Natural Language Recognition Using Context Information
A method of recognising digital ink input by a user into a computer-based digital ink recognition system is disclosed. The user interacts with a paper-based document. The paper-based document has disposed thereon coded data indicative of a particular field of the paper-based document and of at least one reference point of the paper-based document. An image sensor in a sensing device captures images of at least some of the coded data when the sensing device is placed in an operative position relative to the paper-based document. The sensing device then decodes at least some of the coded data to form indicating data indicative of the identity of the field of the paper-based document containing the coded data and at least one of a position and a movement of the sensing device relative to the paper-based document. A server receives the indicating data from the sensing device, and processes the indicating data using a recognizer residing on the server to produce intermediate format data. The intermediate format data is then transmitted to an application which decodes the intermediate format data into computer-readable format data using context information associated with the paper-based document.
Latest Patents:
- TOSS GAME PROJECTILES
- BICISTRONIC CHIMERIC ANTIGEN RECEPTORS DESIGNED TO REDUCE RETROVIRAL RECOMBINATION AND USES THEREOF
- CONTROL CHANNEL SIGNALING FOR INDICATING THE SCHEDULING MODE
- TERMINAL, RADIO COMMUNICATION METHOD, AND BASE STATION
- METHOD AND APPARATUS FOR TRANSMITTING SCHEDULING INTERVAL INFORMATION, AND READABLE STORAGE MEDIUM
The present application is a continuation application of U.S. patent application Ser. No. 10/510,392 filed on Oct. 7, 2004, which is a 371 of PCT/AU04/001088 filed on Aug. 16, 2004, the content of is herein incorporated by reference.
TECHNICAL FIELDThe present invention relates to a method of and system for natural language recognition, and in particular, to a method of and system for computer-based recognition of natural language data implemented on a distributed computer network.
CROSS REFERENCESVarious methods, systems and apparatus relating to the present invention are disclosed in the following granted US patents and co-pending US applications filed by the applicant or assignee of the present application: The disclosures of all of these granted US patents and co-pending US applications are incorporated herein by reference.
Recent advances in pattern classification have enabled the development of sophisticated software systems that can recognize natural language data (i.e. natural language user input) such as speech (see for example L. Rabiner and B. Juang, “Fundamentals of Speech Recognition”, Prentice Hall, Englewood Cliffs, N.J., 1993) or handwriting (see for example G. Lorette, “Handwriting Recognition or Reading? Situation At The Dawn of the 3rd Millennium”, Advances In Handwriting Recognition, Series in Machine Perception and Artificial Intelligence, Vol. 34, pp. 3-15, World Scientific Publishing Co. 1999).
These applications allow users to communicate with a computerised system in a natural and convenient way, and permit the automation of tasks that previously required human input. Some examples of such applications include interactive voice response (IVR) systems, automated cheque-processing systems and automated form data-entry systems.
In addition, the growth of networked computing and the Internet has enabled the development of complex distributed systems, and the existence of open, standardized protocols has allowed the integration of end-user devices, centralized servers, and applications. An example of a three-tiered distributed system architecture is depicted in
The basic processing steps of presently known pattern recognition systems are depicted in
The segmented signal 260 is then passed to a classifier 270 where a representative set of features is extracted from the signal and used in combination with a pre-defined model 275 of the input signal to produce a set of symbol hypotheses 280. These hypotheses 280 give an indication of the probability that a sequence of segments within the signal represent a basic symbolic element (e.g. letter, word, phoneme, etc.). After classification 270, the context-processing module 290 uses the symbol hypotheses 280 generated by the classifier 270 to decode the signal according to a specified context model 295 (such as a dictionary or character grammar). The result 297 produced by the context processing 290 is passed to the application 299 for interpretation and further processing.
Natural language input is inconsistent, noisy, and ambiguous, leading to potential recognition and decoding errors. However, high recognition accuracy is required for pattern recognition applications to operate successfully, since mistakes can be expensive and frustrating to users. As a result, recognition systems should make use of as much contextual information as possible to increase the possibility of correctly recognizing the natural language input. For example, when recognizing a signal that must represent a country name, the recognition system can use a pre-defined list of valid country names to guide the recognition procedure. Similarly, when recognizing a phone number, a limited symbol set (i.e. digits) can be used to constrain the recognition results. The problem domain for many pattern recognition systems is inherently ambiguous (i.e. many of the input patterns encountered during processing cannot be accurately classified without further information from a different source).
The following discussion refers to handwriting by way of background information, however, the present invention should not be considered to be limited to application to only handwriting as the form of natural language data input.
Digital ink is a digital representation of the information generated by a pen-based input device. Generally, digital ink is structured as a sequence of strokes that begin when the pen device makes contact with a drawing surface and ends when the pen-based input device is lifted. Each stroke comprises a set of sampled coordinates that define the movement of the pen-based input device whilst the pen-based input device is in contact with the drawing surface.
As an example, one of the major issues faced in the development of highly accurate handwriting recognition systems is the inherent ambiguity of handwriting (e.g. the letters ‘u’ and ‘v’, ‘t’ and ‘f’, and ‘g’ and ‘y’ are often written with a very similar appearance and are thus easily confused). Human readers rely on contextual knowledge to correctly decode handwritten text, and as a result a large amount of research has been directed at applying syntactic and linguistic constraints to handwritten text recognition (see for example: H. Beigi and T. Fujisaki, “A Character Level Predictive Language Model and Its Application to Handwriting Recognition”, Proceedings of the Canadian Conference on Electrical and Computer Engineering, Toronto, Canada, Sep. 13-16, 1992; U. Marti and H. Bunke, “Handwritten Sentence Recognition”, Proceedings of the 15th International Conference on Pattern Recognition, Barcelona, Spain, Volume 3, pp. 467-470, 2000; D. Bouchaffra, V. Govindaraju, and S. Srihari, “Postprocessing of Recognized Strings Using Nonstationary Markovian Models”, IEEE Transactions Pattern Analysis and Machine Intelligence, 21(10), pp. 990-999, October 1999; J. Pitrelli and E. Ratzlaff, “Quantifying the Contribution of Language Modeling to Writer-Independent On-line Handwriting Recognition”, Proceedings of the Seventh International Workshop on Frontiers in Handwriting Recognition, Amsterdam, September 11-13, 2000; R. Srihari, “Use of Lexical and Syntactic Techniques in Recognizing Handwritten Text”, ARPA Workshop on Human Language Technology, Princeton, N.J., March 1994; and L. Yaeger, B. Webb, and R. Lyon, “Combining Neural Networks and Context-Driven Search for On-Line, Printed Handwriting Recognition in the Newton”, AI Magazine, Volume 19, No. 1, pp. 73-89, AAAI 1998).
The increasing use of pen-based computing and the emergence of paper-based interfaces to networked computing resources (see for example: Anoto, “Anoto, Ericsson, and Time Manager Take Pen and Paper into the Digital Age with the Anoto Technology”, Press Release, 6 Apr. 2000; and Y. Chans, Z. Lei, D. Lopresti, and S. Kung, “A Feature Based Approach For Image Retrieval by Sketch”, Proceedings of SPIE Volume 3229: Multimedia Storage and Archiving Systems II, 1997) has highlighted the need for techniques to interpret digital ink. Pen-based computing allows users to interact with applications.
As a result of the progress in pen-based interface research, handwritten digital ink documents, represented by time-ordered sequences of sampled pen strokes, are becoming increasingly popular (J. Subrahmonia and T. Zimmerman: Pen Computing: Challenges and Applications. Proceedings of the ICPR, 2000, pp. 2060-2066). Handwriting typically involves writing in a mixture of writing styles (e.g. cursive, discrete, run-on etc.), a variety of fonts and scripts and different layouts (e.g. mixing drawings with text, various text line orientations etc.).
Presently, handwriting recognition accuracy remains relatively low, and the number of errors introduced by recognition (both for the database entries and for the handwritten query) means that present techniques do not work well. The process of converting handwriting into text results in the loss of a significant amount of information regarding the general shape and dynamic properties of the ink. In many handwriting styles (particularly cursive writing), the identification of individual characters is highly ambiguous.
Similar work has been performed in the field of speech recognition, natural language processing, and machine translation.
Some known natural language recognition systems currently exist. Paragraph, Inc. offers a network-based distributed handwriting recognition system called “NetCalif” (ParaGraph, Handwriting Recognition for Internet Connected Device, November 1999) that is based on their Calligraphy handwriting recognition software. The user's natural handwriting—cursive, print, or a combination of both—is captured by client software, then transmitted from an Internet-connected device to the NetCalif servers where it is converted and returned as typewritten text to the client device.
Philips has developed “SpeechMagic”, a client/server-based, professional speech recognition software package (Philips, SpeechMagic 4.0, 2000). This system supports specialized vocabularies (called ConTexts) and dictation, recognition, and correction can be done, independently of the location, across a LAN, WAN, or the Internet.
In a networked information or data communications system, a user has access to one or more terminals which are capable of requesting and/or receiving information or data from local or remote information sources. The information source, in the present context, may be a database associated with an application. In such a communications system, a terminal may be a type of processing system, computer or computerised device, personal computer (PC), mobile, cellular or satellite telephone, mobile data terminal, portable computer, Personal Digital Assistant (PDA), pager, thin client, or any other similar type of digital electronic device. The capability of such a terminal to request and/or receive information or data can be provided by software, hardware and/or firmware. A terminal may include or be associated with other devices, for example a pen-based input device for handwriting input or a microphone for speech input.
An information source can include a server, or any type of terminal, that may be associated with one or more storage devices that are able to store information or data, such as digital ink, for example in one or more databases residing on a storage device. The exchange of information (i.e., the request and/or receipt of information or data) between a terminal and an information source, or other terminal(s), is facilitated by a communication means. The communication means can be realised by physical cables, for example a metallic cable such as a telephone line, semi-conducting cables, electromagnetic signals, for example radio-frequency signals or infra-red signals, optical fibre cables, satellite links or any other such medium or combination thereof connected to a network infrastructure.
The reference to any prior art in this specification is not, and should not be taken as, an acknowledgment or any form of suggestion that such prior art forms part of the common general knowledge.
DISCLOSURE OF INVENTIONAccording to an aspect of the present invention there is provided a method of recognising digital ink input by a user into a computer-based digital ink recognition system, the user interacting with a paper-based document, the paper-based document having disposed therein or thereon coded data indicative of a particular field of the paper-based document and of at least one reference point of the paper-based document, the method including the steps of:
capturing by an image sensor in a sensing device images of at least some of the coded data when the sensing device is placed in an operative position relative to the paper-based document;
decoding by the sensing device at least some of the coded data to form indicating data indicative of the identity of the field of the paper-based document containing the coded data and at least one of a position and a movement of the sensing device relative to the paper-based document;
receiving in a server the indicating data from the sensing device;
processing the indicating data using a recognizer residing on the server to produce intermediate format data;
receiving the intermediate format data in an application; decoding the intermediate format data by the application into computer-readable format data using context information associated with the paper-based document. Other aspects are also disclosed.
The present invention should become apparent from the following description, which is given by way of example only, of a preferred but non-limiting embodiment thereof, described in connection with the accompanying figures.
The following modes, given by way of example only, are described in order to provide a more precise understanding of the subject matter of the present invention.
A particular embodiment of the present invention can be realised using a processing system, an example of which is shown in
Input device 306, for example a pen-based input device or a microphone, is normally remote to the system 300. Input device 306 is used by a user to generate natural language data 318 which is preferably transmitted over network 307 to system 300 for processing. Output device 308 produces or generates intermediate format data 320, for example for transmission over a network, to be transmitted to application 324, which could be remote or local to the system 300. The storage device 314 can be any form of data or information storage means, for example, volatile or non-volatile memory, solid state storage devices, magnetic devices, etc.
In use, the processing system 300 may be a server and is adapted to allow data or information to be stored in and/or retrieved from, via wired or wireless communication means, the at least one database 316, which may be remote and accessed via a further network. The processor 302 receives natural language data 318 from input device 306, preferably via network 307, and outputs intermediate format data 320 by utilising output device 308, for example a network interface. The application 324 may return decoded data to the processing system. The application 324 may cause information to be printed, for example on a Netpage™ printer, at a user's location. More than one input device 306 can be provided. It should be appreciated that the processing system 300 may be any form of terminal, server, specialised hardware, or the like. The processing system 300 may be a part of a networked communications system. Also, the application 324 may initiate transfer of natural language data 318 from the input device 306 to server 300.
In a particular embodiment, the server 300 is part of a system for computer-based recognition of natural language data, the system implemented on a network and comprising: the input device 306 to obtain natural language data; server 300 to receive the natural language data 318 via a network 307; a recognizer residing on the server 300 to process, in processor 302, the natural language data 318 to produce intermediate format data 320; and, an application 324 to receive the intermediate format data 320 and to decode the intermediate format data 320 into computer-readable format data using context information associated with the application 324.
The following example provides a more detailed discussion of a particular embodiment of the present invention. The example is intended to be merely illustrative and not limiting to the scope of the present invention.
In a particular preferred embodiment, the present invention is configured to work with the Netpage™ networked computer system, a detailed description of which is given in the applicant's co-pending applications, including in particular, PCT Publication No. WO0242989 entitled “Sensing Device” filed 30 May 2002, PCT Publication No. WO0242894 entitled “Interactive Printer” filed 30 May 2002, PCT Publication No. WO0214075 “Interface Surface Printer Using Invisible Ink” filed 21 Feb. 2002, PCT Publication No. WO0242950 “Apparatus For Interaction With A Network Computer System” filed 30 May 2002, and PCT Publication No. WO03034276 entitled “Digital Ink Database Searching Using Handwriting Feature Synthesis” filed 24 Apr. 2003.
It will be appreciated that not every implementation will necessarily embody all or even most of the specific details and extensions described in these applications in relation to the basic system. However, the system is described in its most complete form to assist in understanding the context in which the preferred embodiments and aspects of the present invention operate.
In brief summary, the preferred form of the Netpage system provides an interactive paper-based interface to online information by utilizing pages of invisibly coded paper and an optically imaging pen. Each page generated by the Netpage system is uniquely identified and stored on a network server, and all user interaction with the paper using the Netpage pen is captured, interpreted, and stored. Digital printing technology facilitates the on-demand printing of Netpage documents, allowing interactive applications to be developed. The Netpage printer, pen, and network infrastructure provide a paper-based alternative to traditional screen-based applications and online publishing services, and supports user-interface functionality such as hypertext navigation and form input.
Typically, a printer receives a document from a publisher or application provider via a broadband connection, which is printed with an invisible pattern of infrared tags that each encodes the location of the tag on the page and a unique page identifier. As a user writes on the page, the imaging pen decodes these tags and converts the motion of the pen into digital ink. The digital ink is transmitted over a wireless channel to a relay base station, and then sent to the network for processing and storage. The system uses a stored description of the page to interpret the digital ink, and performs the requested actions by interacting with an application.
Applications provide content to the user by publishing documents, and process the digital ink interactions submitted by the user. Typically, an application generates one or more interactive pages in response to user input, which are transmitted to the network to be stored, rendered, and finally printed as output to the user. The Netpage system allows sophisticated applications to be developed by providing services for document publishing, rendering, and delivery, authenticated transactions and secure payments, handwriting recognition and digital ink searching, and user validation using biometric techniques such as signature verification.
Distributed Pattern Recognition
An example architecture for a distributed pattern recognition system 400 is depicted in
Performing pattern recognition on a centralized server (e.g. processing system 300) also offers an advantage to pattern-recognition systems that employ user-specific adaptation to achieve higher recognition rates. For example, some handwriting recognition techniques develop a handwriting model for each user of the system based on previous recognition results, which is then used to improve the future accuracy of the system for that user (see for example L. Schomaker, H. Teulings, E. Helsper, and G. Abbink, “Adaptive Recognition Of Online, Cursive Handwriting”, Proceedings of the Sixth International Conference on Handwriting and Drawing. Paris, July, 4-7 Telecom, (pp. 19-21), 1993 and S. Connell and A. K. Jain, “Writer Adaptation of Online Handwritten Models,” Proc. 5th International Conference on Document Analysis and Recognition, Bangalore, India, pp. 434-437, September 1999).
This adaptation is more effective if a single server, or set of servers, performs all recognition for a user (rather than a large number of individual applications each performing their own recognition), since the server is able to perform adaptation based on the input generated by all applications. In addition to this, centralized server-based pattern recognition simplifies the management of the recognition system 400 by allowing recognizers to be reconfigured and upgraded without interaction with the distributed client devices 415 and applications 450, and allows training and test data to be easily collected.
However, the information required to perform the context processing stage of a pattern recognition system is generally application specific and is often very large (e.g. entries in a large application-specific database), making it impractical to transmit the context information to a centralized server for processing. A solution to this problem is to use a mechanism for distributed recognition as depicted in
Symbol DAG
One method of returning the intermediate recognition results (i.e. intermediate format data) to an application is to use a symbol DAG (Directed Acyclic Graph), which is a generic data structure that contains symbol and associated scores as vertices, and valid transitions between symbols as edges. The structure can be implemented as a two-dimensional array of elements, each of which defines the output generated by the pattern classifier for a single segment of the signal and the associated valid transitions for that segment. This structure represents all the potential recognition alternatives that may be derived from the input signal based on the results of the classifier. The application uses this structure, in combination with a context model, to decode the input signal.
The symbol DAG is equivalent to a matrix where each column contains the results of the classification of a single segment of the input signal. Each element in the column represents the probability that the classified segment is a particular symbol, and includes an offset that indicates the next possible segment (column) in the input signal that can follow this symbol. Thus, the matrix represents all the possible decoding paths based on the output of the pattern classifier. These paths and associated classification scores can be combined with a context model to fully decode the input signal.
Note that the symbol DAG is applicable in any pattern recognition task where a sequence of classification results is decoded using a context or set of constraints. The symbols contained in the symbol DAG may be any primitive element that is generated as the output of a pattern classifier, including the output from a time-series classifier. Examples of such recognition systems include handwriting and speech recognition, protein sequencing (see A. C. Camproux, P. Tuffery, S. Hazout, “Hidden Markov Model Approach For Identifying The Modular Framework Of The Protein Backbone”, Protein engineering, 12(12), pp. 1063, December 1999), image processing and computer vision (see Y. He, A. Kundu, “2-D Shape Classification Using Hidden Markov Model”, IEEE Transactions on Pattern Analysis, 13(11), November 1991), and econometrics (see T. Ryden, T. Terasvirta, S. Asbrink, “Stylized Facts of Daily Return Series and the Hidden Markov Model”, Journal of Applied Econometrics, 13(3), pp. 217, May 1998).
Symbol DAG Example
As an example, Table 1 shows a symbol DAG that represents the output from a handwritten character recognizer generated by the ambiguous text given in
To decode the alternatives, the context processor starts with the first entry in the DAG (i.e. the character ‘c’). The score for this entry is added to the accumulated total (since log-probabilities are added rather than multiplied), and processing moves to the column given by the offset value in the entry (in this example, column 1). In column 1, two alternatives exist (i.e. “cl” or “cb”), and the scores for these alternatives are found by adding the scores to the previous total. The decoding continues until the end of the DAG is reached. Similarly, the second entry in column 0 (i.e. the character ‘d’) is decoded; note however, that column 1 is skipped in this traversal of the DAG, as indicated by the offset value of 2 in the character score entry. This is due to the letter ‘d’ being constructed using two strokes, and thus the recognition of the letters ‘l’ and ‘b’ cannot be valid in this alternative. Thus, the potential decoding alternatives in this example are:
c log=0.7*0.8*1.0*1.0=0.56
cbg=0.7*0.2*1.0=0.14
dog=0.3*1.0*1.0=0.30
These values can now be combined with a language model or other contextual information to select the most likely word.
The DAG structure must ensure that strokes are assigned to an individual letter only once. To do this, alternate paths must be defined to ensure that if a stroke is assigned to a letter, no subsequent letter may use that stroke in its construction. An example of this is given in
tile=0.6*1.0*0.6*1.0=0.36
tite=0.6*1.0*1.0*1.0=0.60
lite=0.4*1.0*1.0*1.0=0.40
These values can now be combined with a language model to select the most likely word.
Additionally, the character value of a DAG entry can be set to zero, indicating a NUL character (i.e. a character that does not change the text, but will modify the text probability). This allows word break positions (i.e. spaces) to be modeled as a SPACE/NUL pair, indicating that there is a certain probability that a space appears at that point in the DAG. For example:
The potential decoding alternatives in this example are:
ab=1.0*0.6*1.0=0.6
a b=1.0*0.4*1.0=0.4
Distributed Recognizer Management
Referring to
User-Specific Dictionaries
Distributed recognition systems can also support user dictionaries, which are user-specific word lists (and possibly associated a-priori probabilities) that include words that a user writes frequently but which are unlikely to appear in a standard dictionary (examples include company names, work or personal interest specific terms, etc.). User dictionaries can be stored and managed centrally so that words added to the dictionary when using one application are available to all applications for context processing. Obviously, applications can manage and use their own local user-specific dictionaries if required, since they have full control over context decoding.
When an application requires the recognition of a signal that may contain words found in the user dictionary (e.g. standard handwritten text input such as the subject line of an e-mail or an arbitrary voice message), the centralized recognition system generates the usual intermediate recognition results to be returned to the application for context decoding. However, in addition to this it decodes the intermediate results using the user-dictionary as a language model, the result of which is also returned to the application. These two intermediate results structures can be combined by the application during its context decoding to generate a final decoding that includes the user-specific dictionary information.
User-Specific Training
Distributed recognition systems may also support user-specific training for a recognizer 440, as depicted in
When an existing recognizer is upgraded or a new recognizer is added to the system, the recognition manager 710 queries 1010 the training database 930 to determine if any training data 1020 of the format required by the recognizer 440 exists. If so, the training data 1020 is submitted to the newly registered recognizer 440 for processing, as depicted in
The invention may also be said to broadly consist in the parts, elements and features referred to or indicated herein, individually or collectively, in any or all combinations of two or more of the parts, elements or features, and wherein specific integers are mentioned herein which have known equivalents in the art to which the invention relates, such known equivalents are deemed to be incorporated herein as if individually set forth.
Although a preferred embodiment has been described in detail, it should be understood that various changes, substitutions, and alterations can be made by one of ordinary skill in the art without departing from the scope of the present invention. We claim:
Claims
1. A method of recognising digital ink input by a user into a computer-based digital ink recognition system, the user interacting with a paper-based document, the paper-based document having disposed thereon coded data indicative of a particular field of the paper-based document and of at least one reference point of the paper-based document, the method including the steps of:
- capturing by an image sensor in a sensing device images of at least some of the coded data when the sensing device is placed in an operative position relative to the paper-based document;
- decoding by the sensing device at least some of the coded data to form indicating data indicative of the identity of the field of the paper-based document containing the coded data and at least one of a position and a movement of the sensing device relative to the paper-based document;
- receiving in a server the indicating data from the sensing device;
- processing the indicating data using a recognizer residing on the server to produce intermediate format data;
- receiving the intermediate format data in an application;
- decoding the intermediate format data by the application into computer-readable format data using context information associated with the paper-based document.
2. The method as claimed in claim 1, wherein the particular field of the paper-based document is associated with at least one zone of the paper-based document, and the method includes identifying the context information from the at least one zone.
3. The method as claimed in claim 1, wherein the intermediate format data is a Directed Acyclic Graph (DAG) data structure.
4. The method as claimed in claim 3, wherein the DAG data structure is a matrix containing the processing results of segments of the indicating data.
5. The method as claimed in claim 1, wherein the intermediate format data includes segmented time-series classifier data.
6. The method as claimed in claim 1, wherein the context information is a user dictionary.
7. The method as claimed in claim 1, wherein the recognizer is trained for a specific user.
Type: Application
Filed: Jan 21, 2010
Publication Date: May 20, 2010
Applicant:
Inventors: Jonathon Leigh Napper (Balmain), Paul Lapstun (Balmain), Kia Silverbrook (Balmain)
Application Number: 12/691,725
International Classification: G06F 17/27 (20060101); G06K 9/62 (20060101);
