WO2022119702A1 - Vectorisation de corps de document et formation par contraste de bruit - Google Patents

Vectorisation de corps de document et formation par contraste de bruit Download PDF

Info

Publication number
WO2022119702A1
WO2022119702A1 PCT/US2021/059302 US2021059302W WO2022119702A1 WO 2022119702 A1 WO2022119702 A1 WO 2022119702A1 US 2021059302 W US2021059302 W US 2021059302W WO 2022119702 A1 WO2022119702 A1 WO 2022119702A1
Authority
WO
WIPO (PCT)
Prior art keywords
document
documents
search query
corpus
query
Prior art date
Application number
PCT/US2021/059302
Other languages
English (en)
Inventor
Junaid Ahmed
Li Xiong
Arnold OVERWIJK
Chenyan XIONG
Original Assignee
Microsoft Technology Licensing, Llc
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority claimed from US17/207,103 external-priority patent/US11829374B2/en
Application filed by Microsoft Technology Licensing, Llc filed Critical Microsoft Technology Licensing, Llc
Priority to EP21820792.6A priority Critical patent/EP4256443A1/fr
Publication of WO2022119702A1 publication Critical patent/WO2022119702A1/fr

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/332—Query formulation
    • G06F16/3325—Reformulation based on results of preceding query
    • G06F16/3326—Reformulation based on results of preceding query using relevance feedback from the user, e.g. relevance feedback on documents, documents sets, document terms or passages
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/3331—Query processing
    • G06F16/334—Query execution
    • G06F16/3347—Query execution using vector based model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/93—Document management systems
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning

Definitions

  • Machine learning techniques may be used to identify documents that are responsive to a search query. However, depending on the volume of search queries, the length of or data within documents, and/or the amount of documents, the required computing resources may be prohibitive or may otherwise result in unsatisfactory performance.
  • embedding vectors for each document of a document corpus are generated by combining embedding vectors for document subparts of a given document to yield a final embedding vector for the document.
  • a machine learning model is trained using a query corpus and the document corpus, where the model generates a ranking score for a given (query, document) pair.
  • the model is used to generate ranking scores for (query, document) pairs, such that the training dataset used during training is further refined according to the generated ranking scores. For example, a set of top documents and a negative document may be determined for a query in the query corpus and subsequently used as training data accordingly. As training iteratively progresses, multiple negative documents may therefore be determined for a given query. A negative document may be determined for a given query from the associated set of negative documents according to noise-contrastive estimation techniques. Such determined negative documents may then be evaluated as part of a loss function during model training, thereby yielding a more robust machine learning model for machine learning model-based search processing.
  • Figure 1 illustrates an overview of an example system in which the document body vectorization and noise-contrastive training techniques described herein are utilized.
  • Figure 2 illustrates an overview of an example method for training a machine learning model according to the document body vectorization and noise-contrastive training techniques of the present disclosure.
  • Figure 3 illustrates an overview of an example method for processing a document to generate a document score.
  • Figure 4 illustrates an overview of an example method for generating a set of candidate documents responsive to a search query according to aspects of the present disclosure.
  • Figure 5 is a block diagram illustrating example physical components of a computing device with which aspects of the disclosure may be practiced.
  • FIGS. 6A and 6B are simplified block diagrams of a mobile computing device with which aspects of the present disclosure may be practiced.
  • Figure 7 is a simplified block diagram of a distributed computing system in which aspects of the present disclosure may be practiced.
  • Figure 8 illustrates a tablet computing device for executing one or more aspects of the present disclosure.
  • a machine learning model is used to generate a set of documents that is responsive to a search query. For example, an embedding vector may be generated and stored for each document. When a search query is received, a search query embedding vector may be generated, which is compared to the pre-generated document vectors using an evaluation function. As an example, an approximate nearest neighbor (ANN) search may be used. As another example, an interactive machine learning model may be used, where a received query and each document of the set of documents are evaluated online (e.g., rather than using pre-generated document embedding vectors). However, using pre-generated document vectors may yield lower-quality results as compared to interactive techniques, while interactive techniques may be more computationally expensive since each document is evaluated in view of each received search query.
  • ANN approximate nearest neighbor
  • certain approaches may reduce or otherwise limit the amount of data that is used. For example, rather than processing the body of a document, metadata associated with the document (e.g., an anchor, a uniform resource locator (URL), a title of the document, and/or one or more associated clickstreams) may be used to generate a responsive set of documents for a given query instead.
  • metadata associated with the document e.g., an anchor, a uniform resource locator (URL), a title of the document, and/or one or more associated clickstreams
  • URL uniform resource locator
  • aspects of the present disclosure relate to document body vectorization and noise-contrastive training techniques, thereby enabling the efficient use of document bodies for machine learning model-based searching.
  • a machine learning model is trained using a combination of ANCE (Lee Xiong, Chenyan Xiong, Ye Li, Kwok- Fung Tang, Jialin Liu, Paul Bennett, Junaid Ahmed, and Arnold Overwijk. 2020. Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval. arXiv preprint arXiv:2007.00808v2) and NCE (Andriy Mnih and Koray Kavukcuoglu. 2013.
  • the machine learning model is trained using a corpus of search queries and a corpus of documents. Queries and documents may each be encoded using a transformer-based encoding model, such as the universal sentence encoder or BERT (Bidirectional Encoder Representations from Transformers).
  • a transformer-based encoding model such as the universal sentence encoder or BERT (Bidirectional Encoder Representations from Transformers).
  • the disclosed aspects enable a machine learning model to learn distributed representations of search queries and documents, thereby supporting effective and efficient end-to-end machine learning model-based search processing. Additionally, aspects of the present disclosure leverage language model pre-training, such that it may be used as part of search processing. Such aspects are beneficial as compared to the above-discussed interaction-based techniques, which are computationally expensive and may therefore be limited to re-ranking or may comprise costly online inference if used as part of search processing. Further, the disclosed techniques enable the body of a document to be searched effectively, rather than merely using keyword-based searching on a document body or restricting machine learning model-based search processing to document metadata. Finally, such techniques may be generalizable to international markets, even in instances where most training signals may be from English markets and/or markets having other similar languages.
  • FIG. 1 illustrates an overview of an example system 100 in which the document body vectorization and noise-contrastive training techniques described herein are utilized.
  • system 100 comprises server device 102, data source 104, client device 106, and network 108.
  • server device 102, data source 104, and client device 106 communicate using network 108, which may comprise a local area network, a wireless network, or the Internet, or any combination thereof, among other examples.
  • Server device 102 and data source 104 may each be any of a variety of computing devices, including, but not limited to, a server computing device or a set of computing devices that form a distributed computing device.
  • client device 106 may be any of a variety of computing devices, including, but not limited to, a mobile computing device, a laptop computing device, a tablet computing device, or a desktop computing device. It will be appreciated that while system 100 is illustrated as comprising one server device 102, one data source 104, and one client device 106, any number of such elements may be used in other examples. Further, the functionality described herein with respect to server device 102, data source 104, and client device 106 may be distributed among or otherwise implemented on any number of different computing devices in any of a variety of configurations in other examples.
  • Client device 104 is illustrated as comprising client application 120.
  • Client application 120 may be any of a variety of applications, such as a web application executing in a web browser, a native application, or a combination thereof.
  • a user of client device 104 may use client application 120 to identify a set of documents that are responsive to a search query.
  • Client application 120 may receive a search query from a user, which may be provided to server device 102.
  • Server device 102 may process the search query (e.g., using query processor 116) to determine a set of documents that is responsive to the search query. Accordingly, client application 120 may receive an indication of documents that are responsive to the search query, which may be presented to the user.
  • documents identified as being responsive to a search query are stored or otherwise provided by a data source, such as data source 104.
  • document store 118 of data source 104 may store any of a variety of documents, including, but not limited to, text documents, audio files, video files, and/or webpages of a website.
  • a document may comprise any of a variety of data types.
  • a document may have a body (e.g., the content of the document) and associated metadata (e.g., a title or filename, a URL via which the document is available, a last-modified date, etc.).
  • an email document may have a body, a subject line, one or more attachments, a sender, and/or one or more recipients.
  • document store 118 is illustrated as part of data source 104, it will be appreciated that, in other examples, a document store need not be remote from server device 102 and/or client device 106.
  • aspects of the present disclosure may be implemented to provide server-side and/or client-side machine learning-based search processing, among other examples.
  • Server device 102 is illustrated as comprising document vectorizer 110, model training engine 112, data store 114, and query processor 116.
  • a document corpus e.g., as may comprise documents of one or more document stores such as document store 118
  • Documents of the corpus are encoded to generate document encoding vectors (e.g., using a transformer-based encoding model, such as the universal sentence encoder or BERT) that may be used as an input during model training and subsequent inference.
  • a transformer-based encoding model such as the universal sentence encoder or BERT
  • the length of a document may be such that it cannot be encoded into a single encoding vector.
  • document vectorizer 110 may generate a bag of encodings for a given document, where each encoding is associated with a subpart of the document (e.g., according to sentence breaks, paragraph breaks, a predetermined number of words or bytes, etc.). The bag of encodings may then be used to generate a final encoding for the document.
  • the machine learning model may generate a weight for each token, such that the weights are used to sum the vector of each respective token in the bag of encodings.
  • Encodings for a given document e.g., an associated final encoding and/or bag of encodings for a given document
  • Model training engine 112 trains a model according to the aspects described herein.
  • the model may be trained according to a corpus of search queries and a corpus of documents.
  • the model may take an input comprising a search query and a document.
  • the model may subsequently generate an output comprising a ranking score for the (search query, document) pair.
  • the corpus of search queries with which the model is trained (e.g., as may be stored by data store 114) may comprise historical search queries and/or representative search queries, among other examples.
  • each query of the corpus of search queries further comprises an association with a “positive” document, which is a document that was determined to be responsive to the search query (e.g., a document that ended a user’s clickstream for that search or a document explicitly indicated by the user as being responsive to the user’s search query).
  • a “positive” document is a document that was determined to be responsive to the search query (e.g., a document that ended a user’s clickstream for that search or a document explicitly indicated by the user as being responsive to the user’s search query).
  • model training engine 112 uses ANCE and NCE techniques to train the model.
  • model training engine 112 may perform a predetermined number of training steps, after which the model is used to perform inference using the full document corpus to generate ranking scores for queries in the corpus of training queries.
  • model training engine 112 generates a set of relevant documents for the queries using the model in its current state.
  • a subset of top documents may be selected for each query (e.g., according to a predetermined number or above a predetermined ranking score threshold). Additionally, a negative document for each query may be randomly determined from the subset of top documents.
  • the training corpus may be updated to comprise the subset of top documents and the corpus of training queries, such that model training resumes for the predetermined number of training steps.
  • model training engine 112 uses the same data distribution for training as is ultimately used for inference generation. Additionally, such noise-contrastive techniques may improve model performance as a result of learning from stronger negative examples (e.g., as were ranked using the current state of the model). Further, as a result of using ANCE, each training batch yields, for each query, a positive document and a set of negative documents (e.g., as are selected above). NCE may further be used, where the set of negative documents for a search query is processed to select a final negative document for the query from the set of negative documents, which may then be used in an associated loss function. For example, the loss function used by model training engine 112 may leverage binary cross-entropy, as shown in the example loss function below:
  • M(d + ) is a positive document vector (e.g., a document that is responsive to a given query q, as may be determined from a user’s clickstream or based on a user indication)
  • M(d_) is a negative document vector (e.g., as may be obtained using ACNE and/or NCE, as described above)
  • M(q) is a query vector
  • cos (it, v) denotes the cosine similarity of u and v.
  • Variables w, and k 2 are constants that can be used to tune the loss function. For example, a larger value for w may ultimately yield a less robust model.
  • query processor 116 uses the trained model to process queries (e.g., as may be received from client application 120 of client device 106) and generate a set of documents that is responsive to the query accordingly.
  • query processor 116 generates an embedding vector for a received query.
  • data store 114 may store embedding vectors for documents, such that query processor 116 uses the trained model to process the query embedding vector in view of the document embedding vectors stored by data store 114, thereby generating a responsive set of documents.
  • a dot-product ANN search may be used, such that query processor 116 generates a set of documents responsive to the query that may be returned to client device 106.
  • the returned set of documents may comprise references to the identified documents and/or excerpts from the documents, among other examples.
  • the excerpts may be relevant keywords or excerpts from the documents.
  • Figure 2 illustrates an overview of an example method 200 for training a machine learning model according to the document body vectorization and noise-contrastive training techniques of the present disclosure.
  • aspects of method 200 are performed by a model training engine, such as model training engine 112 in Figure 1.
  • Method 200 begins at operation 202, where a training dataset comprising a query corpus and a document corpus is obtained.
  • operation 202 comprises accessing the query corpus and/or document corpus from a data store, such as data store 114 in Figure 1.
  • the query corpus comprises an association between each query and a positive document in the document corpus, as discussed above.
  • operation 204 comprises performing a predetermined number of training steps, after which it may be determined that the predetermined number of training steps (or other threshold) has been met.
  • the current model e.g., as was trained at operation 204 is used to perform inference of documents in the document corpus for queries of the query corpus.
  • the model may generate a ranking score for each (query, document) pair, such that a set of highest-ranked documents may be determined for a given query (e.g., comprising a predetermined number of documents or with ranking scores above a predetermined ranking score threshold).
  • a negative (query, document) pair is selected for each query of the query corpus.
  • the negative (query, document) pair is randomly selected from the set of highest-ranked documents that was generated at operation 206.
  • the negative (query, document) pair may be determined based on an associated ranking score that was generated at operation 206.
  • any of a variety of techniques may be used to generate the set of highest-ranked documents at operation 206 and, similarly, to select a negative (query, document) pair for a given query at operation 208.
  • the new training dataset may comprise the set of highest-ranked documents for each query of the query corpus (e.g., as was generated at operation 206).
  • the training dataset may further comprise the negative documents that were selected at operation 208, which may be used as negative examples for training the model at operation 204.
  • operations 206-210 are an example of using the noise-contrastive techniques of ANCE to improve model performance as a result of learning from stronger negative examples (e.g., as were ranked using the current state of the model generated at operation 204).
  • An arrow is illustrated from operation 210 to operation 204 to indicate that flow may loop between operations 204-210, such that the described ANCE techniques are applied after every N training steps performed at operation 204.
  • operation 208 comprises selecting a negative (query, document) pair for each query.
  • operation 208 comprises selecting a negative (query, document) pair for each query.
  • a final negative document is determined for each search query at operation 214.
  • the final negative document may be sampled from the set of negative documents for a given query that was generated at operation 212 according to NCE sampling. It will be appreciated that any of a variety of other techniques may be used to determine a negative document for a given query using the set of negative documents generated at operation 212.
  • An example loss function is described above, where a positive document vector (e.g., as may be indicated by the query corpus), a negative document vector (e.g., as was selected at operation 214), and a query vector are evaluated according to a cosine similarity function.
  • the example loss function is tuned using several hyperparameters, which may adjust the respective contributions of the query / positive document similarity and query / negative document similarity.
  • any of a variety of other loss functions may be used.
  • An arrow is illustrated from operation 216 to operation 204 to similarly indicate that the training illustrated by method 200 is iterative, such that operations 204-216 may be performed multiple times to train the machine learning model.
  • operations 204- 216 are performed a predetermined number of times or, as another example, operations 204- 216 are performed until model performance converges. Flow eventually terminates at operation 216.
  • Figure 3 illustrates an overview of an example method 300 for processing a document to generate a document score.
  • aspects of method 300 are performed by a document vectorizer, such as document vectorizer 110 in Figure 1.
  • Method 300 begins at operation 302, where a body of a document is obtained.
  • the body is obtained from a data source, such as document store 118 of data source 104.
  • the document may be from a data store (e.g., data store 114 in Figure 1).
  • the document body may have been cached from the data source in the data store.
  • a document body may be obtained from any of a variety of sources using any of a variety of techniques.
  • an embedding vector is generated for each subpart of the document body, thereby yielding a bag of encodings for the document.
  • the length of a document may be such that a single embedding vector may not be generated without summarizing, downscaling, or otherwise potentially losing information therein.
  • multiple embedding vectors may be used for the document.
  • the document may be split into multiple subparts (e.g., according to sentence breaks, paragraph breaks, a predetermined number of words or bytes, etc.).
  • a vector may be projected into a different dimension using linear projection, such that an average or other combination of the encoding vectors in the different dimension may be generated at operation 306.
  • Embedding vectors for the document subparts may be generated using a transformer-based encoding model, such as the universal sentence encoder or BERT.
  • a final document embedding vector is generated based on the embedding vectors for the constituent subparts of the document that were generated at operation 304. For example, max-pooling techniques may be used to aggregate embedding vectors for the document subparts into a single embedding vector for the document.
  • the encoding model that was applied at operation 304 may further generate a weight for each token of the embedding vector, such that the scalars are passed into softmax operation to generate a probability distribution for each token accordingly. These weights may then be used to sum the vector of each respective token to form the final single vector representation.
  • the final document vector is provided.
  • the document vector may be provided for use while training a model, such as at operation 206 of method 200 in Figure 2.
  • the final document vector may be stored for later use (e.g., during training and inference) in a data store, such as data store 114 in Figure 1. Flow terminates at operation 308.
  • a document body need not be split into multiple subparts according to method 300. Rather, in other instances, a document instead be summarized or otherwise processed to generate a single embedding vector that is representative of the document body.
  • Figure 4 illustrates an overview of an example method 400 for generating a set of candidate documents responsive to a search query according to aspects of the present disclosure.
  • aspects of method 400 are performed by a query processor, such as query processor 116 in Figure 1.
  • Method 400 begins at operation 402, where a search query is obtained.
  • a search query may be received from a client device, such as client device 106 in Figure 1.
  • the search query may comprise one or more keywords or may be a sentence, among other examples.
  • the query is encoded to generate a query embedding vector.
  • the query embedding vector may be generated using a transformer-based encoding model, such as the universal sentence encoder or BERT.
  • Flow progresses to operation 406, where a set of documents is determined using the query embedding vector.
  • a dot-product ANN search may be used, such that document embedding vectors are identified using the query embedding vector, thereby generating a set of documents that is responsive to the search query.
  • the document vectors processed at operation 406 may each be a final document vector based on a set of encoding vectors or bag of encodings, as discussed above with respect to operations 304 and 306 of method 300 in Figure 3.
  • the document vectors may have been pre-generated (e.g., by a document vectorizer, such as document vectorizer 110 in Figure 1).
  • the generated set of document sis provided.
  • the set of documents may be provided to the client device in response to the search query that was received at operation 402.
  • Providing the set of documents may comprise providing references to the identified documents and/or excerpts from the documents, among other examples.
  • the excerpts may be relevant keywords from the documents. Flow terminates at operation 408.
  • Figures 5-8 and the associated descriptions provide a discussion of a variety of operating environments in which aspects of the disclosure may be practiced. However, the devices and systems illustrated and discussed with respect to Figures 5-8 are for purposes of example and illustration and are not limiting of a vast number of computing device configurations that may be utilized for practicing aspects of the disclosure, described herein.
  • Figure 5 is a block diagram illustrating physical components (e.g., hardware) of a computing device 500 with which aspects of the disclosure may be practiced.
  • the computing device components described below may be suitable for the computing devices described above, including devices 102, 104, and 106 in Figure 1.
  • the computing device 500 may include at least one processing unit 502 and a system memory 504.
  • the system memory 504 may comprise, but is not limited to, volatile storage (e.g., random access memory), nonvolatile storage (e.g., read-only memory), flash memory, or any combination of such memories.
  • the system memory 504 may include an operating system 505 and one or more program modules 506 suitable for running software application 520, such as one or more components supported by the systems described herein.
  • system memory 504 may store document vectorizer 524 and training engine 526.
  • the operating system 505, for example, may be suitable for controlling the operation of the computing device 500.
  • inventions of the disclosure may be practiced in conjunction with a graphics library, other operating systems, or any other application program and is not limited to any particular application or system.
  • This basic configuration is illustrated in Figure 5 by those components within a dashed line 508.
  • the computing device 500 may have additional features or functionality.
  • the computing device 500 may also include additional data storage devices (removable and/or non-removable) such as, for example, magnetic disks, optical disks, or tape.
  • additional storage is illustrated in Figure 5 by a removable storage device 509 and a non-removable storage device 510.
  • program modules 506 may perform processes including, but not limited to, the aspects, as described herein.
  • Other program modules may include electronic mail and contacts applications, word processing applications, spreadsheet applications, database applications, slide presentation applications, drawing or computer-aided application programs, etc.
  • embodiments of the disclosure may be practiced in an electrical circuit comprising discrete electronic elements, packaged or integrated electronic chips containing logic gates, a circuit utilizing a microprocessor, or on a single chip containing electronic elements or microprocessors.
  • embodiments of the disclosure may be practiced via a system-on-a-chip (SOC) where each or many of the components illustrated in Figure 5 may be integrated onto a single integrated circuit.
  • SOC system-on-a-chip
  • Such an SOC device may include one or more processing units, graphics units, communications units, system virtualization units and various application functionality all of which are integrated (or “burned”) onto the chip substrate as a single integrated circuit.
  • the functionality, described herein, with respect to the capability of client to switch protocols may be operated via application-specific logic integrated with other components of the computing device 500 on the single integrated circuit (chip).
  • Embodiments of the disclosure may also be practiced using other technologies capable of performing logical operations such as, for example, AND, OR, and NOT, including but not limited to mechanical, optical, fluidic, and quantum technologies.
  • embodiments of the disclosure may be practiced within a general purpose computer or in any other circuits or systems.
  • the computing device 500 may also have one or more input device(s) 512 such as a keyboard, a mouse, a pen, a sound or voice input device, a touch or swipe input device, etc.
  • the output device(s) 514 such as a display, speakers, a printer, etc. may also be included.
  • the aforementioned devices are examples and others may be used.
  • the computing device 500 may include one or more communication connections 516 allowing communications with other computing devices 550. Examples of suitable communication connections 516 include, but are not limited to, radio frequency (RF) transmitter, receiver, and/or transceiver circuitry; universal serial bus (USB), parallel, and/or serial ports.
  • RF radio frequency
  • USB universal serial bus
  • Computer readable media may include computer storage media.
  • Computer storage media may include volatile and nonvolatile, removable and nonremovable media implemented in any method or technology for storage of information, such as computer readable instructions, data structures, or program modules.
  • the system memory 504, the removable storage device 509, and the non-removable storage device 510 are all computer storage media examples (e.g., memory storage).
  • Computer storage media may include RAM, ROM, electrically erasable read-only memory (EEPROM), flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other article of manufacture which can be used to store information and which can be accessed by the computing device 500. Any such computer storage media may be part of the computing device 500.
  • Computer storage media does not include a carrier wave or other propagated or modulated data signal.
  • Communication media may be embodied by computer readable instructions, data structures, program modules, or other data in a modulated data signal, such as a carrier wave or other transport mechanism, and includes any information delivery media.
  • modulated data signal may describe a signal that has one or more characteristics set or changed in such a manner as to encode information in the signal.
  • communication media may include wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, radio frequency (RF), infrared, and other wireless media.
  • RF radio frequency
  • FIGs 6A and 6B illustrate a mobile computing device 600, for example, a mobile telephone, a smart phone, wearable computer (such as a smart watch), a tablet computer, a laptop computer, and the like, with which embodiments of the disclosure may be practiced.
  • the client may be a mobile computing device.
  • FIG 6A one aspect of a mobile computing device 600 for implementing the aspects is illustrated.
  • the mobile computing device 600 is a handheld computer having both input elements and output elements.
  • the mobile computing device 600 typically includes a display 605 and one or more input buttons 610 that allow the user to enter information into the mobile computing device 600.
  • the display 605 of the mobile computing device 600 may also function as an input device (e.g., a touch screen display).
  • an optional side input element 615 allows further user input.
  • the side input element 615 may be a rotary switch, a button, or any other type of manual input element.
  • mobile computing device 600 may incorporate more or less input elements.
  • the display 605 may not be a touch screen in some embodiments.
  • the mobile computing device 600 is a portable phone system, such as a cellular phone.
  • the mobile computing device 600 may also include an optional keypad 635.
  • Optional keypad 635 may be a physical keypad or a “soft” keypad generated on the touch screen display.
  • the output elements include the display 605 for showing a graphical user interface (GUI), a visual indicator 620 (e.g., a light emitting diode), and/or an audio transducer 625 (e.g., a speaker).
  • GUI graphical user interface
  • the mobile computing device 600 incorporates a vibration transducer for providing the user with tactile feedback.
  • the mobile computing device 600 incorporates input and/or output ports, such as an audio input (e.g., a microphone jack), an audio output (e.g., a headphone jack), and a video output (e.g., a HDMI port) for sending signals to or receiving signals from an external device.
  • FIG. 6B is a block diagram illustrating the architecture of one aspect of a mobile computing device. That is, the mobile computing device 600 can incorporate a system (e.g., an architecture) 602 to implement some aspects.
  • the system 602 is implemented as a “smart phone” capable of running one or more applications (e.g., browser, e-mail, calendaring, contact managers, messaging clients, games, and media clients/players).
  • the system 602 is integrated as a computing device, such as an integrated personal digital assistant (PDA) and wireless phone.
  • PDA personal digital assistant
  • One or more application programs 666 may be loaded into the memory 662 and run on or in association with the operating system 664. Examples of the application programs include phone dialer programs, e-mail programs, personal information management (PIM) programs, word processing programs, spreadsheet programs, Internet browser programs, messaging programs, and so forth.
  • the system 602 also includes a non-volatile storage area 668 within the memory 662. The non-volatile storage area 668 may be used to store persistent information that should not be lost if the system 602 is powered down.
  • the application programs 666 may use and store information in the non-volatile storage area 668, such as e-mail or other messages used by an e-mail application, and the like.
  • a synchronization application (not shown) also resides on the system 602 and is programmed to interact with a corresponding synchronization application resident on a host computer to keep the information stored in the non-volatile storage area 668 synchronized with corresponding information stored at the host computer.
  • other applications may be loaded into the memory 662 and run on the mobile computing device 600 described herein (e.g., search engine, extractor module, relevancy ranking module, answer scoring module, etc.).
  • the system 602 has a power supply 670, which may be implemented as one or more batteries.
  • the power supply 670 might further include an external power source, such as an AC adapter or a powered docking cradle that supplements or recharges the batteries.
  • the system 602 may also include a radio interface layer 672 that performs the function of transmitting and receiving radio frequency communications.
  • the radio interface layer 672 facilitates wireless connectivity between the system 602 and the “outside world,” via a communications carrier or service provider. Transmissions to and from the radio interface layer 672 are conducted under control of the operating system 664. In other words, communications received by the radio interface layer 672 may be disseminated to the application programs 666 via the operating system 664, and vice versa.
  • the visual indicator 620 may be used to provide visual notifications, and/or an audio interface 674 may be used for producing audible notifications via the audio transducer 625.
  • the visual indicator 620 is a light emitting diode (LED) and the audio transducer 625 is a speaker. These devices may be directly coupled to the power supply 670 so that when activated, they remain on for a duration dictated by the notification mechanism even though the processor 660 and other components might shut down for conserving battery power.
  • the LED may be programmed to remain on indefinitely until the user takes action to indicate the powered-on status of the device.
  • the audio interface 674 is used to provide audible signals to and receive audible signals from the user.
  • the audio interface 674 may also be coupled to a microphone to receive audible input, such as to facilitate a telephone conversation.
  • the microphone may also serve as an audio sensor to facilitate control of notifications, as will be described below.
  • the system 602 may further include a video interface 676 that enables an operation of an on-board camera 630 to record still images, video stream, and the like.
  • a mobile computing device 600 implementing the system 602 may have additional features or functionality.
  • the mobile computing device 600 may also include additional data storage devices (removable and/or non-removable) such as, magnetic disks, optical disks, or tape. Such additional storage is illustrated in Figure 6B by the non-volatile storage area 668.
  • Data/information generated or captured by the mobile computing device 600 and stored via the system 602 may be stored locally on the mobile computing device 600, as described above, or the data may be stored on any number of storage media that may be accessed by the device via the radio interface layer 672 or via a wired connection between the mobile computing device 600 and a separate computing device associated with the mobile computing device 600, for example, a server computer in a distributed computing network, such as the Internet.
  • a server computer in a distributed computing network such as the Internet.
  • data/information may be accessed via the mobile computing device 600 via the radio interface layer 672 or via a distributed computing network.
  • data/information may be readily transferred between computing devices for storage and use according to well-known data/information transfer and storage means, including electronic mail and collaborative data/information sharing systems.
  • Figure 7 illustrates one aspect of the architecture of a system for processing data received at a computing system from a remote source, such as a personal computer 704, tablet computing device 706, or mobile computing device 708, as described above.
  • Content displayed at server device 702 may be stored in different communication channels or other storage types.
  • various documents may be stored using a directory service 722, a web portal 724, a mailbox service 726, an instant messaging store 728, or a social networking site 730.
  • a query processor 720 may be employed by a client that communicates with server device 702, and/or training engine 721 may be employed by server device 702.
  • the server device 702 may provide data to and from a client computing device such as a personal computer 704, a tablet computing device 706 and/or a mobile computing device 708 (e.g., a smart phone) through a network 715.
  • client computing device such as a personal computer 704, a tablet computing device 706 and/or a mobile computing device 708 (e.g., a smart phone) through a network 715.
  • the computer system described above may be embodied in a personal computer 704, a tablet computing device 706 and/or a mobile computing device 708 (e.g., a smart phone).
  • FIG. 8 illustrates an exemplary tablet computing device 800 that may execute one or more aspects disclosed herein.
  • the aspects and functionalities described herein may operate over distributed systems (e.g., cloud-based computing systems), where application functionality, memory, data storage and retrieval and various processing functions may be operated remotely from each other over a distributed computing network, such as the Internet or an intranet.
  • User interfaces and information of various types may be displayed via on-board computing device displays or via remote display units associated with one or more computing devices.
  • user interfaces and information of various types may be displayed and interacted with on a wall surface onto which user interfaces and information of various types are projected.
  • Interaction with the multitude of computing systems with which embodiments of the invention may be practiced include, keystroke entry, touch screen entry, voice or other audio entry, gesture entry where an associated computing device is equipped with detection (e.g., camera) functionality for capturing and interpreting user gestures for controlling the functionality of the computing device, and the like.
  • some embodiments include a system (e.g., 500, 600) comprising: at least one processor (e.g., 502, 660, 661); and memory (e.g., 504, 662) storing instructions that, when executed by the at least one processor, causes the system to perform a set of operations.
  • a system e.g., 500, 600
  • processor e.g., 502, 660, 661
  • memory e.g., 504, 662
  • the set of operations comprises: training (e.g., Figure 2) a machine learning model based on a training dataset comprising a search query corpus and the document corpus, wherein training the machine learning model comprises: generating (e.g., 206), using the machine learning model, a set of ranking scores for documents of the document corpus based on a first search query of the search query corpus; refining (e.g., 208, 210) the training dataset based on the generated set of ranking scores; determining (e.g., 214) a first negative document from a set of negative documents for the first search query; and evaluating (e.g., 216) a loss function using the first negative document to train the machine learning model; obtaining (e.g., 402) a request comprising a second search query; generating (e.g., 404, 406), using the trained machine learning model, a set of documents from the document corpus that is responsive to the second search query; and providing (e.g., 408), in response to the request
  • the set of operations further comprises: generating (e.g., Figure 3), for each document of the document corpus, a document embedding vector based on an embedding vector of at least one subpart of the document.
  • refining the training dataset comprises: retaining (e.g., 210), for the first search query, a subset of documents of the document corpus in the training dataset based on the set of ranking scores; and determining (e.g., 214) a second negative document for the first search query from the document corpus, wherein the second negative document is part of the set of negative documents for the first search query.
  • the second negative document is randomly determined.
  • the first negative document is determined (e.g., 214) from the set of negative documents for the first search query using noise-contrastive estimation.
  • the loss function evaluates a first cosine similarity between a query embedding vector for the first search query and a first document embedding vector for the first negative document.
  • the loss function further evaluates a second cosine similarity between the query embedding vector and a second document embedding vector for a positive document associated with the first search query.
  • generating (e.g., 406) the set of documents that is responsive to the second search query comprises: performing an approximate nearest neighbor search using a query embedding vector for the second search query and document embedding vectors for documents of the document corpus to generate the set of documents; and ranking the set of documents according to associated ranking scores.
  • some embodiments include a method (e.g., Figure 4) for generating a set of documents responsive to a search query.
  • the method comprises: obtaining (e.g., 402) a request comprising a search query; generating (e.g., 404) a query embedding vector for the search query; generating (e.g., 406), based on the query embedding vector and document embedding vectors for documents of a document corpus, a set of documents responsive to the search query; ranking (e.g., 406) the set of documents according to associated ranking scores; and providing (e.g., 408), in response to the request, the ranked set of documents that is responsive to the search query.
  • generating the set of documents responsive to the search query comprises processing the query embedding vector and the document embedding vectors using an approximate nearest neighbor search.
  • a document embedding vector for a document of the document corpus is a pre-generated document embedding vector based on a plurality of embedding vectors, wherein each embedding vector of the plurality of embedding vectors is associated with a subpart of the document.
  • a document embedding vector for a document of the document corpus is associated with a body of the document.
  • providing the ranked set of documents comprises providing a subpart of a document in the ranked set of documents.
  • some embodiments include a method for machine learning model-based search processing.
  • the method comprises: training (e.g., Figure 2) a machine learning model based on a training dataset comprising a search query corpus and the document corpus, wherein training the machine learning model comprises: generating (e.g., 206), using the machine learning model, a set of ranking scores for documents of the document corpus based on a first search query of the search query corpus; refining (e.g., 208, 210) the training dataset based on the generated set of ranking scores; determining (e.g., 214) a first negative document from a set of negative documents for the first search query; and evaluating (e.g., 216) a loss function using the first negative document to train the machine learning model; obtaining (e.g., 402) a request comprising a second search query; generating (e.g., 404, 406), using the trained machine learning model, a set of documents from the document
  • the method further comprises: generating (e.g., Figure 3), for each document of the document corpus, a document embedding vector based on an embedding vector of at least one subpart of the document.
  • refining the training dataset comprises: retaining (e.g., 210), for the first search query, a subset of documents of the document corpus in the training dataset based on the set of ranking scores; and determining (e.g., 214) a second negative document for the first search query from the document corpus, wherein the second negative document is part of the set of negative documents for the first search query.
  • the first negative document is determined (e.g., 214) from the set of negative documents for the first search query using noise-contrastive estimation.
  • the loss function evaluates a first cosine similarity between a query embedding vector for the first search query and a first document embedding vector for the first negative document.
  • the loss function further evaluates a second cosine similarity between the query embedding vector and a second document embedding vector for a positive document associated with the first search query.
  • generating (e.g., 406) the set of documents that is responsive to the second search query comprises: performing an approximate nearest neighbor search using a query embedding vector for the second search query and document embedding vectors for documents of the document corpus to generate the set of documents; and ranking the set of documents according to associated ranking scores.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Databases & Information Systems (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Business, Economics & Management (AREA)
  • General Business, Economics & Management (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

Des vecteurs d'incorporation de document associés à chaque document d'un corpus peuvent être générés en combinant des vecteurs d'incorporation associés à des sous-parties de document, ce qui produit un vecteur d'incorporation final associé au document. Un modèle d'apprentissage machine est formé en utilisant un corpus d'interrogations et le corpus de documents. Le modèle génère un score de classement pour une paire donnée (interrogation, document). Pendant la formation, les scores de classement sont générés au moyen du modèle, de sorte que l'ensemble de données de formation est encore affiné en utilisant les scores de classement générés. Par exemple, des documents supérieurs et un document négatif peuvent être déterminés pour une interrogation donnée puis utilisés comme données de formation. De multiples documents négatifs peuvent donc être déterminés pour une interrogation donnée. Un document négatif pour une interrogation donnée peut être déterminé à partir des documents négatifs à l'aide d'une estimation de contraste de bruit. Ces documents négatifs déterminés peuvent être évalués au moyen d'une fonction de perte pendant une formation du modèle, ce qui produit un modèle de traitement de recherche plus robuste.
PCT/US2021/059302 2020-12-04 2021-11-15 Vectorisation de corps de document et formation par contraste de bruit WO2022119702A1 (fr)

Priority Applications (1)

Application Number Priority Date Filing Date Title
EP21820792.6A EP4256443A1 (fr) 2020-12-04 2021-11-15 Vectorisation de corps de document et formation par contraste de bruit

Applications Claiming Priority (4)

Application Number Priority Date Filing Date Title
US202063121516P 2020-12-04 2020-12-04
US63/121,516 2020-12-04
US17/207,103 US11829374B2 (en) 2020-12-04 2021-03-19 Document body vectorization and noise-contrastive training
US17/207,103 2021-03-19

Publications (1)

Publication Number Publication Date
WO2022119702A1 true WO2022119702A1 (fr) 2022-06-09

Family

ID=78824985

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2021/059302 WO2022119702A1 (fr) 2020-12-04 2021-11-15 Vectorisation de corps de document et formation par contraste de bruit

Country Status (1)

Country Link
WO (1) WO2022119702A1 (fr)

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190205761A1 (en) * 2017-12-28 2019-07-04 Adeptmind Inc. System and method for dynamic online search result generation
JP2019164409A (ja) * 2018-03-19 2019-09-26 株式会社日立ソリューションズ 文書検索装置、文書検索方法、及び文書検索プログラム

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190205761A1 (en) * 2017-12-28 2019-07-04 Adeptmind Inc. System and method for dynamic online search result generation
JP2019164409A (ja) * 2018-03-19 2019-09-26 株式会社日立ソリューションズ 文書検索装置、文書検索方法、及び文書検索プログラム

Similar Documents

Publication Publication Date Title
US11200269B2 (en) Method and system for highlighting answer phrases
US9965465B2 (en) Distributed server system for language understanding
US20230186094A1 (en) Probabilistic neural network architecture generation
US11603017B2 (en) Query rewriting and interactive inquiry framework
US10528632B2 (en) Systems and methods for responding to an online user query
US20190012373A1 (en) Conversational/multi-turn question understanding using web intelligence
US11899675B2 (en) Machine reading comprehension system for answering queries related to a document
US10635733B2 (en) Personalized user-categorized recommendations
US20160335261A1 (en) Ranking for efficient factual question answering
US20220100676A1 (en) Dynamic cache management in beam search
US11829374B2 (en) Document body vectorization and noise-contrastive training
EP4359950A1 (fr) Plate-forme de données hétérogènes
WO2023003675A1 (fr) Système de base de connaissances d'entreprise pour médiation communautaire
US11762863B2 (en) Hierarchical contextual search suggestions
US11921728B2 (en) Performing targeted searching based on a user profile
WO2022119702A1 (fr) Vectorisation de corps de document et formation par contraste de bruit
WO2018125770A1 (fr) Système d'aperçu contextuel
WO2022099566A1 (fr) Modèle d'injection de connaissances pour raisonnement commun génératif
US20230343329A1 (en) Keyword Detection for Audio Content
US20240184790A1 (en) Performing targeted searching based on a user profile
WO2023204898A1 (fr) Détection de mot-clé pour contenu audio
WO2023235072A1 (fr) Apprentissage contrastif supervisé pour recommandation de contenu associé

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21820792

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2021820792

Country of ref document: EP

Effective date: 20230704