Classifications

• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F40/00—Handling natural language data
  • G06F40/20—Natural language analysis
  • G06F40/263—Language identification

Definitions

• the present disclosure relates to language detection and, in particular, to systems and methods for detecting languages in short text messages.
• language detection or identification is a process in which a language present in a body of text is detected automatically based on the content of the text.
• Language detection is useful in the context of automatic language translation, where the language of a text message must generally be known before the message can be translated accurately into a different language.
• chat text domain a particularly challenging domain
• messages often include only a few words (e.g., four or less), some or all of which can be informal and/or misspelled.
• chat text domain existing language detection approaches have proven to be inaccurate and/or slow, given the lack of information and the informalities present in such messages.
• Embodiments of the systems and methods described herein are used to detect the language in a text message based on, for example, content of the message, information about the keyboard used to generate the message, and/or information about the language preferences of the user who generated the message. Compared to previous language detection techniques, the systems and methods described herein are generally faster and more accurate, particularly for short text messages (e.g., of four words or less).
• the systems and methods use a plurality of language detection tests and classifiers to determine probabilities associated with possible languages in a text message.
• Each language detection test can output a set or vector of probabilities associated with the possible languages.
• the classifiers can combine the output from the language detection tests to determine a most likely language for the message.
• the particular language detection test(s) and classifier(s) chosen for the message can depend on a predicted accuracy, a confidence score, and/or a linguistic domain for the message.
• Certain examples of the systems and methods described herein perform an initial classification of a language in a text message so that more focused language detection techniques can be performed to make a final determination of the language.
• the systems and methods can perform initial language detection testing on a text message to identify a group or category (e.g., Cyrillic languages or Latin languages) for the language in the text message. Once the language category is identified, language detection techniques designed for the language category can be used to identify the specific language in the message. In preferred examples, extraneous elements (e.g., emoji or numerical digits or characters) are removed from the text message prior to language detection, thereby resulting in faster and more accurate language detection.
• the systems and methods described herein are generally more accurate and efficient than prior language detection approaches.
• the systems and methods can be configured to use any one or more of the language detection methods described herein.
• the subject matter of this disclosure relates to a computer-implemented method of identifying a language in a message.
• the method includes: obtaining a text message; removing non-language characters from the text message to generate a sanitized text message; and detecting at least one of an alphabet and a script present in the sanitized text message, wherein detecting includes at least one of: (i) performing an alphabet-based language detection test to determine a first set of scores, wherein each score in the first set of scores represents a likelihood that the sanitized text message includes the alphabet for one of a plurality of different languages; and (ii) performing a script-based language detection test to determine a second set of scores, wherein each score in the second set of scores represents a likelihood that the sanitized text message includes the script for one of the plurality of different languages.
• the method also includes identifying the language in the sanitized text message based on at least one of the first set of scores, the second set of scores, and a combination of the first and second sets of scores.
• the non-language characters include an emoji and/or a numerical character.
• the combination can include an interpolation between the first and second sets of scores.
• identifying the language in the sanitized text message includes performing a language detection test on the sanitized text message to generate a third set of scores, wherein each score in the third set of scores represents a likelihood that the sanitized text message includes one of a plurality of different languages.
• the language detection test can be selected from a plurality of language detection tests, based on the at least one of the first set of scores, the second set of scores, and the combination of the first and second sets of scores.
• the language detection test includes a language detection method and one or more classifiers.
• the language detection method can include, for example, a dictionary-based language detection test, an n-gram language detection test, an alphabet-based language detection test, a script-based language detection test, a user language profile language detection test, or any combination thereof.
• the one or more classifiers can include, for example, a supervised learning model, a partially supervised learning model, an unsupervised learning model, an interpolation, or any combination thereof.
• the method includes processing the third set of scores using one or more classifiers to identify the language in the sanitized text message.
• the method can include outputting, from the one or more classifiers, an indication that the sanitized text message is in the identified language.
• the indication can include a confidence score.
• the subject matter of this disclosure relates to a computer- implemented system for identifying a language in a message.
• the system includes a sanitizer module, a grouper module, and a language detector module.
• the sanitizer module obtains a text message and removes non-language characters from the text message to generate a sanitized text message.
• the grouper module detects at least one of an alphabet and a script present in the sanitized text message and is operable to perform operations including at least one of: performing an alphabet-based language detection test to determine a first set of scores, wherein each score in the first set of scores represents a likelihood that the sanitized text message includes the alphabet for one of a plurality of different languages; and performing a script-based language detection test to determine a second set of scores, wherein each score in the second set of scores represents a likelihood that the sanitized text message includes the script for one of the plurality of different languages.
• the language detector module identifies the language in the sanitized text message based on at least one of the first set of scores, the second set of scores, and a combination of the first and second sets of scores.
• the non-language characters include an emoji and/or a numerical character.
• the combination can include an interpolation between the first and second sets of scores.
• the grouper module can be operable to perform operations that include selecting the language detector module from a plurality of language detector modules based on the at least one of the first set of scores, the second set of scores, and the combination of the first and second sets of scores.
• the language detector module can include a language detection methods module.
• the language detection methods module can be operable to perform operations that include performing a language detection test on the sanitized text message to generate a third set of scores, wherein each score in the third set of scores represents a likelihood that the sanitized text message includes one of a plurality of different languages.
• the language detection test can include, for example, a dictionary -based language detection test, an n-gram language detection test, an alphabet-based language detection test, a script-based language detection test, a user language profile language detection test, or any combination thereof.
• the language detector module includes a classifier module operable to perform operations that include processing the third set of scores using one or more classifiers to identify the language in the sanitized text message.
• the one or more classifiers can include, for example, a supervised learning model, a partially supervised learning model, an unsupervised learning model, an interpolation, or any combination thereof.
• the classifier module can be operable to perform operations that include outputting an indication that the sanitized text message is in the identified language.
• the indication can include a confidence score.
• the article includes: a non-transitory computer-readable medium having instructions stored thereon that, when executed by one or more computers, cause the computers to perform operations including: obtaining a text message; removing non-language characters from the text message to generate a sanitized text message; detecting at least one of an alphabet and a script present in the sanitized text message, wherein detecting includes at least one of: (i) performing an alphabet-based language detection test to determine a first set of scores, wherein each score in the first set of scores represents a likelihood that the sanitized text message includes the alphabet for one of a plurality of different languages; and (ii) performing a script-based language detection test to determine a second set of scores, wherein each score in the second set of scores represents a likelihood that the sanitized text message includes the script for one of the plurality of different languages.
• the operations further include identifying the language in the sanitized text message based on at least one of the first set of scores, the second set of scores, and a combination of the first and second sets of scores.
• FIG. 1 A is a diagram of an example system for performing language detection.
• FIG. IB is a flowchart of an example method of detecting a language in a text message.
• FIG. 2 is a flowchart of an example n-gram method of detecting a language in a text message.
• FIG. 3 is a flowchart of an example dictionary-based method of detecting a language in a text message.
• FIG. 4 is a flowchart of an example alphabet-based method of detecting a language in a text message.
• FIG. 5 is a flowchart of an example script-based method of detecting a language in a text message.
• FIG. 6 is a flowchart of an example user language profile method of detecting a language in a text message.
• FIG. 7 is a schematic diagram of an example language detection module.
• FIG. 8 is a schematic diagram of an example classifier module.
• FIG. 9 is a flowchart of an example method of detecting a language in a text message using the language detection module of FIG. 7 and the classifier module of FIG. 8.
• FIG. 10 is a flowchart of an example method of detecting a language in a text message.
• FIG. 11 is a flowchart of an example method of detecting a language in a text message.
• FIG. 12 is a flowchart of an example method of detecting a language in a text message.
• FIG. 13 is a schematic diagram of an example system for detecting a language in a text message.
• FIG. 14 is a flowchart of an example method of detecting a language in a text message.
• FIG. 15 is a flowchart of an example method of detecting a language in a text message.
• the language detection systems and methods described herein can be used to identify the language in a text message when language information for the message (e.g., keyboard information from a client device) is absent, malformed or unreliable.
• language information for the message e.g., keyboard information from a client device
• the systems and methods improve the accuracy of language translation methods used to translate text messages from one language to another.
• Language translation generally requires the source language to be identified accurately, otherwise the resulting translation can be inaccurate.
• FIG. 1A illustrates an example system 10 for detecting a language in a message, such as a text message or an audio message.
• a server system 12 provides message analysis and language detection functionality.
• the server system 12 comprises software components and databases that can be deployed at one or more data centers 14 in one or more geographic locations, for example.
• the server system 12 software components can comprise a detection module 16, a classifier module 18, and a manager module 20.
• the software components can comprise subcomponents that can execute on the same or on different individual data processing apparatus.
• the server system 12 databases can comprise training data 22, dictionaries 24, alphabets 26, scripts 28, and user profile information 30.
• the databases can reside in one or more physical storage systems. The software components and data will be further described below.
• An application such as a web-based application, can be provided as an end-user application to allow users to provide messages to the server system 12.
• the end-user applications can be accessed through a network 32 by users of client devices, such as a personal computer 34, a smart phone 36, a tablet computer 38, and a laptop computer 40.
• client devices such as a personal computer 34, a smart phone 36, a tablet computer 38, and a laptop computer 40.
• client devices are possible.
• the user messages can be accompanied by information about the devices used to create the messages, such as information about the keyboard, client device, and/or operating system used to create the messages.
• FIG. 1A depicts the classifier module 18 and the manager module 20 as being connected to the databases (i.e., training data 22, dictionaries 24, alphabets 26, scripts 28, and user profile information 30), the classifier module 18 and/or the manager module 20 are not necessarily connected to some or all of the databases.
• the classifier module 18 can receive input from the detection module 16, and the manager module 20 can receive input from the classifier module 18. No other inputs are required to be received by the classifier module 18 and/or the manager module 20.
• FIG. IB illustrates an example method 100 that uses the system 10 to detect a language in a message.
• the method 100 begins by receiving or obtaining (step 102) a text message generated by a user.
• the text message is analyzed (step 104) using one or more language detection methods (e.g., by the detection module 16) that each provide an indication of the language or languages present in the message.
• the output from the language detection methods is then combined (step 106) using one or more classifiers (e.g., by the classifier module 18) that provide a further indication of the language present in the message.
• the one or more classifiers can include, for example, a supervised learning model, a partially supervised learning model, an unsupervised learning model, and/or an interpolation.
• the output from the one or more classifiers is then used to determine (step 108) the language in the message (e.g., using the manager module 20).
• the language indication from the one or more classifiers can be selected by the manager module 20 according to a computed confidence score and/or a linguistic domain.
• the classifiers can compute a confidence score indicating a degree of confidence associated with the language prediction.
• certain classifier output can be selected according to the linguistic domain associated with the user or the message. For example, if the message originated in a computer gaming
• a particular classifier output can be selected as providing the most accurate language prediction.
• a different classifier output can be selected as being more appropriate for the sports linguistic domain.
• Other possible linguistic domains include, for example, news, parliamentary proceedings, politics, health, travel, web pages, newspaper articles, microblog messages, and the like.
• certain language detection methods or combinations of language detection methods can be more accurate for certain linguistic domains, when compared to other linguistic domains.
• the domain can be determined based on the presence of words from a domain - Si - vocabulary in a message.
• a domain vocabulary for computer gaming could include common slang words used by gamers.
• the language detection methods used by the detection module 16 can include, for example, an n-gram method (e.g., a byte n-gram method), a dictionary-based method, an alphabet-based method, a script-based method, and a user language profile method. Other language detection methods are possible. Each of these language detection methods can be used to detect a language present in a message. The output from each method can be, for example, a set or vector of probabilities associated with each possible language in the message. In some instances, two or more of the language detection methods can be performed simultaneously, using parallel computing, which can reduce computation times considerably.
• a byte n-gram method uses byte n-grams instead of word or character n-grams to detect languages.
• the byte n-gram method is preferably trained over a mixture of byte n-grams (e.g., with 1 ⁇ n ⁇ 4), using a naive Bayes classifier having a multinomial event model.
• the model preferably generalizes to data from different linguistic domains, such that the model's default configuration is accurate over a diverse set of domains, including newspaper articles, online gaming, web pages, and microblog messages. Information about the language identification task can be integrated from a variety of domains.
• the task of attaining high accuracy can be relatively easy for language identification in a traditional text categorization setting, for which in-domain training data is available.
• This task can be more difficult when attempting to use learned model parameters for one linguistic domain to classify or categorize data from a separate linguistic domain.
• This problem can be addressed by focusing on important features that are relevant to the task of language identification. This can be based on, for example, a concept called information gain, which was originally introduced for decision trees as a splitting criteria, and later found to be useful for selecting features in text categorization.
• a detection score can be calculated that represents the difference in information gain relative to domain and language. Features having a high detection score can provide information about language without providing information about domain.
• an example byte n-gram method 200 begins by training the method using the training data 22.
• the method can be trained over a mixture of byte n-grams, using the naive Bayes classifier having the multinomial event model.
• Training data 22 is preferably collected (step 202) for a wide number and variety of languages and adjusted so that the amount of data available per language is made uniform.
• a small portion of the training data 22 is set aside (step 204) as a test set.
• a byte n-gram model is trained (step 206) on the training data 22 with appropriate smoothing and backoff techniques.
• the input features to the model are byte streams from each input sentence, and since the source language label is known for these sentences, the model adjusts its parameters to learn byte sequences typical to a given language.
• the test set that was isolated in the beginning is then used to predict (step 208) language labels based on the model that had been trained.
• the accuracy of prediction gives the performance of this byte n-gram language identification system.
• it is difficult to train such a byte n-gram system for each linguistic domain by collecting data across numerous languages for said domain. The difficulty arises from a lack of sufficient data per domain.
• these byte n-gram systems are typically trained to cater to a generic domain and not to any specific domains.
• the trained model can be compiled (step 210) into a program along with intermediate machine parameters.
• the program can serve as a general purpose language identification system.
• the dictionary-based language detection method counts the number of tokens or words belonging to each language by looking up words in a dictionary or other word listing associated with the language.
• the language having the most words in the message is chosen as the best language. In the case of multiple best languages, the more frequent or commonly used of the best languages can be chosen.
• the language dictionaries can be stored in the dictionaries database 24.
• FIG. 3 is a flowchart of an example dictionary-based language detection method 300.
• a text message is provided (step 302) and a set of possible languages for the text message is identified (step 304) using, for example, the detection module 16.
• a first possible language is then chosen (step 306) from the set.
• the words in the text message that are present in a dictionary for the possible language are counted (step 308). If additional possible languages from the set have not yet been considered (step 310), a new possible language is selected (step 312), and step 308 is repeated. Once all possible languages from the set have been considered, the language with the most words in the text message can be identified (step 314) as the language in the message.
• the method can be used to compute the likelihood that a language is in the message, for each language in the set.
• the output from the dictionary -based method can be a vector of probabilities for each language in the set.
• dictionaries that include informal words or chat words (e.g., abbreviations, acronyms, slang words, and profanity), in addition to formal words.
• informal words or chat words e.g., abbreviations, acronyms, slang words, and profanity
• Informal words are commonly used in short text communications and in chat rooms.
• the dictionaries are preferably augmented to include informal words on an ongoing basis, as new informal words are developed and used.
• the alphabet-based method is generally based on character counts for each language's alphabet and relies on the observation that many languages have unique alphabets or different sets of characters. For example, Russian, English, Korean, and Japanese each use a different alphabet. Although the alphabet-based method can be unable to distinguish some languages precisely (e.g., languages that use similar alphabets, such as Latin languages), the alphabet- based method can generally detect certain languages quickly. In some instances, it is preferable to use the alphabet-based method in combination with one or more other language detection methods (e.g., using a classifier), as discussed herein.
• the language alphabets can be stored in the alphabets database 26.
• FIG. 4 is a flowchart of an example alphabet-based language detection method 400.
• a text message is provided (step 402) and a set of possible languages for the text message is identified (step 404) using, for example, the detection module 16.
• a first possible language is then chosen (step 406) from the set.
• the characters in the text message that are present in an alphabet for the possible language are counted (step 408). If additional possible languages from the set have not yet been considered (step 410), a new possible language is selected (step 412), and step 408 is repeated. Once all possible languages from the set have been considered, the language with the most characters in the text message can be identified (step 414) as the language in the message.
• the alphabet-based method can be used to compute a likelihood that a language is in the message, for each language in the set.
• the output from the alphabet-based method can be a vector of probabilities for each language in the set.
• the script-based language detection method determines the character counts for each possible script (e.g. Latin script, CJK script, etc.) that are present in the message.
• the script-based method relies on the observation that different languages can use different scripts, e.g., Chinese and English.
• the method preferably uses a mapping that maps a script to a list of languages that use the script. For example, the mapping can consider the UNICODE values for the characters or symbols present in the message, and these UNICODE values can be mapped to a corresponding language or set of possible languages for the message.
• the language scripts and UNICODE values or ranges can be stored in the scripts database 28.
• a text message is provided (step 502) and the scripts present in the message are identified (step 504) using, for example, the detection module 16.
• the number of characters falling into each script is then counted (step 506).
• the script with the highest number of characters is considered to be the best script (step 508), and the languages corresponding to the best script are identified (step 510).
• the output from the script-based method is a set of probabilities (e.g., in vector form) for each possible language in the message.
• the user language profile based method uses the user profile information database 30, which stores historical messages sent by various users.
• the languages of these stored messages are detected using, for example, one or more other language detection methods described herein (e.g., the byte n-gram method), to identify the language(s) used by each user. For example, if all of a user's prior messages are in Spanish, the language profile for that user can indicate the user's preferred language is Spanish. Likewise, if a user's prior messages are in a mixture of different languages, the language profile for the user can indicate probabilities associated with the different languages (e.g., 80% English, 15% French, and 5% Spanish).
• the user language profile based method addresses language detection issues associated with very short messages, which often do not have enough information in them to make an accurate language determination.
• the language preference of a user can be used to predict the language(s) in the user's messages, by assuming the user will continue to use the language(s) he or she has used previously.
• an example user language profile detection method 600 begins by storing (step 602) a user's previous messages and detecting (step 604) the language(s) present in the stored messages. The frequency with which different language(s) appear in the user's message is determined (step 606) and a use probability for such language(s) is outputted (step 608).
• each language detection method can provide a vector of probabilities, where each probability in the vector is associated with a possible language in the message and represents the likelihood that the message is in the given language. Due to the different methods employed and the information available in the message, the probabilities from each language detection method may not be consistent.
• the detection module 16 can include or utilize, for example, an n-gram module 702 for performing an n-gram detection method (e.g., the byte n- gram detection method 200), a dictionary-based module 704 for performing the dictionary- based method 300, an alphabet-based module 706 for performing the alphabet-based method 400, a script-based module 708 for performing the script-based method 500, and a user language profile module 710 for performing the user language profile method 600. Additional or alternative language detection methods can be incorporated into the detection module 16, as desired. Some known methods include using, for example, word level n-grams, Markov models, and predictive modeling techniques.
• the classifier module 18 can include an interpolation module 802, a support vector machines (SVM) module 804, and a linear SVM module 806.
• the interpolation module 802 is used to perform a linear interpolation of the results from two or more language detection methods.
• the language of a text message can be determined by interpolating between results from the byte n-gram method and the dictionary -based method.
• the byte n-gram method can determine the likelihood of English is 0.3, the likelihood of French is 0.4, and the likelihood of Polish is 0.3 (e.g., the output from the byte n-gram method can be ⁇ en:0.3, fr:0.4, pl:0.3 ⁇ ).
• the dictionary-based method can determine the likelihood of English is 0.1, the likelihood of French is 0.2, and the likelihood of Polish is 0.7 (e.g., the output can be ⁇ en:0.1, fr:0.2, pl:0.7 ⁇ ).
• the output from the byte n-gram method is multiplied by a first weight and the output from the dictionary -based method is multiplied by a second weight, such that the first and second weights add to one.
• the optimal weights for interpolating between two or more values can be determined numerically through trial and error. Different weights can be tried to identify the best set of weights for a given set of messages. In some instances, the weights can be a function of the number of words or characters in the message. Altematively or additionally, the weights can depend on the linguistic domain of the message. For example, the optimal weights for a gaming environment can be different than the optimal weights for a sports environment. For a combination of the byte n-gram method and the dictionary-based method, good results can be obtained using a weight of 0.1 on the byte n-gram method and a weight of 0.9 on the dictionary-based method.
• the SVM module 804 can be or include a supervised learning model that analyzes language data and recognizes language patterns.
• the SVM module 804 can be a multi-class SVM classifier, for example.
• the feature vector can be the concatenation of the two distributions above (i.e., ⁇ en:0.3, fr:0.4, pl:0.3, en:0.1, fr:0.2, pl:0.7 ⁇ ).
• the SVM classifier is preferably trained on labeled training data.
• the trained model acts as a predictor for an input.
• the features selected in the case of language detection can be, for example, sequences of bytes, words, or phrases.
• Input training vectors can be mapped into a multi-dimensional space.
• the SVM algorithm can then use kernels to identify the optimal separating hyperplane between these dimensions, which will give the algorithm a
• the kernel can be, for example, a linear kernel, a polynomial kernel, or a radial basis function (RBF) kernel, although other suitable kernels are possible.
• a preferred kernel for the SVM classifier is the RBF kernel. After training the SVM classifier using training data, the classifier can be used to output a best language among all the possible languages.
• the training data can be or include, for example, the output vectors from different language detection methods and an indication of the correct language, for a large number of messages having, for example, different message lengths, linguistic domains, and/or languages.
• the training data can include a large number of messages for which the language in each message is known.
• the linear SVM module 806 can be or include a large-scale linear classifier.
• An SVM classifier with a linear kernel can perform better than other linear classifiers, such as linear regression.
• the linear SVM module 806 differs from the SVM module 804 at the kernel level. There are some cases when a polynomial model works better than a linear model, and vice versa.
• the optimal kernel can depend on the linguistic domain of the message data and/or the nature of the data.
• Other possible classifiers used by the systems and methods described herein include, for example, decision tree learning, association rule learning, artificial neural networks, inductive logic programming, random forests, clustering, Bayesian networks, reinforcement learning, representation learning, similarity and metric learning, and sparse dictionary leaming. One or more of these classifiers, or other classifiers, can be incorporated into and/or form part of the classifier module 18.
• an example method 900 uses the detection module 16, the classifier module 18, and the manager module 20 to detect the language in a message.
• the message is provided or delivered (step 902) to the detection module 16.
• the message can be accompanied by information about the message and/or the user who generated the message.
• the information can include, for example, an identification number for the user, information about the keyboard used by the user to generate the message, and/or information about the operating system controlling the software used by the user to generate the message.
• the message can be accompanied by data indicating the user used a French keyboard to generate the message and that user's operating system is in English.
• one or more language detection methods are used (step 904) to detect a language in the message.
• Each method used by the detection module 16 can output a prediction regarding the language present in the message.
• the prediction can be in the form of a vector that includes a probability for each possible language that can be in the message.
• the output from the detection module 16 is then delivered to the classifier module 18 where the results from two or more language detection methods can be combined (step 906).
• Various combinations of the results from the language detection methods can be obtained.
• the results from the byte n-gram method and the dictionary-based method are combined in the classifier module 18 by interpolation.
• a SVM SVM
• combination or classification is performed on the results from the byte n-gram method, the dictionary-based method, the alphabet method, and the user profile method.
• the combination can include or consider results from the script-based method.
• a further example includes a large linear combination of the byte n-gram method, the language profile method, and the dictionary method. In general, however, the results from any two or more of the language detection methods can be combined in the classifier module 18.
• the method 900 uses the manager module 20 to select output (step 908) from a particular classifier.
• the output can be selected based on, for example, a confidence score computed by a classifier, an expected language detection accuracy, and/or a linguistic domain for the message.
• a best language is then chosen (step 910) from the selected classifier output.
• a method 1000 includes receiving or providing a message (step 1002) that can include information about the keyboard language used to generate the message. If the message is greater than a threshold length (e.g., 25 bytes or 25 characters, although any suitable threshold length is possible) (step 1004), the language can be detected (step 1006) using the byte n-gram method (or other method or combination of methods). The language of the message can then be chosen (step 1008), based on the results from the byte n-gram method.
• a threshold length e.g. 25 bytes or 25 characters, although any suitable threshold length is possible
• the system can determine if the keyboard language is available (step 1010). If the keyboard language is available, the language of the message can be chosen (step 1012) to be the same as the keyboard language. Alternatively, if the keyboard language is not available, the method 1000 can again consider the length of the message. For example, if the message length is less than a second threshold (e.g., 4 bytes or 4 characters, although any suitable second threshold is possible) (step 1014), the language can be detected (step 1016) and chosen using the dictionary-based method. Otherwise, if the message length is greater than the second threshold, the byte n-gram method (or other method or combination of methods) can be used (step 1018) to detect the language in the message. The results from the byte n-gram method and the dictionary-based method can be combined (e.g., using an interpolator or other classifier), and the language of the message can be determined based on the combination (step 1020).
• a second threshold e.g., 4 bytes or 4 characters, although any suitable second threshold is possible
• FIG. 1 1 is an example method 1 100 of identifying a language in a text message.
• a text message that was generated on a client device of a user is received or provided (step 1102).
• the alphabet-based method and/or the script-based method are used to determine (step 1 104) an alphabet and/or a script associated with the text message.
• a candidate language associated with the alphabet and/or the script is identified. If the candidate language is a language with a unique alphabet and/or script (e.g., Russian, Arabic, Hebrew, Greek, Chinese, Taiwanese, Japanese, Korean, or the like) (step 1 106), then the candidate language is determined (step 1 108) to be the language of the text message.
• a unique alphabet and/or script e.g., Russian, Arabic, Hebrew, Greek, Chinese, Taiwanese, Japanese, Korean, or the like
• the length of the text message is evaluated. If the message length is less than a threshold length (e.g., 4 bytes or 4 characters, although any suitable threshold length is possible) and the text message includes or is accompanied by a keyboard language used by the client device (step 11 10), then the language of the message is chosen (step 11 12) to be the keyboard language. [0066] Alternatively, if the message length is greater than the threshold length or the keyboard language is not available, then the message is processed with an n-gram method (e.g., the byte n-gram method) to identify (step 11 14) a first set of possible languages for the text message.
• a threshold length e.g., 4 bytes or 4 characters, although any suitable threshold length is possible
• the message is also then processed with the dictionary -based method to identify (step 1 116) a second set of possible languages for the text message. If a user language profile exists for the user (step 1 118), then the user language profile is obtained (step 1120) and combined (e.g., using an SVM classifier or a large linear classifier) with the first set of possible languages and the second set of possible languages to obtain a first combination of possible languages (step 1 122). The language of the text message is then chosen (step 1124), based on the first combination of possible languages.
• the first set of possible languages and the second set of possible languages are combined (e.g., using a linear interpolator or other classifier) to obtain a second combination of possible languages (step 1 126).
• the language of the text message is chosen (step 1 128), based on the second combination of possible languages.
• language detection is performed by combining the output from multiple language detection methods in two or more steps.
• a first step can use the alphabet-script based method to detect special languages that use their own unique alphabets or scripts, such as, for example, Chinese (cn), Japanese (ja), Korean (ko), Russian (ru), Hebrew (he), Greek (el), and Arabic (ar).
• the alphabet-script based method refers to, for example, using one or both of the alphabet-based method and the script-based method.
• the second step can use a combination (e.g., from a classifier) of multiple detection methods (e.g., the byte n-gram method, the user language profile based method, and the dictionary-based method) to detect other languages (e.g., Latin languages) in the message.
• a combination e.g., from a classifier
• multiple detection methods e.g., the byte n-gram method, the user language profile based method, and the dictionary-based method
• other languages e.g., Latin languages
• the message provided or received for language detection includes certain digits, characters, or images (e.g., emoticons or emoji) that are not specific to any particular language and/or are recognizable to any user, regardless of language preference.
• the systems and methods described herein can ignore such characters or images when performing language detection and can ignore messages that include only such characters or images.
• the systems and methods can remove such characters or images from messages, prior to performing language detection.
• the process of removing extraneous characters or images from messages can be referred to herein as sanitizing the messages.
• the sanitizing process can result in faster detection times and/or improved language detection accuracy.
• FIG. 12 is a flowchart of an example method 1200 for detecting a language in a message.
• the method uses the detection module 16, the classifier module 18, and the manager module 20 to identify a most likely or best language 1202 for a given input message 1204.
• the input message 1204 can be accompanied by information about the user or the system(s) used to generate the message.
• the input message 1204 can be accompanied by a user identification number (or other user identifier), information about the keyboard (e.g., a keyboard language) used to generate the message, and/or information about the operating system (e.g., an operating system language) used to generate the message.
• a user identification number or other user identifier
• information about the keyboard e.g., a keyboard language
• the operating system e.g., an operating system language
• the detection module 16 includes ten different language detection methods.
• Three of the language detection methods in the detection module 16 are Byte n-gram A 1206, Byte n-gram B 1208, and Byte n-gram C 1210, which are all byte n-gram methods and can be configured to detect a different set or number of languages.
• Byte n-gram A 1206 can be configured to detect 97 languages
• Byte n-gram B 1208 can be configured to detect 27 languages
• Byte n-gram C 1210 can be configured to detect 20 languages.
• Two of the language detection methods in the detection module 16 are
• Dictionary A 1212 and Dictionary B 1214 which are both dictionary-based methods and can be configured to detect a different set or number of languages.
• Dictionary A 1212 can be configured to detect 9 languages
• Dictionary B 1214 can be configured to detect 10 languages.
• Two of the language detection methods in the detection module 16 are Language Profile A 1216 and Language Profile B 1218, which are user language profile methods and can be configured to detect a different set or number of languages.
• Language Profile A 1216 can be configured to detect 20 languages
• Language Profile B 1218 can be configured to detect 27 languages.
• Two of the language detection methods in the detection module 16 are Alphabet A 1220 and Alphabet B 1222, which are alphabet-based methods and can be configured to detect a different set or number of languages.
• Alphabet A 1220 can be configured to detect 20 languages
• Alphabet B 1222 can be configured to detect 27 languages.
• the detection module 16 also includes a script-based language detection method 1224.
• Output from the different language detection methods in the detection module 16 is combined and processed by the classifier module 18.
• an interpolation classifier 1226 combines output from Byte n-gram B 1208 and Dictionary B 1214. Weights for the interpolation can be, for example, 0.1 for Byte n-gram B 1208 and 0.9 for Dictionary B 1214.
• the classifier module 18 can also use an SVM classifier 1228 that combines output from Byte n-gram C 1210, Dictionary B 1214, Language Profile B 1218, and Alphabet B 1222.
• the classifier module 18 can also use a first combination 1230 of the script-based method 1224 and an SVM classifier combination of Byte n-gram C 1210, Dictionary A 1212, Language Profile A 1216, and Alphabet A 1220. Additionally, the classifier module 18 can use a second combination 1232 of the script based method 1224 and a Linear SVM classifier combination of Byte n-gram C 1210, Dictionary A 1212, and Language Profile A 1216. While FIG. 12 shows specific language detection tests, classifiers, and combinations of detection test output being used in the classifier module 18, other language detection tests, classifiers, and/or combinations can be used.
• the script- based method 1224 and the classifier can be used in a tiered approach.
• the script- based method 1224 can be used to quickly identify languages having unique scripts. When such a language is identified in the message 1204, use of the SVM classifier in the first combination 1230 or the Linear SVM classifier in the second combination may not be required.
• the manager module 20 can select specific language detection methods, classifiers, and/or combinations of detection method output to identify the language in the message 1204.
• the manager module 20 can make the selection according to the linguistic domain or according to an anticipated language for the message.
• the manager module 20 can select specific classifiers according to a confidence score determined by the classifiers. For example, the manager module 20 can select the output from the classifier that is the most confident in its prediction.
• the systems and methods described herein are suitable for making language detection available as a service to a plurality of users. Such a service is made possible and/or enhanced by the speed at which the systems and methods identify languages, and by the ability of the systems and methods to handle multiple identification techniques at runtime, based on service requests from diverse clients.
• a language detection system 1300 obtains or receives a text message 1302 and uses a sanitizer module 1304, a grouper module 1306, and one or more language detector modules to identify a language present in the text message 1302.
• the sanitizer module 1304 prepares the text message 1302 for language detection by deleting certain characters or other extraneous elements from the text message 1302.
• the sanitizer module 1304 can remove, for example, one or more numerical characters (e.g., “1 ,” “3,” “15,” “249,” etc.), one or more emoji (e.g., in the form of images and/or combinations of characters, such as “:-D” or “:- ⁇ ”), and/or one or more other non-language characters or elements that may not be useful for language detection, such as, for example, certain punctuation marks (e.g., periods or commas), extra spaces, and/or carriage returns.
• the extraneous elements are not removed from the text message 1302, but are simply ignored or avoided during subsequent language detection.
• the sanitizer module 1304 can be used to flag or identify the extraneous elements in such instances. References to "the text message 1302" in subsequent paragraphs are intended to cover, for example, text messages with extraneous elements removed or ignored. In various examples, the extraneous elements are referred to herein as "non-language characters.”
• the grouper module 1306 is used to perform an initial classification of the language in the text message 1302 and, based on the initial classification, select one or more subsequent language detection methods to make a final determination of the language in the text message 1302.
• the grouper module 1306 performs the initial classification by detecting an alphabet and/or a script present in the text message 1302.
• the alphabet and/or the script can be detected using, for example, the alphabet-based method and/or the script-based method, described herein.
• the alphabet-based method can determine a first set of scores for the text message 1302, with each score representing a probability or likelihood that the alphabet is for one of a plurality of different languages.
• the grouper module 1306 can detect the alphabet in the text message 1302 based on the highest score from the first set of scores.
• the script-based method can determine a second set of scores for the text message 1302, with each score representing a probability or likelihood that the script is for one of a plurality of different languages.
• the grouper module 1306 can detect the script in the text message 1302 based on the highest score from the second set of scores.
• the grouper module 1306 can combine results or scores (e.g., using an interpolator or other classifier) from the alphabet-based method and the script- based method to detect the alphabet and/or the script in the text message 1302.
• the grouper module 1306 selects a language detector module to use for making a final determination of the language in the text message 1302, as described below and herein.
• the grouper module 1306 can pass results or other information (e.g., one or more scores) from the alphabet-based method and/or the script-based method to the selected language detector module.
• the language detection system 1300 can include or utilize the following language detector modules: an alphabet-distinguishable language detector 1308, a Cyrillic language detector 1310, a Latin language detector 1312, and a backoff language detector 1314.
• an alphabet-distinguishable language detector 1308, a Cyrillic language detector 1310, a Latin language detector 1312, and a backoff language detector 1314 can be included or utilized.
• Each of these language detector modules 1308, 1310, 1312, and 1314 can include a detection methods module and a classifier module.
• the alphabet- distinguishable language detector 1308 can include a detection methods module 1316 and a classifier module 1318
• the Cyrillic language detector 1310 can include a detection methods module 1320 and a classifier module 1322
• the Latin language detector 1312 can include a detection methods module 1324 and a classifier module 1326
• the backoff language detector 1314 can include a detection methods module 1328 and a classifier module 1330.
• the detection methods modules 1316, 1320, 1324, and 1328 include or utilize one or more language detection methods, which can be or include, for example, the n- gram method (e.g., the byte n-gram method), the dictionary-based method, the alphabet-based method, the script-based method, and/or the user language profile method. Other language detection methods are contemplated.
• the detection methods modules 1316, 1320, 1324, and 1328 can use the language detection methods to produce output providing an indication of the language present in the text message 1302.
• the output can be or include, for example, one or more scores representing a likelihood that the text message 1302 is in one or more languages.
• the language in the text message 1302 is determined directly from the output of one of the detection methods modules 1316, 1320, 1324, or 1328.
• the language in the text message 1302 can be determined from the output of one of the classifier modules 1318, 1322, 1326, or 1330.
• each classifier module 1318, 1322, 1326, or 1330 processes output from a corresponding detection methods module 1316, 1320, 1324, or 1328 to provide a further indication of the language present in a text message.
• the classifier modules 1318, 1322, 1326, and 1330 preferably use or include one or more classifiers, such as, for example, a supervised learning model, a partially supervised learning model, an unsupervised learning model, and/or an interpolation.
• an alphabet- distinguishable language is a language that has a unique alphabet and/or a unique script, such that the language in the text message 1302 can be determined once the alphabet and/or the script for the language are detected.
• alphabet-distinguishable languages include, for example, Simplified Chinese (cn), Traditional Chinese (tw), Japanese ja), Arabic (ar), Hebrew (he), Greek (el), Korean (ko), and Thai (th).
• the grouper module 1306 passes results (e.g., one or more scores or probabilities, a detected alphabet, and/or a detected script) from the alphabet-based method and/or the script-based method to the alphabet- distinguishable language detector 1308.
• the detection methods module 1316 can perform the alphabet-based method and/or the script-based method to detect the alphabet and/or the script in the text message 1302.
• the alphabet- distinguishable language detector 1308 can determine the language in the text message 1302 once the alphabet and/or the script are detected. In some instances, such a determination can be made using the classifier module 1318 to process any output from the detection methods module 1316.
• the grouper module 1306 selects the Cyrillic language detector 1310.
• Cyrillic languages include, for example, Bulgarian (bg), Ukrainian (uk), and Russian (ru).
• the detection methods module 1320 can include or utilize one or more language detection methods described herein, such as the byte n-gram method and/or the dictionary-based method.
• the detection methods module 1320 utilizes the dictionary -based method, which can use one or more dictionaries specific to Cyrillic languages.
• the dictionary -based method can count the number of tokens or words in the text message 1302 that belong to one or more Cyrillic languages by looking up words in the one or more dictionaries.
• the Cyrillic language having the most tokens or words in the text message 1302 is determined to be the language in the text message 1302.
• the detection methods module 1320 can provide output from one or more language detection methods (e.g., the dictionary -based method) to the classifier module 1322, which can process the output to determine the language in the text message 1302.
• the classifier module 1322 can receive a set of scores from the detection methods module 1320 and can determine the Cyrillic language in the text message 1302 by identifying the language having the highest score.
• the grouper module 1306 selects the alphabet and/or script detected by the grouper module 1306 .
• Latin languages include, for example, English (en), French (fr), Spanish (es), German (de), Portuguese (pt), Dutch (nl), Polish (pi), Italian (it), Vietnamese (tr), Catalan (ca), Czech (cs), Danish (da), Finnish (fi), Hungarian (hu), Indonesian (id), Norwegian (no), Romanian (ro), Slovak (sk), Swedish (sv), Malay (ms), Vietnamese (vi).
• the detection methods module 1324 can include or utilize one or more language detection methods described herein.
• the detection methods module 1324 includes or utilizes the byte n-gram method and/or the dictionary-based method.
• the output from one or both of these preferred methods can be processed or combined using the classifier module 1326 to determine the specific Latin language in the text message 1302.
• the n-gram method and the dictionary-based method can each output a set of scores, with each score representing a likelihood that the text message 1302 is in one of a plurality of different Latin languages.
• the classifier module 1326 can process the sets of scores using, for example, one or more classifiers and/or interpolation techniques described herein, to determine the Latin language in the text message 1302.
• the grouper module 1306 selects the backoff language detector 1314 to detect a language in the text message 1302.
• the backoff language detector 1314 can be selected, for example, when the grouper module 1306 does not select the alphabet- distinguishable language detector 1308, the Cyrillic language detector 1310, or the Latin language detector 1312. Such a situation may occur, for example, when the grouper module 1306 fails to detect an alphabet and/or a script associated with an alphabet-distinguishable language, a Cyrillic language, or a Latin language.
• the detection methods module 1328 and/or the classifier module 1330 can be used to identify the language in the text message 1302.
• the language detection methods used by the detection methods module 1328 can be or include, for example, the n-gram method (e.g., the byte n-gram method), the dictionary-based method, the alphabet-based method, the script-based method, the user language profile method, and any combination thereof.
• the specific classifiers used by the classifier module 1330 can be or include, for example, a supervised learning model, a partially supervised learning model, an unsupervised learning model, an interpolation, and/or any combination thereof. Other language detection methods and/or classifiers can be used.
• the backoff language detector 1314 can use any of the language detection methods and classifiers described herein.
• the backoff language detector 1314 is preferably flexible and can be configured to include or use new detection methods and/or new combinations of detection methods as such new methods and/or combinations are developed or become available. In some instances, by resorting to the backoff language detector 1314, the language detection system 1300 is able to provide a valid output rather than a NULL output.
• FIG. 14 is a flowchart of an example method 1400 for identifying a language in a text message.
• the method 1400 includes receiving the text message (step 1402) and detecting (step 1404) at least one of a Latin alphabet and a Latin script in the text message.
• a dictionary-based language detection test is performed (step 1406) to determine a first set of scores, with each score in the first set of scores representing a likelihood that the text message is in one of a plurality of different Latin languages.
• An n-gram language detection test is performed (step 1408) to determine a second set of scores, with each score in the second set of scores representing a likelihood that the text message is in one of the plurality of different Latin languages.
• the first and second sets of scores are combined (step 1410) using, for example, one or more classifiers and/or interpolation techniques.
• the language in the text message is identified (step 1412) based on the combination.
• FIG. 15 is a flowchart of an example method 1500 for identifying a language in a text message.
• a text message is obtained (step 1502) and non-language characters are removed (step 1504) from the text message to generate a sanitized text message.
• An alphabet and/or a script are detected in the sanitized text message by performing at least one of: (i) an alphabet- based language detection test to determine a first set of scores (step 1506) and (ii) a script- based language detection test to determine a second set of scores (step 1508).
• Each score in the first set of scores represents a likelihood that the sanitized text message includes the alphabet for one of a plurality of different languages.
• Each score in the second set of scores represents a likelihood that the sanitized text message includes the script for one of the plurality of different languages.
• the language in the sanitized text message is identified (step 1510) based on the first set of scores, the second set of scores, and/or a combination of the first and second sets of scores.
• Embodiments of the subject matter described in this specification can be implemented as one or more computer programs, i.e., one or more modules of computer program instructions, encoded on computer storage medium for execution by, or to control the operation of, data processing apparatus.
• the program instructions can be encoded on an artificially generated propagated signal, e.g., a machine-generated electrical, optical, or electromagnetic signal, that is generated to encode information for transmission to suitable receiver apparatus for execution by a data processing apparatus.
• a computer storage medium can be, or be included in, a computer- readable storage device, a computer-readable storage substrate, a random or serial access memory array or device, or a combination of one or more of them.
• a computer storage medium is not a propagated signal
• a computer storage medium can be a source or destination of computer program instructions encoded in an artificially-generated propagated signal.
• the computer storage medium can also be, or be included in, one or more separate physical components or media (e.g., multiple CDs, disks, or other storage devices).
• the operations described in this specification can be implemented as operations performed by a data processing apparatus on data stored on one or more computer-readable storage devices or received from other sources.
• data processing apparatus encompasses all kinds of apparatus, devices, and machines for processing data, including by way of example a programmable processor, a computer, a system on a chip, or multiple ones, or combinations, of the foregoing.
• the apparatus can include special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application-specific integrated circuit).
• the apparatus can also include, in addition to hardware, code that creates an execution environment for the computer program in question, e.g., code that constitutes processor firmware, a protocol stack, a database management system, an operating system, a cross-platform runtime environment, a virtual machine, or a combination of one or more of them.
• code that creates an execution environment for the computer program in question e.g., code that constitutes processor firmware, a protocol stack, a database management system, an operating system, a cross-platform runtime environment, a virtual machine, or a combination of one or more of them.
• the apparatus and execution environment can realize various different computing model infrastructures, such as web services, distributed computing and grid computing infrastructures.
• a computer program (also known as a program, software, software application, script, or code) can be written in any form of programming language, including compiled or interpreted languages, declarative or procedural languages, and it can be deployed in any form, including as a stand-alone program or as a module, component, subroutine, object, or other unit suitable for use in a computing environment.
• a computer program can, but need not, correspond to a file in a file system.
• a program can be stored in a portion of a file that holds other programs or data (e.g., one or more scripts stored in a markup language document), in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files that store one or more modules, sub-programs, or portions of code).
• a computer program can be deployed to be executed on one computer or on multiple computers that are located at one site or distributed across multiple sites and interconnected by a communication network.
• the processes and logic flows described in this specification can be performed by one or more programmable processors executing one or more computer programs to perform actions by operating on input data and generating output.
• the processes and logic flows can also be performed by, and apparatus can also be implemented as, special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application-specific integrated circuit).
• special purpose logic circuitry e.g., an FPGA (field programmable gate array) or an ASIC (application-specific integrated circuit).
• processors suitable for the execution of a computer program include, by way of example, both general and special purpose microprocessors, and any one or more processors of any kind of digital computer.
• a processor will receive instructions and data from a read-only memory or a random access memory or both.
• the essential elements of a computer are a processor for performing actions in accordance with instructions and one or more memory devices for storing instructions and data.
• a computer will also include, or be operatively coupled to receive data from or transfer data to, or both, one or more mass storage devices for storing data, e.g., magnetic disks, magneto-optical disks, optical disks, or solid state drives.
• mass storage devices for storing data, e.g., magnetic disks, magneto-optical disks, optical disks, or solid state drives.
• a computer need not have such devices.
• a computer can be embedded in another device, e.g., a mobile telephone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a Global Positioning System (GPS) receiver, or a portable storage device (e.g., a universal serial bus (USB) flash drive), to name just a few.
• Devices suitable for storing computer program instructions and data include all forms of nonvolatile memory, media and memory devices, including, by way of example, semiconductor memory devices, e.g., EPROM, EEPROM, and flash memory devices; magnetic disks, e.g., internal hard disks or removable disks; magneto-optical disks; and CD-ROM and DVD-ROM disks.
• the processor and the memory can be supplemented by, or incorporated in, special purpose logic circuitry.
• embodiments of the subject matter described in this specification can be implemented on a computer having a display device, e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor, for displaying information to the user and a keyboard and a pointing device, e.g., a mouse, a trackball, a touchpad, or a stylus, by which the user can provide input to the computer.
• a display device e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor
• a keyboard and a pointing device e.g., a mouse, a trackball, a touchpad, or a stylus
• a computer can interact with a user by sending documents to and receiving documents from a device that is used by the user; for example, by sending web pages to a web browser on a user's client device in response to requests received from the web browser.
• Embodiments of the subject matter described in this specification can be implemented in a computing system that includes a back-end component, e.g., as a data server, or that includes a middleware component, e.g., an application server, or that includes a front-end component, e.g., a client computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the subject matter described in this specification, or any combination of one or more such back-end, middleware, or front-end components.
• the components of the system can be interconnected by any form or medium of digital data communication, e.g., a communication network. Examples of communication networks include a local area network ("LAN”) and a wide area network (“WAN”), an internetwork (e.g., the Internet), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks).
• LAN local area network
• WAN wide area network
• Internet internetwork
• peer-to-peer networks
• the computing system can include clients and servers.
• a client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.
• a server transmits data (e.g., an HTML page) to a client device (e.g., for purposes of displaying data to and receiving user input from a user interacting with the client device).
• client device e.g., for purposes of displaying data to and receiving user input from a user interacting with the client device.
• Data generated at the client device e.g., a result of the user interaction

Landscapes

• Engineering & Computer Science (AREA)
• Theoretical Computer Science (AREA)
• Health & Medical Sciences (AREA)
• Artificial Intelligence (AREA)
• Audiology, Speech & Language Pathology (AREA)
• Computational Linguistics (AREA)
• General Health & Medical Sciences (AREA)
• Physics & Mathematics (AREA)
• General Engineering & Computer Science (AREA)
• General Physics & Mathematics (AREA)
• Machine Translation (AREA)

Applications Claiming Priority (2)

Publications (1)

Family

ID=60162256

Family Applications (1)

Country Status (6)

Families Citing this family (2)

Family Cites Families (4)

• 2017
  • 2017-10-02 JP JP2019517966A patent/JP2019535082A/ja active Pending
  • 2017-10-02 EP EP17788004.4A patent/EP3519984A1/de not_active Withdrawn
  • 2017-10-02 CA CA3039085A patent/CA3039085A1/en not_active Abandoned
  • 2017-10-02 AU AU2017339433A patent/AU2017339433A1/en not_active Abandoned
  • 2017-10-02 WO PCT/US2017/054722 patent/WO2018067440A1/en active Application Filing
  • 2017-10-02 CN CN201780074219.8A patent/CN110023931A/zh active Pending

Also Published As

Similar Documents

Legal Events