US20060015326A1

Movatterモバイル変換

Info

Publication number: US20060015326A1
Application number: US11/180,153
Authority: US
Inventors: Shinsuke Mori; Daisuke Takuma
Original assignee: International Business Machines Corp
Current assignee: International Business Machines Corp
Priority date: 2004-07-14
Filing date: 2005-07-13
Publication date: 2006-01-19
Also published as: US20080228463A1; US7917350B2; JP2006031295A; JP4652737B2

Abstract

Calculates a word n-gram probability with high accuracy in a situation where a first corpus), which is a relatively small corpus containing manually segmented word information, and a second corpus, which is a relatively large corpus, are given as a training corpus that is storage containing vast quantities of sample sentences. Vocabulary including contextual information is expanded from words occurring in first corpus of relatively small size to words occurring in second corpus of relatively large size by using a word n-gram probability estimated from an unknown word model and the raw corpus. The first corpus (word-segmented) is used for calculating n-grams and the probability that the word boundary between two adjacent characters will be the boundary of two words (segmentation probability). The second corpus (word-unsegmented), in which probabilistic word boundaries are assigned based on information in the first corpus (word-segmented), is used for calculating a word n-grams.

Description

FIELD OF THE INVENTION

The present invention relates to recognition technology in natural language processing, and improving the accuracy of recognition in natural language processing by using a corpus, in particular by effectively using a corpus to which segmentation is not applied.

BACKGROUND ART

Along with the progress of recognition technology for natural language, various techniques, including kana-kanji conversion, spelling checking (character error correction), OCR, and speech recognition techniques, have achieved a practical-level predication capability. At present, most of the methods for implementing these techniques with high accuracy are based on probabilistic language models and/or statistical language models. Probabilistic language models are based on the frequency of occurrence of words or characters and require a collection of a huge number of texts (corpus) in an application field.

The following documents are considered:

- [Non-patent Document 1] “Natural Language Processing: Fundamentals and applications”, edited by Hozumi Tanaka, 1999, Institute of Electronics, Information and Communication Engineers
- [Non-patent Document 2] W. J. Teahan, and John G. Cleary, 1996, “The entropy of English using ppm-based models”, In DCC.
- [Non-patent Document 3] Leo Breiman, Jerome H. Friedman, Richard A. Olshen, and Chales J. Stone, 1984, Classification and Regression Trees, Chapman & Hall, Inc.
- [Non-patent Document 4] Masaaki Nagata, “A Self-Organizing Japanese Word Segmenter using Heuristic Word Identification and Re-estimation”, 1997

In most speech recognition systems, the most probable character string is selected from among a number of candidates by referring to a probabilistic language model as well as an acoustic model. In spell checking (character error correction), unnatural character strings and their correction candidates are listed based on the likelihood of a probabilistic language model.

Because a practical model treats a word as a unit, it is required that a corpus be provided with information about word boundaries. In order to determine word boundaries, an operation such as segmentation or tagging is performed.

Automatic word segmentation methods have been already known. However, the existing automatic word segmentation systems provide low accuracies in fields such as the medical field, where many technical terms are used. To manually correct the results of automatic word segmentation, the operator needs to have knowledge of technical terms in the application field, and typically, a minimum of tens of thousands sentences are required in order to achieve recognition sufficiently accurate enough for practical use.

In training using a corpus in an application field, it is generally difficult to obtain a huge corpus segmented and tagged manually for the application field, taking much time and cost and thus making it difficult to develop a system in a short period.

Although information segmented into words in a field (for example in the medical field) may works in processing the language in that field, there is no assurance that the information will work also in another application field (for example in the economic field, which is completely different from the medical field). In other words, a correct corpus segmented and tagged in a field may be definitely correct in that field, but may not necessarily correct in other fields because the segmented and/or tagged corpus has been fixed by segmentation and/or tagging.

In this regard, there are many techniques in the background art that are pursuing efficiency and accuracy in word segmentation in Asian languages. However, all of these techniques are aiming to predetermine word boundaries in word segmentation fixedly.

Taking Japanese out of the Asian languages as an example, word information required for analyzing Japanese text relates to the structure of word spelling, which is the information regarding the character configuration (representation form) and pronunciation of entry words, including “spelling information”, “pronunciation information”, and “morphological information”. These items of information may provide important clues mainly in extracting candidate words from Japanese text in morphological analysis.

Although there is no clear definition of the term “word”, attention is directed to two elements of the “word” herein, “spelling” and “pronunciation” and two words are regarded as the same words if and only if they have the same spelling (characters) and pronunciation. Isomorphic words just having the same spelling (characters) or homophonic words just having the same pronunciation are regarded as different words. The spelling of a word is involved in identifying a morphological characteristic and the pronunciation is involved in identifying a phonemic characteristic.

Hence, the Japanese words composed of two Chinese characters
(reporter),
(train),
(return to the office), and
(charity) all have the same pronunciation
(kisha) but different spellings (characters), whereby they are different words. The “word” is symbolized in the computer, in which the correspondence between the symbol as the spelling (character) and the symbol as its meaning is registered. Japanese is one kind of agglutinative language, and has an extremely high word formation power, whereby care must be taken in registering words in the computer as “dictionary”. The pronunciation is given in a string of input symbols (e.g., katakana in Japanese, Roman character representation of katakana) in the computer.
A word is registered in the computer by a method of registering all the possible spellings (characters), or collecting and registering the spellings having high use frequency, a method of registering only typical spellings, and searching for a word in combination with its pronunciation, or a method of providing various sorts of character conversion table apart from the dictionary and investigating the correspondence with headwords, or a combination of these methods.
A plain example for correcting the result of automatic word segmentation is given below. For example, for the pronunciation of
(ha-ki-mo-no), there are two corresponding spellings. One is the word
(footwear) and the other is a sequence of two words
(postpositional particle) and
(kimono). These two spellings are associated with the pronunciation “ha-ki-mo-no”. If there is an occurrence of this pronunciation and the spelling resulting from word segmentation is considered to be improper, the spelling is corrected by re-segmenting. Unlike English, Japanese language does not have a space between words (write with a space between words), therefore an expert must determine word boundaries from the context around an sample sentence, based on the knowledge of technical terms.
As an example indicates that the word
(footwear) is assigned to the pronunciation
(ha-ki-mo-no), it will be found that the word needs to be correctly recognized using the knowledge of vocabulary. Therefore, there is a demand for a method for increasing the accuracy, making effective use of the corpus without segmentation.
For all processes in natural language processing, conversion of character strings or speech data into a string of morphemes is a prerequisite. However, in Asian languages such as Japanese, it is difficult to morphologically analyze even written text because, unlike English text, text in such languages is written without a space between words. Therefore, as part of the accuracy problem described above, there is the need for accurately obtaining candidate morpheme strings (x) when input data (y) such as a hiragana character string, a katakana character string, or speech data is given.
In a statistical approach, this can be formulated as the maximization problem of P(x|y) and Bayes' theorem can be used to decompose it into two models of maximizing, P(y|x) and P(x), as shown in the right-hand side of the equation
P(x|Y)=P(y|x)P(x)/P(y)
where P(y) is a constant as y is given. The model of P(x) is independent of the type of input data (whether it is a symbol string, character string, or speech data), hence called a “language model”. One of the most commonly used probabilistic language models is a word n-gram model.
<Conventional Art Relating to the Use of Unsegmented Corpus>
As conventional art there are methods in which the result of segmentation of an unsegmented corpus based on training with a segmented corpus is used:

- (a) Counting n-grams with weight by using candidate segmentations,
- (b) Using only 1-best of the candidates resulting from automatic segmentation, and
- (c) Using n-best of the candidates resulting from automatic segmentation.

However, methods (a) and (c) require high computational costs for bi-gram and higher and are unrealistic. Advantages of the present invention over method (b) will be descried later with respect to experiments.

SUMMARY OF THE INVENTION

In light of the foregoing, a general aspect of the present invention can be summarized as follows. The invention provides that a word n-gram probability is calculated with high accuracy in a situation where:

- (a) a first corpus (word-segmented), which is a relatively small corpus containing manually segmented word information, and
- (b) a second corpus (word-unsegmented), which is a relatively large corpus containing raw information are given as a training corpus that is storage containing vast quantities of sample sentences.

Vocabulary including contextual information is expanded from words occurring in the first corpus (word-segmented) of relatively small size to words occurring in the second corpus (word-unsegmented) of relatively large size by using a word n-gram probability estimated from an unknown word model and the raw corpus.
<Usage of Word Segmented Corpus>
A first corpus (word-segmented) is used for calculating n-grams and the probability that the boundary between two adjacent characters will be the boundary of two words (segmentation probability). A second corpus (word-unsegmented), in which probabilistic word boundaries are assigned based on information in the first corpus (word-segmented), is used for calculating a word n-gram.
<Calculation of Probabilistic Word Boundaries>
In the second corpus (word-unsegmented), the segmentation probability calculated from the first corpus (word-segmented) is assigned between characters.
<Character-Wise Unknown-Word Model>
The correspondences between each character and its pronunciations are modeled. Thereby, a kana-kanji conversion model for an unknown word is proposed.
Advantages of the invention include that with a word boundary probability estimating device, a probabilistic language model building device, a kana-kanji converting device, and a method therefor according to the present invention as described above, existing vocabulary/linguistic models concerning the first corpus (word-segmented) are combined with vocabulary/linguistic models built by probabilistically segmenting the second corpus (word-unsegmented), which is a raw corpus, whereby the accuracy of recognition in natural language processing can be improved. Because the capability of a probabilistic language model can be improved simply by collecting sample sentences in a field of interest, application of the present invention to fields for which language recognition technique corpuses not provided can be supported.

BRIEF DESCRIPTION OF THE DRAWINGS

For a more complete understanding of the present invention and the advantages there of, reference is now made to the following description taken in conjunction with the accompanying drawings, in which:
FIG. 1 illustrates a configuration of a kana-kanji converting device according to the present invention;
FIG. 2 shows a configuration of a kana-kanji converting device and a kana-kanji conversion program that implement a kana-kanji conversion method according to the present invention;
FIG. 3 shows details of the kana-kanji converting device shown inFIG. 2;
FIG. 4 shows a probabilistic language model building device;
FIG. 5 is a flowchart of a process for calculating the expected frequency of a character string as a word;
FIG. 6 shows the probability that a word boundary will exist between characters when a character of a particular type is followed by the character of different or the same type;
FIG. 7 is a flowchart of a process for calculating word n-gram probability;
FIG. 8 is a flowchart of an operation of the kana-kanji converting device (language decoding section) shown inFIG. 2 in which the weights of the first corpus (word-segmented) and the second corpus (word-unsegmented) are adjusted with λ;
FIG. 9 shows details of the corpuses used in the experiments to demonstrate advantages of introducing the proposed method in writing applicable documents;
FIG. 10 shows the results of the experiments on the corpuses shown inFIG. 9;
FIG. 11 shows details of the corpuses for calculating the precision and the recall as well for evaluating the performance of word segmentation; and
FIG. 12 shows models used inFIG. 11 and the results.

DESCRIPTION OF SYMBOLS

1 . . . Kana-kanji converting device
10 . . . CPU
12 . . . Input device
14 . . . Display device
16 . . . Storage device
18 . . . Recording medium
2 . . . Kana-kanji conversion program
22 . . . Language decoding section
30 . . . Base form pool
300 . . . Vocabulary dictionary
302 . . . Character dictionary
32 . . . Language model
320 . . . First corpus (word-segmented)
322 . . . Second corpus (word-unsegmented)

DETAILED DESCRIPTION OF THE INVENTION:

The present invention provides that a word n-gram probability is calculated with high accuracy in a situation where:

Vocabulary including contextual information is expanded from words occurring in the first corpus (word-segmented) of relatively small size to words occurring in the second corpus (word-unsegmented) of relatively large size by using a word n-gram probability estimated from an unknown word model and the raw corpus.
<Usage of Word Segmented Corpus>
The first corpus (word-segmented) is used for calculating n-grams and the probability that the boundary between two adjacent characters will be the boundary of two words (segmentation probability). The second corpus (word-unsegmented), in which probabilistic word boundaries are assigned based on information in the first corpus (word-segmented), is used for calculating a word n-gram.
<Calculation of Probabilistic Word Boundaries>
In the second corpus (word-unsegmented), the segmentation probability calculated from the first corpus (word-segmented) is assigned between characters.
<Character-Wise Unknown-Word Model>
The correspondences between each character and its pronunciations are modeled. Thereby, a kana-kanji conversion model for an unknown word is proposed.
Thus, with a word boundary probability estimating device, a probabilistic language model building device, a kana-kanji converting device, and a method therefor according to the present invention as described above, existing vocabulary/linguistic models concerning the first corpus (word-segmented) are combined with vocabulary/linguistic models built by probabilistically segmenting the second corpus (word-unsegmented), which is a raw corpus, whereby the accuracy of recognition in natural language processing can be improved. Because the capability of a probabilistic language model can be improved simply by collecting sample sentences in a field of interest, application of the present invention to fields for which language recognition technique corpuses not provided can be supported.
Furthermore, even for words that do not appear in the first corpus (word-segmented), candidates can be listed by referring to character-wise frequency information. Moreover, if the word n-gram probability of the second corpus (word-unsegmented) is used, which is a raw corpus, contextual information about unknown words can be used as well.
The following is an example of an advantageous embodiment:
(Operation of Kana-Kanji Converting Device)
Operation of a kana-kanji converting device1 (inFIGS. 1 and 2) to which the present invention can be applied will be described below.
(Kana-Kanji Converting Device1)
A kana-kanji converting device1 to which the present invention can be applied will be described below.FIG. 1 illustrates an exemplary configuration of the kana-kanji converting device1 to which the present invention can be applied. As shown inFIG. 1, the kana-kanji converting device1 according to the present invention includes aCPU10 including a microprocessor, memory, their peripheral circuits, and the like (not shown),input device12 such as a mouse and a keyboard, adisplay device14 such as a CRT display, and astorage device16 such as an HDD drive, a DVD drive, and a CD drive.
That is, the kana-kanji converting device1 has a typical hardware configuration, executes a kana-kanji conversion program2 (which will be described later with reference toFIG. 2), which is supplied in the form of a program recorded on arecording medium18, such as a DVD, CD-ROM, or CD-RAM, converts input symbols strings which are input from the keyboard120 in theinput devices12 and converted into a digital format to create text data, and records the text data on arecording medium18 placed in thestorage device16 or display the text data on thedisplay device14. The kana-kanji converting device can be construed as a smaller unit than the kana-kanji converting device, such as a word boundary probability estimating device, and a probabilistic language model building device (seeFIG. 4). The same applies to cases where the kana-kanji device is construed as being in the categories of method and program.
(Kana-Kanji Conversion Program2)
FIG. 2 shows a configuration of a kana-kanji conversion program2 that implement a kana-kanji conversion method to which the present invention can be applied. As shown inFIG. 2, the kana-kanji conversion program2 includes alanguage model32 and abase form pool30 described above. Thelanguage model32 includes a first corpus (word-segmented)320 and a second corpus (word-unsegmented)322. The kana-kanji conversion program2 may be stored in theprogram storage device16 or may be loaded into a storage device such as a memory (for example, a random access memory) in the CPU10 (which may be allocated as an array in the program) when being executed.
(Base Form Pool30)
Stored in thebase form pool30 is avocabulary dictionary300 in which the pronunciations of words occurring in the first corpus (word-segmented) are stored correspondingly to the first corpus (word-segmented)320 of thelanguage model32. Also, all characters constituting the words and their pronunciations are stored in acharacter dictionary302. Thecharacter dictionary302 is sometimes referred to as the single-kanji dictionary.
It is novel to consider combinations of all characters and their pronunciations and associate and store the pronunciations with their occurrence probabilities in the character dictionary. In particular, without this provision, it is impossible to conceive of referring to the occurrence probabilities when applying kana-kanji conversion to a pronunciation.
FIG. 3 shows the details of thebase form pool30. For example, the pronunciation /takahashi/ is stored for the word /
/ (Takahashi, a family name) in thevocabulary dictionary300, the pronunciation /kore/ is stored for the word /
/ (this), and the pronunciation /kiyo/ is stored for the word /
/ (contribution). These are stored as already segmented words.
If this vocabulary dictionary is provided correspondingly to the first corpus (word-segmented), statistics on the probabilities of occurrence (the likelihood that each word will appear) can be readily taken like this: the probability that the word /
/ will appear in the first corpus (word-segmented) is 0.010, the probably that /
/ will appear is 0.0300, and the probability that /
/ will appear is 0.020. While the probability of occurrence is shown as being stored in the first corpus (word-segmented)320 inFIG. 3, the storage location is not limited to the example shown inFIG. 3, provided that it is associated with the first corpus (word-segmented).
Stored in thecharacter dictionary302 are combinations of all pronunciations of characters. For example, for the spelling of the character /
/, combinations of all of its pronunciations, including /taka/ and /kou/, are stored; and for the spelling of the character /
/, combinations of all of its pronunciations, including /hashi/ and /kyou/, are stored.
Also stored in thecharacter dictionary302 is a table containing the probabilities of occurrence of pronunciations associated with characters. A 0.7 probability of /taka/ occurring and a 0.3 probability of /kou/ occurring are contained in association with the spelling of the character /
/, a 0.7 probability of /hashi/ occurring and a 0.3 probability of /kyou/ occurring are contained in association with the spelling of the character /
/, and a 0.7 probability of /kore/ occurring and 0.3 probability of /ze/ occurring are contained in association with the spelling of the character /
/, and 0.7 probability of /kiyo/ occurring and a 0.3 probability of /sei/ occurring are contained in association with the spelling of the character /
/.
These probabilities of occurrence may not necessarily be included in thecharacter dictionary302. They may be stored in a location separate from thecharacter dictionary302, provided that correspondences between all the pronunciations and characters are described. In this way, a character-wise unknown-word model is built. Building of the character-wise unknown-word model allows listing of candidates by referring to the occurrence probabilities (frequency information) This is summed inFIG. 4 and will be described later.
(First Corpus (Word-Segmented)320)
Details of the first corpus are shown inFIG. 3. Stored in the first corpus (word-segmented)320 are character strings made up of a number of characters.
(Second Corpus (Word-Unsegmented)322)
Details of the second corpus are shown inFIG. 3. A word boundary probability estimating device (FIG. 4) calculates the probability of a word boundary existing between characters in the first corpus (word-segmented), refers to the probability, applies it to between characters in the second corpus (word-unsegmented), and estimates the probability of a word boundary existing between them.
Stored in the second corpus (word-unsegmented)322 are character strings made up of a number of characters. The second corpus (word-unsegmented), which segmentation has not been applied to, is also called the “raw corpus”. Because segmentation requires manual correction and therefore is troublesome as described above, it is preferable that the large second corpus (word-unsegmented) of large size can be used effectively.
It should be noted that, while a model including the first corpus (word-segmented) and second corpus (word-unsegmented) is called alanguage model32 herein, models, including those stored in thebase form pool30, may be referred to as “languages models” in some places herein. The term “language model” as used herein refers to a storage device in which these items of information.
(Language Decoding Section22)
Alanguage decoding section22 decodes an input symbol string into the word string (W′ in theexpression 2 below) that provides the highest probability calculated using thebase form pool30 and thelanguage model32 to thedisplay device14 or thestorage device16, and outputs it as text data to display or store it on thedisplay device14 or thestorage device16.
According to the present invention, the first corpus (word-segmented) and the second corpus (word-unsegmented) can be linearly interpolated (deleted interpolation) according to the followingexpression 1. This process will be described later with reference toFIG. 8.
Pr(w1|w2,w3)=λP1(w1|w2,w3)+(1−λ)P2 (w1|w2,w3) (1)
In the expression, N=3, 0≦λ≦1, P1 denotes the first corpus (word-segmented), and P2 denotes the second corpus (word-unsegmented).
In thefollowing expression 2, P(Y|W) is given by thebase form model30 and P(W) is given by thelanguage model32. P(W) can be obtained by calculating a weighted average of the first corpus (word-segmented)320 and the second corpus (word-unsegmented)322 according to the value of λ by using theexpression 1.
W′=argmaxP(W|Y)=argmaxP(Y|W)P(W) (2)
where y represents the input symbol string (y1, y2, . . . , yk), W represents the word string (w1, w2, . . . , w1), and W′ represents the word string (w′1, w′2, . . . , w′1).
(Calculation of Word Boundary Probability)
A method for calculating the probability that a word boundary will exist in the first corpus (word-segmented)320 will be described below.
In the first corpus (word-segmented)320, the character string
(Learning linguistics.), is stored as sample sentence. The character string is stored as seven characters,
and
These seven characters are classified into six character types: kanji, symbol, numeric, hiragana, katakana, and alphabetic (characters that cannot be classified into symbol, numeric, nor any of the other character type). The characters in the sequence in the sample sentence can be classified as, “kanji”, “kanji,” “kanji”, “hiragana”, “kanji”, “hiragana”, and “symbol”, respectively.
FIG. 6 shows the probabilities that word boundaries will exist between characters when a character of a particular type is followed by the character of a different or the same type, calculated from the relations in the sequence of the character types in the first corpus (word-segmented)320. The probability can be readily calculated if information indicating whether a word boundary already exists between characters can be obtained from the first corpus.
However, even if such information cannot obtained from the first corpus, the next calculation can be performed by setting a probability of 0.50000000 is set for the entire corpus as preliminary information concerning whether or not word boundaries will exist. It should be noted that, although this action may result in a lower accuracy, the technical idea of the present invention can be widely applied to such a case.
Furthermore, even if words are not segmented and therefore information as to whether or not word boundaries already exist is not available, a probability of 1.00000000 for sentences the endpoints of which are known.
InFIG. 6, it is indicated that the probability that a word boundary will exist between to kanjis is 0.24955045 if one kanji follows the other kanji. Also, it is indicated that the probability of a word boundary will exist between a kanji and a hiragana that follows the kanji is 0.67322202, the probability of a word boundary will exist between a hiragana and a kanji that follows the hiragana is 0.97213218, and the probability of a word boundary exist between a hiragana and a symbol that follow the hiragana is 0.99999955.
Probabilities closer to 1 are higher (likely) and probabilities closer to 0 are lower (unlikely). If word boundaries are already determined (word boundaries already exist), the twovalues 0 and 1 would be enough for discriminating whether or not text is segmented into words. It should be noted that intermediate values between 0 and 1 are used for probabilistically showing the degrees of segmentations. Of course, any other method that shows the degree of probability may be used. (Estimation of the probability of a word boundary existing) In the second corpus (word-unsegmented)322, probabilistic segmentation can be estimated by referring to the probability of a word boundary existing that is calculated from the first corpus (word-segmented)320.
In one of the simplest examples, it is estimated that the probability of word boundaries obtained in the first corpus (word-segmented)320 will directly apply to the second corpus (word-unsegmented). In this case, the values obtained in the first corpus (word-segmented) can be used directly as the probabilities of word boundaries existing, though many other reference methods may be used. It should be noted that the term “reference” or “refer” as used herein has broad meaning in terms of usage.
The probabilities of word boundaries existing that are calculated in the first corpus (word-segmented) are assigned between words in the second corpus (word-unsegmented). For example, if the sample sentence
occurs in the second corpus (word-unsegmented), the following boundaries are applied to between two characters along with the probabilities shown in the square brackets.

[1]
[0.24955045]
[0.24955045]
[0.67322202]
[0.971213218]
[0.67332202]
[1]
[1]

This is based on the relations in the sequence of character types shown inFIG. 6.
That is, if the sample sentence
(Reading a declarative sentence.) occurs in the second corpus (word-unsegmented), the following boundaries along with the same probabilities shown in the square brackets that have been used above are applied to between two characters.

[1]
[0.24055045]
[0.24955045]
[0.67332202]
[0.97213218]
[0.67322202]
[0.99999955]
[1]
(Calculation of Word n-Gram Probability)

An n-gram model is a language model for examining how often n character strings or combinations of words will occur in a character string.
If the word segmentation probability (the probability of segmentation between the i-th character and the i+1-th character is represented by Pi) is calculated, the uni-gram of a word w can be calculated as: $\begin{matrix} fr (w) = \sum_{i \in 0_{1}} P_{i} [\prod_{j = 1}^{k - 1} (1 - P_{i + j})] P_{i + k} \\ O \\ _{1} = {i | x_{i + 1}^{i + k} = w} \end{matrix}$
The frequency of the uni-gram in this example can be calculated as follows.
This means that the probability of the word n-gram is calculated. The uni-gram (N=1) probability can be calculated by using the relation between the preceding and succeeding characters of the character constituting the word uni-gram, in the position of its occurrence.
For example, the uni-gram of the word
can be calculated as:
1×(1−0.24955045)×(1−0.24955045)
Furthermore, the uni-gram of the word having the longer character string
can be calculated as:
1×(1−0.24955045)×(1−0.24955045)×(1−0.67332202)
The uni-gram of the word
can be calculated as:
0.6733202×(1−0.97213218)×0.6733202
However, the expression yields an extremely low value. Accordingly, it can be estimated that the probability of the word
occurring is extremely low, that is, a kanji is unlikely to occur subsequently to a hiragana. This can be understood empirically.
A typical word n-gram probability can be calculated by expanding the above expression. For example bi-gram probability can be calculated as follows: $\begin{matrix} fr (w1W2) = \sum_{i \in 02} (P_{i} [\prod_{j = 1}^{k - 1} (1 - P_{i + j})] P_{i + k} \times [\prod_{j = 1}^{l - 1} (1 - P_{i + K + j})] P_{i + k + 1}) \\ O_{2} = {i | x_{i + 1}^{i + k} = w1 ⋀ x_{i + K + 1}^{i + k + 1} = w2} \end{matrix}$
A method for efficiently calculating the expected frequency of the character string x₁x₂. . . x_kwill be described below with reference toFIG. 5.
FIG. 5 shows a flow chart of a process of calculating the expected frequency of the character string x₁x₂. . . x_kin the second corpus (word-unsegmented)322. In S200, the probability P_intthat a word boundary does not exist in the character string of interest can be calculated.
Here, the probability that a word boundary will exist in a character string of interest,
(consisting of four characters) will be calculated. This word is a proper noun and contains characters of different types. The word boundary probabilities will be represented on a character-type basis.
The character string of interest
consists of four characters of three character types: katakana, katakana, kanji, and alphabetic.
The probability P_intthat a word boundary does not exist in the character string of interest can be calculated as: (1−0.05458673) (1−0.90384580) (1−0.99999955).
At step S210, the position in the second corpus (word-unsegmented)322 at which the character string of interest occurs is sought.
For example, suppose that the character string of interest is found as follows:
. . .
. . . (where in fact there are other character strings preceding and succeeding the character string of interest, which are represented by the suspension points to show the omission; the same applies to the following description).
A hiragana, “
” precedes the character string of interest and a hiragana “
” succeeds the character string of interest. Accordingly, P_sumis calculated as: (1−0.99999955) (1−0.99999955).
At step S230, the next occurrence position of the character string of interest is sought. If at step S240 the character string of interest (a vicinity pattern) is found as . . .
. . . ,
then the process returns to S220. The symbol “┌” precedes the character string of interest and the symbol “┘” succeeds the character string of interest.
Accordingly, P_sumis calculated as (1−0.99999955) (1−0.99999955) and is added to the P_sumcalculated above.
Such addition is repeated until the character string of interest is no longer found in the second corpus (word-unsegmented)322 at S240, then P_int×P_sumis calculated finally at S250. In this way, P_intand P_sumare calculated separately, thereby efficiently calculating the frequency of occurrence of the character string. The calculation process shown inFIG. 5 is called twice in the flow shown inFIG. 7 as a subroutine.
Other methods for calculating the segmentation probability may be used such as methods using a decision tree or PPM. By using these methods, broader range of character strings can be referred to. The technical idea of the present invention is not limited to these. Any other methods can be used that can describe the existence of a word boundary between characters.
Using the word segmentation probabilities, the second corpus (word-unsegmented)322 can be treated as a corpus in which characters are segmented at a character boundary (between xi and xi+1) with a probability Pi.
All occurrences of the spelling of a word w in a raw corpus is given by $O_{1} = {i | x_{i + 1}^{i + k} = w}$
Then, the probabilistic frequency fr of a word w in a raw corpus can be defined as the sum of probabilistic frequencies as follow: $fr (w) = \sum_{i \in 0_{1}} P_{i} [\prod_{j = 1}^{k - 1} (1 - P_{i + j})] P_{i + k}$
This shows that fr is the expected frequency of w in the raw corpus.
Therefore, the word uni-gram probability can be represented as:
P_r(w)=f_r(w)/f_r(•)
where $f_{r} (\cdot) = 1 + \sum_{i = 1}^{nr - 1} P_{i}$
FIG. 7 shows a method for calculating the word n-gram probability P(W_n|W₁, W₂, . . . , W_n-1). At steps S400 and S430, the subroutine shown inFIG. 5 is called. In the process, f2/f1 is calculated. If f1 is 0, the f2/f1 is indeterminate and therefore 0 is returned at S420. On the other hand, if f1 is not 0, the expected frequency f2 of W₁, W₂, . . . , W_nis calculated at S430 and f2/f1 is returned at S440.
(Kana-Kanji Conversion Using the Second Corpus (Word-Unsegmented))
Thelanguage decoding section22 refers to both of thevocabulary dictionary300 and thecharacter dictionary302 in thebase form pool300. At step S100, thelanguage decoding section22 receives an input symbols string from the keyboard. At step S102, thelanguage decoding section22 lists possible input symbol strings and their probabilities. As summarized inFIG. 4, the fact that the probabilistic language model has been built by using the second corpus (word-unsegmented) contributes to the listing of new probabilities for selecting conversion candidates. At step S120, λ is set. At step S122, thelanguage decoding section22 refers to the first corpus (word-segmented)320 and the second corpus (word-unsegmented)322 by assigning weights to them according to theexpression 1 given earlier, where λ≠1. At step S124, thelanguage decoding section22 sequentially outputs the word string with the highest occurrence probability as text data representing the result of the kana-kanji conversion. It should be noted that the word string /
/ can be correctly provided even though it has a probability as low as 0.001. This means that the input symbol string /takahashikorekiyo/ can be correctly converted into the proper noun /
/ even though the input symbol string represents an unknown word.
FIG. 4 summarizes the relations. The words /
/, /
/, and /
/ occur with probabilities of 0.010, 0.030, and 0.020, respectively, in the first corpus (word-segmented) only. However, the multiplication of these probabilities produces a value of 0.0006, which is the probably of these words being candidates. Because the word string /
/ has a higher probability of 0.001, it is evident that the symbol string was correctly converted.
This is because it can be estimated from the character-wise unknown word model that the character sting
occurs in the second corpus (word-unsegmented) and that the input symbol string /takahashikorekiyo/ corresponds to
with a constant probability of 0.001. If the word string were not correctly provide even with a probability as low as 0.001, it would be incorrectly converted to
which is a string of known words with high occurrence frequencies. This is because a sample (bi-gram) in which “
/kore/” is followed by “
/kiyo/” does not occur.
<First Experiment>
Advantages of introducing the proposed method in writing applicable documents will be demonstrated below.FIG. 9 shows details of the corpuses used in the experiment.
FIG. 10 shows the results of comparative experiments on eight models using the corpuses. Model C represents the proposed method. The testing was conducted using the investigation323 sentences. Both in the testing and training, documents similarly relating with each other are not include in both of the corpuses.
Comparing with model A, model A′, and Model B, it is evident that the accuracy was increased by the introduction of the first corpus (word-segmented). Also, comparing model B with model B′, model C with model C′, and model D with model D′, it can be seen that that the advantage of the models that allow bi-grams or higher order to be calculate.
Furthermore, comparing model C with model D, it can be seen that the proposed method efficiently uses the unsegmented corpus in the application field.
<Second Experiment>
An experiment was conducted to also calculate the precision and the recall for evaluating the performance of word segmentation. The evaluation is based on the result of kana-kanji conversion of a katakana spelling and the numbers of the characters in the correct longest common subsequence (LCS). Letting the number of the characters contained in a spelling in the first corpus (word-segmented) be N_C, the number of the characters contained in the result of kana-kanji conversion be N_SYS, and the number of the characters in the longest common subsequence be N_LCS, then the recall can be defied as N_LCS/N_Cand the precision can be defined as N_LCS/N_SYS.
FIG. 11 shows the corpuses used in this experiment.FIG. 12 shows the models used in the experiment and the result. The test was conducted by using 509,261 characters in the EDR corpus. These experiments show that the proposed method (model C) has an advantage over models A and B both in precision and recall.
(Variation)
One may want to use the second corpus (word-unsegmented) more often in some technical fields than in others. A method according to the present invention can readily control this by adjusting the weight in linear interpolation with the first corpus (word-segmented). This is based on the concept that the n-gram probability estimated from the second corpus (word-unsegmented) is less accurate than a language model estimated from a corpus in which words are manually and precisely segmented.
Variations described for the present invention can be realized in any combination desirable for each particular application. Thus particular limitations, and/or embodiment enhancements described herein, which may have particular advantages to a particular application need not be used for all applications. Also, not all limitations need be implemented in methods, systems and/or apparatus including one or more concepts of the present invention. Methods may be implemented as signal methods employing signals to implement one or more steps. Signals include those emanating from the Internet, etc.
The present invention can be realized in hardware, software, or a combination of hardware and software. A visualization tool according to the present invention can be realized in a centralized fashion in one computer system, or in a distributed fashion where different elements are spread across several interconnected computer systems. Any kind of computer system—or other apparatus adapted for carrying out the methods and/or functions described herein—is suitable. A typical combination of hardware and software could be a general purpose computer system with a computer program that, when being loaded and executed, controls the computer system such that it carries out the methods described herein. The present invention can also be embedded in a computer program product, which comprises all the features enabling the implementation of the methods described herein, and which—when loaded in a computer system—is able to carry out these methods.
Computer program means or computer program in the present context include any expression, in any language, code or notation, of a set of instructions intended to cause a system having an information processing capability to perform a particular function either directly or after conversion to another language, code or notation, and/or reproduction in a different material form.
Thus the invention includes an article of manufacture which comprises a computer usable medium having computer readable program code means embodied therein for causing a function described above. The computer readable program code means in the article of manufacture comprises computer readable program code means for causing a computer to effect the steps of a method of this invention. Similarly, the present invention may be implemented as a computer program product comprising a computer usable medium having computer readable program code means embodied therein for causing a function described above. The computer readable program code means in the computer program product comprising computer readable program code means for causing a computer to affect one or more functions of this invention. Furthermore, the present invention may be implemented as a program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for causing one or more functions of this invention.
It is noted that the foregoing has outlined some of the more pertinent objects and embodiments of the present invention. This invention may be used for many applications. Thus, although the description is made for particular arrangements and methods, the intent and concept of the invention is suitable and applicable to other arrangements and applications. It will be clear to those skilled in the art that modifications to the disclosed embodiments can be effected without departing from the spirit and scope of the invention. The described embodiments ought to be construed to be merely illustrative of some of the more prominent features and applications of the invention. Other beneficial results can be realized by applying the disclosed invention in a different manner or modifying the invention in ways known to those familiar with the art.

Claims

1) A word boundary probability estimating device comprising:

means for calculating a probability that a word boundary will exist between a first plurality of characters constituting a first character string stored in a first corpus by invoking information as to whether a word boundary already exists between the first plurality of characters from the first corpus containing the first character string comprising the first plurality of characters or setting up the preliminary information as to whether the word boundary exists, and

means for estimating the probability that the word boundary will exist in a second plurality of characters constituting a second character string stored in a second corpus by referring to a calculated probability between the second plurality of characters constituting the second character string stored in the second corpus.

2) The word boundary probability estimating device according toclaim 1, wherein the probability that a word boundary will exist between the plurality of characters is calculated on the basis of relations in the sequence of character types of successive characters in the character string in the first corpus by using the information as to whether a word boundary already exists between characters of those character types.

3) The word boundary probability estimating device according toclaim 1, wherein the calculated probability that a word boundary will exist includes values between 0 and 1 capable of probabilistically indicating the degree of segmentation as well as values of 0 and 1 indicating whether or not segmentation is already determined, in order to discriminate whether the character string including the plurality of characters is segmented or not.

4) The word boundary probability estimating device according toclaim 1; further comprising:

word n-gram probability calculating means for calculating the probability that a word boundary will exist on the basis of the relation between the preceding character and the subsequent character at the position of each occurrence of a character string including one or more characters constituting a word n-gram, the probability being a word n-gram probability, and thereby forming a probabilistic language model building device.

5) The word boundary probability estimating device according toclaim 1; further comprising:

word n-gram probability calculating means for calculating the probability that a word boundary will exist on he basis of a decision tree in which relations between a plurality of characters constituting a word n-gram are described or on the basis of PPM (Prediction by Partial Match), the probability being a word n-gram probability, and thereby forming a probabilistic language model building device.

6) The word boundary probability estimating device according toclaim 4; further comprising:

a first corpus storing a character string which includes a plurality of characters and is segmented into at least two words as a word including one or more characters, and storing an occurrence probability indicating the likelihood of occurrence of each of segmented words in association with the segmented word;

a second corpus storing a character string including a plurality of characters and a word n-gram probability calculated by the probabilistic language model building device;

a vocabulary dictionary storing, correspondingly to the first corpus, pronunciations associated with each of words segmented as known words;

a character dictionary storing, correspondingly to the first corpus, a pronunciation and spelling of each character in a plurality of conversion candidates that can be converted from an unknown word in association with each other so that that the plurality of conversion candidates can be listed for the unknown word, and storing the occurrence probability of a pronunciation of each character; and

a language decoding section converting a input spelling into a conversion candidate by referring to the occurrence probabilities associated with each word stored in the first corpus and the occurrence probabilities of pronunciations of each character stored in the character dictionary, and word the n-gram probability stored in the second corpus for an input pronunciation, and thereby forming a kana-kanji converting device.

7) An unknown word model building method comprising the steps of:

selecting one of the words occurring in a character string stored in a storage device;

assigning a pronunciation to the selected word;

storing the word and the pronunciation in a vocabulary dictionary;

assigning pronunciations to each of the characters constituting the word; and

associating and storing a plurality of different occurrence probabilities with the associated pronunciations into a character dictionary.

8) A word boundary probability estimating method comprising the steps of:

calculating a probability that a word boundary will exist between a plurality of characters constituting a character string stored in a first corpus by invoking information as to whether a word boundary already exists between the plurality of characters from the first corpus containing the character string including the plurality of characters or setting up the preliminary information as to whether the word boundary exists; and

estimating the probability that a word boundary will exist between a plurality of characters constituting a character string stored in a second corpus by referring to the calculated probability between the plurality of characters constituting the character string stored in the second corpus.

9) The word boundary probability estimating method according toclaim 8, wherein the step of calculating the probability that a word boundary will exist between a plurality of characters calculates the probability on the basis of relations in the sequence of character types of successive characters in the character string in the first corpus by using the information as to whether a word boundary already exists between characters of those character types.

10) The word boundary probability estimating method according toclaim 8, wherein the calculated probability that a word boundary will exist includes values between 0 and 1 capable of probabilistically indicating the degree of segmentation as well as values of 0 and 1 indicating whether or not segmentation is already determined, in order to discriminate whether the character string including the plurality of characters is segmented or not.

11) The word boundary probability estimating method according toclaim 8, further comprising the step of calculating the probability that a word boundary will exist on the basis of the relation between the preceding character and the subsequent character at the position of each occurrence of a character string including at least one character constituting a word n-gram, the probability being a word n-gram probability, thereby constituting a probabilistic language model building method.

12) The word boundary probability estimating method according toclaim 8, further comprising the step of calculating the probability that a word boundary will exist on the basis of a decision tree in which relations between a plurality of characters constituting a word n-gram are described or on the basis of PPM (Prediction by Partial Match), the probability being a word n-gram probability, thereby constituting a probabilistic language model building method.

13) The word boundary probability estimating method according toclaim 11, further comprising the steps of:

referring to a first corpus storing a character string which includes a plurality of characters and is segmented into at least two words as the word including at least one character, and storing an occurrence probability indicating the likelihood of occurrence of each of segmented words in association with the segmented word;

referring to a second corpus storing a character string including a plurality of characters and a word n-gram probability calculated by the probabilistic language model building device according to the probabilistic language model building method;

referring to a vocabulary dictionary storing, correspondingly to the first corpus, pronunciations associated with each of words segmented as known words;

referring to a character dictionary storing, correspondingly to the first corpus, pronunciations and spelling of each character in a plurality of conversion candidates that can be converted from an unknown word in association with each other so that that the plurality of conversion candidates can be listed for the unknown word and storing the occurrence probability of a pronunciation of each character; and

converting a input spelling into a conversion candidate by referring to the occurrence probabilities associated with each word stored in the first corpus and the occurrence probabilities of pronunciations of each character stored in the character dictionary, and the word n-gram probability stored in the second corpus for an input pronunciation, thereby forming a kana-kanji conversion method.

14) The word boundary probability estimating method according toclaim 12, further comprising the steps of:

15) The word boundary probability estimating device according toclaim 5; further comprising:

a first corpus storing a character string which includes a plurality of characters and is segmented into at least two words as a word including at least one character, and storing an occurrence probability indicating the likelihood of occurrence of each of segmented words in association with the segmented word;

15) An article of manufacture comprising a computer usable medium having computer readable program code means embodied therein for causing unknown word model building, the computer readable program code means in said article of manufacture comprising computer readable program code means for causing a computer to effect the steps ofclaim 7.

16) An article of manufacture comprising a computer usable medium having computer readable program code means embodied therein for causing word boundary probability estimating, the computer readable program code means in said article of manufacture comprising computer readable program code means for causing a computer to effect the steps ofclaim 8.

17) A program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for unknown word model building, said method steps comprising the steps ofclaim 7.

18) A program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for word boundary probability estimating, said method steps comprising the steps ofclaim 8.

19) A computer program product comprising a computer usable medium having computer readable program code means embodied therein for causing functions of a word boundary probability estimating device, the computer readable program code means in said computer program product comprising computer readable program code means for causing a computer to effect the functions ofclaim 1.

20) A computer program product comprising a computer usable medium having computer readable program code means embodied therein for causing functions of a kana-kanji converting device, the computer readable program code means in said computer program product comprising computer readable program code means for causing a computer to effect the functions ofclaim 6.