, 2011||Dec 17, 2013||International Business Machines Corporation||Compression algorithm incorporating dynamic selection of a predefined huffman dictionary|
|US8676858 *||Jan 8, 2010||Mar 18, 2014||Novell, Inc.||Grouping and differentiating volumes of files|
|US8791843||Oct 15, 2012||Jul 29, 2014||Lsi Corporation||Optimized bitstream encoding for compression|
|US8811611||Oct 8, 2009||Aug 19, 2014||Novell, Inc.||Encryption/decryption of digital data using related, but independent keys|
|US8832103||Apr 13, 2010||Sep 9, 2014||Novell, Inc.||Relevancy filter for new data based on underlying files|
|US8874529 *||Mar 16, 2009||Oct 28, 2014||Bert A. Silich||User-determinable method and system for manipulating and displaying textual and graphical information|
|US8874578||Dec 30, 2009||Oct 28, 2014||Novell, Inc.||Stopping functions for grouping and differentiating files based on content|
|US8933829 *||Sep 23, 2013||Jan 13, 2015||International Business Machines Corporation||Data compression using dictionary encoding|
|US8983959||Dec 30, 2009||Mar 17, 2015||Novell, Inc.||Optimized partitions for grouping and differentiating files of data|
|US9035809||Oct 15, 2012||May 19, 2015||Seagate Technology Llc||Optimizing compression engine throughput via run pre-processing|
|US9053120||Dec 15, 2009||Jun 9, 2015||Novell, Inc.||Grouping and differentiating files based on content|
|US9160611 *||Apr 9, 2010||Oct 13, 2015||Webroot Inc.||System and method for performing longest common prefix strings searches|
|US9231615||Oct 24, 2012||Jan 5, 2016||Seagate Technology Llc||Method to shorten hash chains in Lempel-Ziv compression of data with repetitive symbols|
|US9298722||Sep 28, 2009||Mar 29, 2016||Novell, Inc.||Optimal sequential (de)compression of digital data|
|US9348835||Sep 30, 2014||May 24, 2016||Novell, Inc.||Stopping functions for grouping and differentiating files based on content|
|US9363339||Mar 15, 2014||Jun 7, 2016||Hughes Network Systems, Llc||Staged data compression, including block level long range compression, for data streams in a communications system|
|US9369721 *||Sep 14, 2012||Jun 14, 2016||Red Hat Israel, Ltd.||Data compression of images using a shared dictionary|
|US9390098||Oct 8, 2009||Jul 12, 2016||Novell, Inc.||Fast approximation to optimal compression of digital data|
|US9438413||Mar 18, 2011||Sep 6, 2016||Novell, Inc.||Generating and merging keys for grouping and differentiating volumes of files|
|US9479383||Mar 16, 2014||Oct 25, 2016||Hughes Network Systems, Llc||Data compression for priority based data traffic, on an aggregate traffic level, in a multi stream communications system|
|US9513813||Dec 18, 2015||Dec 6, 2016||International Business Machines Corporation||Determining prefix codes for pseudo-dynamic data compression utilizing clusters formed based on compression ratio|
|US9558241 *||Oct 13, 2015||Jan 31, 2017||Webroot Inc.||System and method for performing longest common prefix strings searches|
|US9680500||Jun 7, 2016||Jun 13, 2017||Hughes Network Systems, Llc||Staged data compression, including block level long range compression, for data streams in a communications system|
|US9716734||Jun 10, 2014||Jul 25, 2017||Hughes Network Systems, Llc||System and method for long range and short range data compression|
|US20070162957 *||Mar 2, 2007||Jul 12, 2007||Andrew Bartels||Methods, systems and devices for securing supervisory control and data acquisition (SCADA) communications|
|US20070255570 *||Jan 22, 2007||Nov 1, 2007||Annaz Fawaz Y||Multi-platform visual pronunciation dictionary|
|US20090183067 *||Jan 13, 2009||Jul 16, 2009||Canon Kabushiki Kaisha||Processing method and device for the coding of a document of hierarchized data|
|US20100235331 *||Mar 16, 2009||Sep 16, 2010||Silich Bert A||User-determinable method and system for manipulating and displaying textual and graphical information|
|US20100274786 *||Apr 9, 2010||Oct 28, 2010||Brightcloud Inc.||System And Method For Performing Longest Common Prefix Strings Searches|
|US20100293248 *||May 12, 2009||Nov 18, 2010||Yaniv Kamay||Data compression of images using a shared dictionary|
|US20110013777 *||Oct 8, 2009||Jan 20, 2011||Teerlink Craig N||Encryption/decryption of digital data using related, but independent keys|
|US20110016098 *||Jan 8, 2010||Jan 20, 2011||Teerlink Craig N||Grouping and differentiating volumes of files|
|US20110016124 *||Dec 30, 2009||Jan 20, 2011||Isaacson Scott A||Optimized Partitions For Grouping And Differentiating Files Of Data|
|US20110016135 *||Nov 11, 2009||Jan 20, 2011||Teerlink Craig N||Digital spectrum of file based on contents|
|US20110016136 *||Dec 29, 2009||Jan 20, 2011||Isaacson Scott A||Grouping and Differentiating Files Based on Underlying Grouped and Differentiated Files|
|US20110016138 *||Dec 15, 2009||Jan 20, 2011||Teerlink Craig N||Grouping and Differentiating Files Based on Content|
|US20110043387 *||Aug 20, 2009||Feb 24, 2011||International Business Machines Corporation||Data compression using a nested hierachy of fixed phrase length static and dynamic dictionaries|
|US20110173166 *||Mar 18, 2011||Jul 14, 2011||Teerlink Craig N||Generating and merging keys for grouping and differentiating volumes of files|
|US20120130965 *||Jun 1, 2010||May 24, 2012||Yun Sik Oh||Data compression method|
|US20130011064 *||Sep 14, 2012||Jan 10, 2013||Red Hat Israel, Ltd.||Data compression of images using a shared dictionary|
|US20130018932 *||Jul 12, 2011||Jan 17, 2013||Hughes Network Systems, Llc||System and method for long range and short range data compression|
|US20140085115 *||Sep 23, 2013||Mar 27, 2014||International Business Machines Corporation||Data compression using dictionary encoding|
|US20160055213 *||Oct 13, 2015||Feb 25, 2016||Webroot Inc.||System and method for performing longest common prefix strings searches|
|U.S. Classification||708/203, 341/106, 358/426.01, 708/212, 341/51, 341/87, 380/217, 341/50, 341/95|
|International Classification||G06T9/00, G06F5/00, H03M7/30, H03M7/40, H03M7/00, G06F7/00, G06F15/00|
|Cooperative Classification||H04N19/93, H03M7/3088, H04N19/91, H04N19/13, H03M7/3086, G06T9/005|
|European Classification||H04N7/26Z12, H04N7/26A4V, G06T9/00S, H03M7/30Z2, H03M7/30Z1|
|Mar 18, 2011||FPAY||Fee payment|
Year of fee payment: 12
|Mar 18, 2011||SULP||Surcharge for late payment|
Year of fee payment: 11