WO2019223362A1 - 自动问答方法及装置 - Google Patents

自动问答方法及装置 Download PDF

Info

Publication number
WO2019223362A1
WO2019223362A1 PCT/CN2019/073662 CN2019073662W WO2019223362A1 WO 2019223362 A1 WO2019223362 A1 WO 2019223362A1 CN 2019073662 W CN2019073662 W CN 2019073662W WO 2019223362 A1 WO2019223362 A1 WO 2019223362A1
Authority
WO
WIPO (PCT)
Prior art keywords
question
sentence
question sentence
sample
user
Prior art date
Application number
PCT/CN2019/073662
Other languages
English (en)
French (fr)
Inventor
陈华杰
Original Assignee
北京国双科技有限公司
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
Application filed by 北京国双科技有限公司 filed Critical 北京国双科技有限公司
Publication of WO2019223362A1 publication Critical patent/WO2019223362A1/zh

Links

Images

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

Definitions

  • the present invention relates to the field of information processing, and more specifically, to a method and a device for automatic question answering.
  • the automatic question answering technology through the form of intelligent assistants or intelligent customer service, automatically answers questions encountered by users in handling related affairs, which greatly reduces the workload of manual customer service, and also allows users to enjoy faster Q & A services.
  • the traditional automatic question answering scheme converts the automatic question answering process into the user question retrieval process in the common question database. Specifically, it uses natural language processing tools to analyze user question sentences, identifies the keywords of user question sentences, and according to the obtained keywords The keywords retrieve the matching questions and answers from the FAQ database.
  • the machine learning methods used in traditional automated question answering schemes are usually based on the bag-of-words model, which ignores the order relationship between the words in the question sentence and loses some of the semantic information in the question sentence, making the machine learning process.
  • the lack of interpretability with machine learning results also leads to the lack of interpretability of the automated question answering process and results.
  • the present invention is provided in order to provide an automatic question answering method and device that overcome the above problems or at least partially solve the above problems.
  • the present invention provides the following technical solutions:
  • An automatic question answering method including:
  • a question sentence matching the user question sentence is determined from the candidate question set as a target question sentence; wherein the question model is a history question sentence as a training sample to focus attention Neural network training;
  • the acquiring the user question sentence and the candidate question set includes:
  • At least one question sentence associated with the user question sentence is retrieved from a preset question bank as a candidate question set; the question sentences in the preset question bank all correspond to preset answer information.
  • the use of a preset question model to determine a question sentence that matches the user question sentence from the candidate question set, and the target question sentence includes:
  • a question sentence in which the similarity meets a preset similarity condition in the candidate question set is determined as a question sentence that matches the user question sentence as a target question sentence.
  • the training process of the problem model includes:
  • training samples including sample question sentences
  • the sample problem sentence is used to train an attention neural network to obtain the problem model.
  • the obtaining training samples includes:
  • the third question sentence as a negative sample of the first question sentence; wherein the positive sample and the negative sample are training samples .
  • training the attention neural network by using the sample question sentence to obtain the question model includes:
  • a recurrent neural network with a two-way gate structure is used to extract a feature vector corresponding to each moment in the sample question sentence;
  • the method further includes:
  • the obtaining a word vector corresponding to each word in the sample question sentence includes:
  • a word vector corresponding to each word in the sample question sentence is obtained.
  • An automatic question answering device includes:
  • a question obtaining unit for obtaining a user question sentence and a candidate question set, wherein the question sentences in the candidate question set each correspond to preset answer information
  • a question determination unit is configured to determine a question sentence that matches the user question sentence from the candidate question set by using a preset question model as a target question sentence, wherein the question model is a history question sentence It is obtained by training the attention neural network as a training sample;
  • a question response unit is configured to respond to the user question sentence by using answer information corresponding to the target question sentence.
  • a storage medium includes a stored program, and when the program runs, a device where the storage medium is located is controlled to execute the foregoing automatic question answering method.
  • a processor is used to run a program, and when the program runs, the above-mentioned automatic question answering method is performed.
  • the automatic question answering method and device provided by the present invention first obtain user question sentences and candidate question sets, and then use the question model obtained by training the question sentence training samples on the attention neural network to extract the candidate questions
  • the question sentence matching the user question sentence is collectively determined as the target question sentence, and finally the answer information corresponding to the target question sentence is used to respond to the user question sentence, and the characteristics of the attention neural network are used to make
  • the automatic question-and-answer process and results realized by using the question model are more interpretable and improve the effect of the automatic question-and-answer.
  • FIG. 1 is a flowchart of an automatic question answering method provided by an embodiment of the present application
  • FIG. 2 is another flowchart of an automatic question answering method provided by an embodiment of the present application.
  • FIG. 3 is a flowchart of a problem model training process according to an embodiment of the present application.
  • FIG. 5 is a schematic diagram of sentence encoding provided by an embodiment of the present application.
  • FIG. 6 is another flowchart of a problem model training process according to an embodiment of the present application.
  • FIG. 7 is a schematic structural diagram of an automatic question answering device provided by an embodiment of the present application.
  • FIG. 8 is another schematic structural diagram of an automatic question-answering apparatus according to an embodiment of the present application.
  • FIG. 9 is an exemplary flowchart of an automatic question answering service provided by an embodiment of the present application.
  • FIG. 1 is a flowchart of an automatic question answering method provided by an embodiment of the present application.
  • the method includes:
  • S101 Acquire a user question sentence and a candidate question set.
  • the user question sentence refers to a sentence characterizing a question posed by a user.
  • the user question sentence may be a question sentence entered by a user directly in a text form, or may be a question sentence obtained by converting a question voice uttered by a user into a text form.
  • the candidate question set is a question set for matching user question sentences, and the question sentences in the candidate question set all correspond to preset answer information.
  • S102 Use a preset question model to determine a question sentence that matches the user question sentence from the candidate question set as a target question sentence.
  • the problem model is obtained by training the attention neural network using historical question sentences as training samples.
  • different positions of the words in the question sentence correspond to different moments.
  • Using attention neural networks can learn the weight information of different moments in the question sentence and can effectively increase the weight of important words in the question sentence. Reduce the interference of useless information in question sentences.
  • sentence vectors When generating sentence vectors based on weight information at different times in question sentences, it can make the sentence vectors have the semantic information of the question sentences, thereby improving the interpretability of sentence vectors.
  • the problem model is actually an attention neural network model, which has the characteristics of an attention neural network. Therefore, using the question model to match the user question sentence from the candidate question set can improve the interpretability of the matching result.
  • the question sentence matching the user question sentence in the candidate question set shall have the same semantics as the user question sentence and shall have the same meaning, that is, the target question sentence and the user question sentence are actually the same question .
  • the target question sentence and the user question sentence should have the same semantics, but the expression form can be the same or different.
  • the target question sentence may be a modified sentence of the user question sentence, so that the target question sentence and the user question sentence have the same semantics but different expression forms.
  • the target question sentence matches the user question sentence, it indicates that the two are actually the same question, and the target question sentence corresponds to the preset answer information. Therefore, the answer information corresponding to the target question sentence is actually the user question sentence. Answer information.
  • the automatic question answering method provided in this embodiment first obtains a user question sentence and a candidate question set, and then uses a question model obtained by training the attention neural network by using the question sentence training samples to determine from the candidate question set and the user The question sentence that matches the question sentence is used as the target question sentence. Finally, the answer information corresponding to the target question sentence is used to respond to the user question sentence.
  • the characteristics of the attention neural network are used to make the problem sentence realized by the question model. The process and result of automatic question answering are more interpretable, which improves the effect of automatic question answering.
  • FIG. 2 is another flowchart of an automatic question answering method provided by an embodiment of the present application.
  • the method includes:
  • S202 Use the BM25 algorithm to retrieve at least one question sentence associated with the user question sentence from a preset question library as a candidate question set.
  • the preset question library (FAQ, Frequently Asked Questions) is a question library constructed according to previously collected question sentences.
  • the preset question library may be a question library specific to a certain field, or a multi-purpose question library. Question base for each area.
  • the question statements in the preset question library may include: frequently asked questions in a target domain, and variants of the frequently asked questions.
  • the variant sentence of the frequently asked question refers to a question sentence with the same semantics but a different expression form as the frequently asked question sentence, and it may specifically be obtained by crawling the frequently asked question sentence on the Internet in a retrieval manner.
  • the expression forms of common question sentences can be enriched and the hit rate of user question sentences can be improved.
  • the target field may refer to the field of justice, or the field of finance, computer technology, or other fields.
  • the question sentences in the preset question database all correspond to preset answer information, and the preset answer information may be written by a manual summary, or may be obtained by other methods.
  • BM25 Best Match25
  • BM25 Best Match25
  • BM25 is an algorithm used to evaluate the correlation between search words and documents, and has a high search efficiency and effect.
  • Using the BM25 algorithm at least one question sentence associated with the user question sentence is retrieved from the preset question bank as a candidate question set, instead of directly using the preset question bank as a candidate question set, thereby reducing candidate questions
  • the amount of data in the set improves the processing speed of matching the user question sentence from the candidate question set.
  • S203 Use a preset question model to obtain the similarity between the user question sentence and the question sentence in the candidate question set.
  • the user question sentence and any question sentence in the candidate question set can be converted into corresponding sentence vectors.
  • the user question sentence and the The similarity between any problem sentences in the candidate problem set is described.
  • the problem model is obtained by training an attention neural network using historical problem sentences as training samples.
  • the sentence vectors obtained by using the problem model include the semantic information of the problem sentences. Therefore, the closer the semantic information between the problem sentences is, The higher the similarity of the sentence vectors.
  • S204 Determine, as the target question sentence, a question sentence in which the similarity meets a preset similarity condition in the candidate question set as a question sentence matching the user question sentence.
  • the highest similarity question sentence in the candidate question set may be determined as the target question sentence, or the candidate The question sentence with the highest similarity and greater than a preset similarity threshold in the question set is determined as the target question sentence.
  • the preset similarity condition is: the highest similarity, or the highest similarity and greater than a preset similarity threshold.
  • the automatic question answering method uses the BM25 algorithm to retrieve at least one question sentence associated with the user question sentence from a preset question library as a candidate question set after obtaining the user question sentence, and It does not directly use the entire preset question bank as a candidate question set, thereby reducing the amount of data in the candidate question set, improving the processing speed of matching user question sentences from the candidate question set using the question model, and reducing the response time of the entire automatic question answering process , Can fully meet the needs of high concurrency and rapid response in the field of automatic question answering.
  • FIG. 3 is a flowchart of a problem model training process according to an embodiment of the present application.
  • the problem model training process provided in this embodiment refers to the process of training the attention neural network by using the problem sentence training samples.
  • the problem model training process includes:
  • the training sample includes a sample question sentence, and the sample question sentence is a question sentence as a training sample.
  • the sample question sentence may be a question sentence in a preset question library (FAQ), or a question sentence obtained through other channels.
  • FQ preset question library
  • the process of obtaining training samples may include:
  • the first question sentence may be any question sentence in the preset question library.
  • the second question sentence is used as a positive sample of the first question sentence
  • the third question sentence is used as a negative sample of the first question sentence; wherein the positive sample and the negative sample are Training samples.
  • the second question sentence and the third question sentence may also be question sentences in the preset question library; the number of the second question sentences may be one or more; similarly, the first question sentence The number of three question sentences can also be one or more.
  • step a1 to a3 for each question sentence in the preset question library, question sentences with the same semantics can be used as positive samples, and question sentences with different semantics can be used as negative samples.
  • the unsupervised method may include at least one of a TF-IDF (term frequency-inverse document frequency) vector space model, a BM25 and a WMD (word shift's distance) algorithm.
  • TF-IDF term frequency-inverse document frequency
  • WMD word shift's distance
  • S302 Use the sample question sentence to train an attention neural network to obtain the question model.
  • tensorflow can be used to train the attention neural network.
  • the training goal is to make semantically similar problem sentences have similar sentence vectors (same semantic problem sentences should have the same sentence vector ), Similarity can be measured by cosine similarity.
  • Theano, Keras, Torch, etc. can also be used to train the attention neural network.
  • TensorFlow, Theano, Keras, and Torch are all machine learning frameworks.
  • the attention neural network is trained by using the sample problem sentences, so that problem sentences with similar semantics can have similar sentence vectors, so that the machine learning process and learning The results have strong interpretability, and interpretable problem models are obtained.
  • FIG. 4 is another flowchart of a problem model training process provided by an embodiment of the present application.
  • the problem model training process includes:
  • the training sample includes a sample question sentence, and the sample question sentence is a question sentence as a training sample.
  • each word in the sample question sentence needs to be converted into a word vector form as input data of the attention neural network.
  • a recurrent neural network with a two-way gate structure is used to extract a feature vector corresponding to each moment in the sample question sentence.
  • the words in the sentence can be regarded as a time series, and each word in the sentence corresponds to a moment.
  • a recurrent neural network with a two-way gate structure may include a long short-term memory network (LSTM, Long Short-Term Memory), a gated recurrent unit (GRU, Gated Recurrent Unit), and the like.
  • LSTM long short-term memory network
  • GRU gated recurrent unit
  • the attention information can be used to learn the weight information at different times in the sentence.
  • S405 Determine a sentence vector corresponding to the sample question sentence according to the feature vector and weight information corresponding to each moment in the sample question sentence.
  • the sentence vector corresponding to the sample question sentence can be obtained; of course, other calculation methods can also be used to determine the sample
  • the sentence vector corresponding to the question sentence is not repeated here.
  • steps S403-S405 can be as shown in FIG. 5.
  • the word vector corresponding to each word in any sample question sentence is subjected to feature extraction of the two-way gate structure, and the weights w1 to wn corresponding to each word vector are determined.
  • a sentence vector of the sample question sentence is determined according to the weights w1 to wn and each of the extracted feature vectors.
  • S406 Determine network parameters of the attention neural network according to the sentence vector, and obtain the problem model.
  • the sentence vector encoded by the recurrent neural network and the attention mechanism using the bidirectional gate structure fully contains the semantic information of the sample question sentence, and reduces the interference of useless information.
  • Sample question sentences with the same semantics should have the same or similar sentence vectors, so the network parameters of the attention neural network can be continuously adjusted according to the similarity between the sentence vectors of each sample question sentence, until a set of realizable Sample problem sentences with the same semantics have the same network parameters as the target of the same or similar sentence vectors, and then a trained problem model is obtained.
  • a recurrent neural network with a two-way gate structure is used to extract feature vectors corresponding to each moment in the sample question sentence; and using an attention mechanism Determine the weight information corresponding to each moment in the sample question sentence; then determine the sentence vector corresponding to the sample question sentence according to the feature vector and weight information corresponding to each moment in the sample question sentence; finally, determine according to the sentence vector
  • the network parameters of the attention neural network are used to obtain the problem model.
  • the semantic information of the sample problem sentence is fully mined, which further improves the interpretability of the machine learning process and the learning result and obtains an interpretable More sexual problem models.
  • FIG. 6 is another flowchart of a problem model training process according to an embodiment of the present application.
  • the problem model training process includes:
  • the target field is the application field of the automatic question answering device.
  • a large number of text sentences in the judicial field can be collected as a text sentence database in the judicial field.
  • S502 Perform word segmentation processing on the text sentence to obtain a word segmentation result of the text sentence.
  • Word segmentation processing can be performed using open source word segmentation tool software. For example, using Harbin Institute of Technology's Language Technology Platform (LTP, Language, Technology, Plantform) to segment words in the text sentence library to obtain segmentation results.
  • LTP Harbin Institute of Technology's Language Technology Platform
  • the word segmentation result of the text sentence includes each word in the text sentence.
  • S503 Perform word vector training on the word segmentation result of the text sentence to obtain a word vector model.
  • a word vector training tool may be used for training to generate a corresponding word vector model.
  • Word vector training tools can use Word2vec, or other word vector training tools.
  • S504 Obtain a training sample, where the training sample includes a sample question sentence.
  • S505 Perform word segmentation processing on the sample question sentence to obtain a word segmentation result of the sample question sentence.
  • the word segmentation result of the sample question sentence includes each word in the sample question sentence.
  • S506 Obtain a word vector corresponding to each word in the sample question sentence according to the word vector model.
  • a word vector corresponding to a word can be obtained by using a word vector model. Therefore, for each word in a sample question sentence, a word vector corresponding to each word in the sample question sentence can be obtained from the word vector model.
  • a random word vector (such as the word vector of UNK) is specified for these out-of-set words as the word vector corresponding to the out-of-set words.
  • a recurrent neural network with a two-way gate structure is used to extract a feature vector corresponding to each moment in the sample question sentence.
  • S509 Determine a sentence vector corresponding to the sample question sentence according to the feature vector and weight information corresponding to each moment in the sample question sentence;
  • S510 Determine network parameters of the attention neural network according to the sentence vector, and obtain the problem model.
  • the problem model training process obtained in this embodiment obtains a text sentence in a target field and performs word segmentation processing on the text sentence, and then performs word vector training on the word segmentation result of the text sentence to obtain a word vector model for improving the The efficiency of obtaining the word vector corresponding to each word in the sample problem sentence improves the efficiency of the problem model training.
  • an embodiment of the present invention further provides a corresponding automatic question answering device.
  • the technical content of the automatic question answering device described below can be cross-referenced with the technical content of the automatic question answering method described above.
  • FIG. 7 is a schematic structural diagram of an automatic question-answering device provided by an embodiment of the present application.
  • the automatic question answering device of this embodiment is configured to implement the automatic question answering method of the foregoing embodiment. As shown in FIG. 7, the device includes:
  • the question obtaining unit 100 is configured to obtain a user question sentence and a candidate question set, where the question sentences in the candidate question set each correspond to preset answer information.
  • the question determining unit 200 is configured to determine a question sentence that matches the user question sentence from the candidate question set by using a preset question model as a target question sentence, wherein the question model is a history question
  • the sentence is used as a training sample to train the attention neural network;
  • the question response unit 300 is configured to use the answer information corresponding to the target question sentence to respond to the user question sentence.
  • the problem obtaining unit 100 is specifically configured to:
  • At least one question sentence associated with the user question sentence is retrieved from a preset question bank as a candidate question set; the question sentences in the preset question bank all correspond to preset answer information.
  • the problem determination unit 200 is specifically configured to:
  • a question sentence in which the similarity meets a preset similarity condition in the candidate question set is determined as a question sentence that matches the user question sentence as a target question sentence.
  • the automatic question answering device provided in this embodiment first obtains a user question sentence and a candidate question set, and then uses a question model obtained by training the attention sentence neural network through the question sentence training samples to determine from the candidate question set and the user The question sentence that matches the question sentence is used as the target question sentence. Finally, the answer information corresponding to the target question sentence is used to respond to the user question sentence.
  • the characteristics of the attention neural network are used to make the problem sentence realized by the question model. The process and result of automatic question answering are more interpretable, which improves the effect of automatic question answering.
  • FIG. 8 is another schematic structural diagram of an automatic question answering device provided by an embodiment of the present application.
  • the automatic question-answering device of this embodiment further includes a model training unit 400 and a question base unit 500.
  • the model training unit 400 is configured to obtain a training sample, where the training sample includes a sample question sentence; and training the attention neural network by using the sample question sentence to obtain the question model.
  • the question library unit 500 is configured to collect historical question statements, configure the question library according to the collected historical question statements, and obtain a preset question library.
  • the process by which the model training unit 400 obtains training samples includes:
  • the third question sentence as a negative sample of the first question sentence; wherein the positive sample and the negative sample are training samples .
  • the model training unit 400 uses the sample question sentence to train an attention neural network, and the process of obtaining the question model includes:
  • a recurrent neural network with a two-way gate structure is used to extract a feature vector corresponding to each moment in the sample question sentence;
  • model training unit 400 uses the sample question sentence to train an attention neural network, it is further configured to:
  • the word vector training is performed on the word segmentation result of the text sentence to obtain a word vector model.
  • the process by which the model training unit 400 obtains a word vector corresponding to each word in the sample question sentence includes:
  • a word vector corresponding to each word in the sample question sentence is obtained.
  • the automatic question answering device uses a recurrent neural network with a bidirectional gate structure according to the word vector corresponding to each word in the sample question sentence to extract the feature vector corresponding to each moment in the sample question sentence; and uses the attention mechanism to determine Weight information corresponding to each moment in the sample question sentence; then determining a sentence vector corresponding to the sample question sentence according to the feature vector and weight information corresponding to each moment in the sample question sentence, and finally determining the sentence vector according to the sentence vector.
  • the network parameters of the attention neural network are described, and the problem model is obtained.
  • the semantic information of the sample problem sentence is fully mined, which further improves the interpretability of the machine learning process and the learning results and obtains the interpretability. Stronger problem model.
  • FIG. 9 is an exemplary flowchart of an automatic question and answer business process provided by an embodiment of the present application.
  • the automatic question and answer business process in this embodiment includes:
  • the user query problem is obtained, and the user query problem is preprocessed and the user intention is recognized to obtain the user question sentence.
  • the preprocessing may include personalized word segmentation and keyword extraction.
  • At least one question sentence associated with the user question sentence is initially retrieved from the knowledge base (that is, a preset question library) as a candidate question set.
  • the way to obtain problem sentences in the knowledge base can include at least one of network intelligent crawling, user data mining, and user-defined.
  • An automatic question answering device provided by an embodiment of the present invention includes a processor and a memory.
  • the above-mentioned problem obtaining unit 100, problem determination unit 200, question response unit 300, model training unit 400, and question library unit 500 are all stored in the memory as program units.
  • the processor executes the above program unit stored in the memory to achieve the corresponding function.
  • the processor contains a kernel, and the kernel retrieves the corresponding program unit from the memory.
  • the kernel can set one or more, and adjust the kernel parameters to solve the technical problem that the current automatic question answering scheme lacks interpretability.
  • Memory may include non-permanent memory, random access memory (RAM), and / or non-volatile memory in computer-readable media, such as read-only memory (ROM) or flash memory (RAM).
  • Memory includes at least one Memory chip.
  • An embodiment of the present invention provides a storage medium on which a program is stored, and the program is implemented by a processor to implement the automatic question answering method.
  • An embodiment of the present invention provides a processor, where the processor is configured to run a program, and the automatic question answering method is executed when the program is run.
  • An embodiment of the present invention provides a device.
  • the device includes a processor, a memory, and a program stored on the memory and executable on the processor.
  • the processor executes the program, the following steps are implemented:
  • a question sentence matching the user question sentence is determined from the candidate question set as a target question sentence; wherein the question model is a history question sentence as a training sample to focus attention Neural network training;
  • the acquiring the user question sentence and the candidate question set includes:
  • At least one question sentence associated with the user question sentence is retrieved from a preset question bank as a candidate question set; the question sentences in the preset question bank all correspond to preset answer information.
  • the use of a preset question model to determine a question sentence that matches the user question sentence from the candidate question set, and the target question sentence includes:
  • a question sentence in which the similarity meets a preset similarity condition in the candidate question set is determined as a question sentence that matches the user question sentence as a target question sentence.
  • the training process of the problem model includes:
  • training samples including sample question sentences
  • the sample problem sentence is used to train an attention neural network to obtain the problem model.
  • the obtaining training samples includes:
  • the third question sentence as a negative sample of the first question sentence; wherein the positive sample and the negative sample are training samples .
  • training the attention neural network by using the sample question sentence to obtain the question model includes:
  • a recurrent neural network with a two-way gate structure is used to extract a feature vector corresponding to each moment in the sample question sentence;
  • the method further includes:
  • the obtaining a word vector corresponding to each word in the sample question sentence includes:
  • a word vector corresponding to each word in the sample question sentence is obtained.
  • the equipment in this article can be server, PC, PAD, mobile phone, etc.
  • This application also provides a computer program product, which when executed on a data processing device, is suitable for executing a program having the following method steps for initialization:
  • a question sentence matching the user question sentence is determined from the candidate question set as a target question sentence; wherein the question model is a history question sentence as a training sample for attention Neural network training;
  • the acquiring the user question sentence and the candidate question set includes:
  • At least one question sentence associated with the user question sentence is retrieved from a preset question bank as a candidate question set; the question sentences in the preset question bank all correspond to preset answer information.
  • the use of a preset question model to determine a question sentence that matches the user question sentence from the candidate question set, and the target question sentence includes:
  • a question sentence in which the similarity meets a preset similarity condition in the candidate question set is determined as a question sentence that matches the user question sentence as a target question sentence.
  • the training process of the problem model includes:
  • training samples including sample question sentences
  • the sample problem sentence is used to train an attention neural network to obtain the problem model.
  • the obtaining training samples includes:
  • the third question sentence as a negative sample of the first question sentence; wherein the positive sample and the negative sample are training samples .
  • training the attention neural network by using the sample question sentence to obtain the question model includes:
  • a recurrent neural network with a two-way gate structure is used to extract a feature vector corresponding to each moment in the sample question sentence;
  • the method further includes:
  • the obtaining a word vector corresponding to each word in the sample question sentence includes:
  • a word vector corresponding to each word in the sample question sentence is obtained.
  • this application may be provided as a method, a system, or a computer program product. Therefore, this application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application may take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
  • computer-usable storage media including, but not limited to, disk storage, CD-ROM, optical storage, etc.
  • These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to work in a specific manner such that the instructions stored in the computer-readable memory produce a manufactured article including an instruction device, the instructions
  • the device implements the functions specified in one or more flowcharts and / or one or more blocks of the block diagram.
  • These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of steps can be performed on the computer or other programmable device to produce a computer-implemented process, which can be executed on the computer or other programmable device.
  • the instructions provide steps for implementing the functions specified in one or more flowcharts and / or one or more blocks of the block diagrams.
  • a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.
  • processors CPUs
  • input / output interfaces output interfaces
  • network interfaces network interfaces
  • memory volatile and non-volatile memory
  • the memory may include non-permanent memory, random access memory (RAM), and / or non-volatile memory in computer-readable media, such as read-only memory (ROM) or flash memory (flash RAM).
  • RAM random access memory
  • ROM read-only memory
  • flash RAM flash memory
  • Computer-readable media includes permanent and non-persistent, removable and non-removable media.
  • Information storage can be accomplished by any method or technology.
  • Information may be computer-readable instructions, data structures, modules of a program, or other data.
  • Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), and read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, read-only disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, Magnetic tape cartridges, magnetic tape storage or other magnetic storage devices or any other non-transmitting medium may be used to store information that can be accessed by a computing device.
  • computer-readable media does not include temporary computer-readable media, such as modulated data signals and carrier waves.
  • this application may be provided as a method, a system, or a computer program product. Therefore, this application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application may take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
  • computer-usable storage media including, but not limited to, disk storage, CD-ROM, optical storage, etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

本发明提供的自动问答方法及装置，先获取用户问题语句与候选问题集，再利用通过问题语句训练样本对注意力神经网络进行训练得到的问题模型，从所述候选问题集中确定出与所述用户问题语句相匹配的问题语句，作为目标问题语句，最后采用所述目标问题语句对应的答案信息，对所述用户问题语句进行回应，利用注意力神经网络的特性，使利用所述问题模型实现的自动问答过程及结果更具有可解释性，提高了自动问答效果。

Description

自动问答方法及装置
本申请要求于2018年5月23日提交中国专利局、申请号为201810502726.3、发明名称为“自动问答方法及装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本发明涉及信息处理领域，更具体的说，涉及自动问答方法及装置。
背景技术
自动问答技术通过智能助手或者智能客服的形式，自动回答用户在办理相关事务时遇到的问题，大幅减少了人工客服的工作量，同时也让用户享受到更快捷的问答服务。
传统的自动问答方案，将自动问答过程转化成用户问题在常见问题库中的检索过程，具体会利用自然语言处理工具对用户问题语句进行分析，识别出用户问题语句的关键词，并根据获得的关键词从常见问题库中检索出相匹配问题及其答案。然而，在传统自动问答方案中所采用的机器学习方法通常是基于词袋模型的，其中忽略了问题语句中各个词语之间的先后顺序关系，丢失了问题语句中部分语义信息，使得机器学习过程与机器学习结果缺乏可解释性，同样也导致自动问答过程及结果缺乏可解释性。
因此，目前迫切需要一种切实有效的技术方案，以提高自动问答过程及结果的可解释性。
发明内容
鉴于上述问题，提出了本发明以便提供一种克服上述问题或者至少部分地解决上述问题的自动问答方法及装置。
为实现上述目的，本发明提供如下技术方案：
一种自动问答方法，包括：
获取用户问题语句与候选问题集，其中，所述候选问题集中的问题语句均对应有预设的答案信息；
利用预置的问题模型，从所述候选问题集中确定出与所述用户问题语句相 匹配的问题语句，作为目标问题语句；其中，所述问题模型为，将历史问题语句作为训练样本对注意力神经网络进行训练得到的；
采用所述目标问题语句对应的答案信息，对所述用户问题语句进行回应。
优选的，所述获取用户问题语句与候选问题集包括：
获取用户问题语句；
采用BM25算法，从预置问题库中检索出与所述用户问题语句相关联的至少一条问题语句，作为候选问题集；所述预置问题库中的问题语句均对应有预设的答案信息。
优选的，所述利用预置的问题模型，从所述候选问题集中确定出与所述用户问题语句相匹配的问题语句，作为目标问题语句包括：
利用预置的问题模型，获取所述用户问题语句与所述候选问题集中的问题语句的相似度；
将所述候选问题集中所述相似度符合预设相似度条件的问题语句，确定为与所述用户问题语句相匹配的问题语句，作为目标问题语句。
优选的，所述问题模型的训练过程包括：
获取训练样本，所述训练样本中包括样本问题语句；
采用所述样本问题语句对注意力神经网络进行训练，得到所述问题模型。
优选的，所述获取训练样本包括：
获取第一问题语句；
获取与所述第一问题语句语义相同的第二问题语句，以及，与所述第一问题语句语义不同的第三问题语句；
将所述第二问题语句作为所述第一问题语句的正样本，将所述第三问题语句作为所述第一问题语句的负样本；其中，所述正样本与所述负样本为训练样本。
优选的，所述采用所述样本问题语句对注意力神经网络进行训练，得到所述问题模型包括：
获取所述样本问题语句中各个词对应的词向量；
根据所述词向量，利用双向门结构的循环神经网络，抽取所述样本问题语句中各个时刻对应的特征向量；
利用注意力机制，确定所述样本问题语句中各个时刻对应的权重信息；
根据所述样本问题语句中各个时刻对应的特征向量与权重信息，确定所述样本问题语句对应的句向量；
根据所述句向量，确定所述注意力神经网络的网络参数，得到所述问题模型。
优选的，在所述采用所述样本问题语句对注意力神经网络进行训练之前，所述方法还包括：
获取目标领域的文本语句；
对所述文本语句进行分词处理，得到所述文本语句的分词结果；
对所述文本语句的分词结果进行词向量训练，得到词向量模型；
相应的，
所述获取所述样本问题语句中各个词对应的词向量包括：
对所述样本问题语句进行分词处理，得到所述样本问题语句的分词结果；
根据所述词向量模型，获取所述样本问题语句中各个词对应的词向量。
一种自动问答装置，包括：
问题获取单元，用于获取用户问题语句与候选问题集，其中，所述候选问题集中的问题语句均对应有预设的答案信息；
问题确定单元，用于利用预置的问题模型，从所述候选问题集中确定出与所述用户问题语句相匹配的问题语句，作为目标问题语句；其中，所述问题模型为，将历史问题语句作为训练样本对注意力神经网络进行训练得到的；
问题回应单元，用于采用所述目标问题语句对应的答案信息，对所述用户问题语句进行回应。
一种存储介质，所述存储介质包括存储的程序，其中，在所述程序运行时控制所述存储介质所在设备执行上述的自动问答方法。
一种处理器，所述处理器用于运行程序，其中，所述程序运行时执行上述的自动问答方法。
借由上述技术方案，本发明提供的自动问答方法及装置，先获取用户问题语句与候选问题集，再利用通过问题语句训练样本对注意力神经网络进行训练得到的问题模型，从所述候选问题集中确定出与所述用户问题语句相匹配 的问题语句，作为目标问题语句，最后采用所述目标问题语句对应的答案信息，对所述用户问题语句进行回应，利用注意力神经网络的特性，使利用所述问题模型实现的自动问答过程及结果更具有可解释性，提高了自动问答效果。
上述说明仅是本发明技术方案的概述，为了能够更清楚了解本发明的技术手段，而可依照说明书的内容予以实施，并且为了让本发明的上述和其它目的、特征和优点能够更明显易懂，以下特举本发明的具体实施方式。
附图说明
通过阅读下文优选实施方式的详细描述，各种其他的优点和益处对于本领域普通技术人员将变得清楚明了。附图仅用于示出优选实施方式的目的，而并不认为是对本发明的限制。而且在整个附图中，用相同的参考符号表示相同的部件。在附图中：
图1为本申请实施例提供的自动问答方法的一种流程图；
图2为本申请实施例提供的自动问答方法的另一种流程图；
图3为本申请实施例提供的问题模型训练过程的一种流程图；
图4为本申请实施例提供的问题模型训练过程的另一种流程图；
图5为本申请实施例提供的语句编码示意图；
图6为本申请实施例提供的问题模型训练过程的又一种流程图；
图7为本申请实施例提供的自动问答装置的一种结构示意图；
图8为本申请实施例提供的自动问答装置的另一种结构示意图；
图9为本申请实施例提供的自动问答业务流程示例图。
具体实施方式
下面将参照附图更详细地描述本公开的示例性实施例。虽然附图中显示了本公开的示例性实施例，然而应当理解，可以以各种形式实现本公开而不应被这里阐述的实施例所限制。相反，提供这些实施例是为了能够更透彻地理解本公开，并且能够将本公开的范围完整的传达给本领域的技术人员。
请参阅图1，图1为本申请实施例提供的自动问答方法的一种流程图。
如图1所示，所述方法包括：
S101：获取用户问题语句与候选问题集。
用户问题语句是指表征用户提出的问题的语句，用户问题语句可以是用户直接输入的文本形式的问题语句，也可以是将用户说出的问题语音转换成文本形式后得到的问题语句。
候选问题集是用于匹配用户问题语句的一个问题集合，候选问题集中的问题语句均对应有预设的答案信息。
S102：利用预置的问题模型，从所述候选问题集中确定出与所述用户问题语句相匹配的问题语句，作为目标问题语句。
其中，所述问题模型为，将历史问题语句作为训练样本对注意力神经网络进行训练得到的。基于时序来对问题语句进行编码时，问题语句中词的不同位置对应不同时刻，采用注意力神经网络可以学习到问题语句中不同时刻的权重信息，并可以有效提高问题语句中重要词的权重，降低问题语句中无用信息的干扰，在依据问题语句中不同时刻的权重信息生成句向量时，能够使句向量具备问题语句的语义信息，从而提高句向量的可解释性。
所述问题模型实际上是一种注意力神经网络模型，其具备注意力神经网络的特性。所以，利用所述问题模型从所述候选问题集中去匹配所述用户问题语句，能够提高匹配结果的可解释性。
候选问题集中与所述用户问题语句相匹配的问题语句，应与所述用户问题语句的语义相同，应表达同一含义，也即，所述目标问题语句与所述用户问题语句实际上为同一问题。所述目标问题语句与所述用户问题语句的语义应相同，但表达形式可以相同，也可以不同。例如，所述目标问题语句可以是所述用户问题语句的变形语句，这样所述目标问题语句与所述用户问题语句的语义相同，但表达形式不同。
S103：采用所述目标问题语句对应的答案信息，对所述用户问题语句进行回应。
由于目标问题语句与用户问题语句相匹配，表明两者实际上是同一问题，而目标问题语句又对应有预设的答案信息，所以，目标问题语句对应的答案信息，实际上就是用户问题语句的答案信息。
在确定出用户问题语句的答案信息后，采用所述答案信息对用户问题语句进行回应，便实现了自动问答过程。
本实施例提供的自动问答方法，先获取用户问题语句与候选问题集，再利用通过问题语句训练样本对注意力神经网络进行训练得到的问题模型，从所述候选问题集中确定出与所述用户问题语句相匹配的问题语句，作为目标问题语句，最后采用所述目标问题语句对应的答案信息，对所述用户问题语句进行回应，利用注意力神经网络的特性，使利用所述问题模型实现的自动问答过程及结果更具有可解释性，提高了自动问答效果。
请参阅图2，图2为本申请实施例提供的自动问答方法的另一种流程图。
如图2所示，所述方法包括：
S201：获取用户问题语句。
S202：采用BM25算法，从预置问题库中检索出与所述用户问题语句相关联的至少一条问题语句，作为候选问题集。
预置问题库(FAQ，Frequently Asked Questions)是根据预先收集到的问题语句构建得到的问题库，具体地，所述预置问题库可以具体为针对某一领域的问题库，也可以是针对多个领域的问题库。
一示例中，所述预置问题库中的问题语句可包括：目标领域的常见问题语句，以及，所述常见问题语句的变形语句。
所述常见问题语句的变形语句，是指与所述常见问题语句语义相同但表达形式不同的问题语句，其具体可以是采用检索的方式在互联网中爬取所述常见问题语句而得到的。通过常见问题语句的变形语句，可以丰富常见问题语句的表达形式，提高用户问题语句的命中率。
一示例中，所述目标领域可以是指司法领域，也可以是指金融领域、计算机技术领域或其他领域。
所述预置问题库中的问题语句均对应有预设的答案信息，所述预设的答案信息可以是人工总结编写的，也可以是采取其他方式获取到的。
BM25(Best Match25)算法是一种用来评价搜索词和文档之间相关性的算法，具有较高的搜索效率与效果。采用BM25算法，从预置问题库中检索出与所述用户问题语句相关联的至少一条问题语句，作为候选问题集，而并 不是将预置问题库直接作为候选问题集，从而减少了候选问题集的数据量，提高了从所述候选问题集中匹配所述用户问题语句的处理速度。
S203：利用预置的问题模型，获取所述用户问题语句与所述候选问题集中的问题语句的相似度。
利用所述问题模型，可以将所述用户问题语句与所述候选问题集中任一问题语句均转换成相应的句向量，通过计算句向量之间的相似度，可以得到所述用户问题语句与所述候选问题集中任一问题语句之间的相似度。
所述问题模型是将历史问题语句作为训练样本对注意力神经网络进行训练得到的，利用所述问题模型得到的句向量包含了问题语句的语义信息，所以，语义信息越相近的问题语句之间的句向量的相似度就越高。
S204：将所述候选问题集中所述相似度符合预设相似度条件的问题语句，确定为与所述用户问题语句相匹配的问题语句，作为目标问题语句。
在获取到所述用户问题语句与所述候选问题集中每个问题语句的相似度后，可以将所述候选问题集中所述相似度最高的问题语句确定为目标问题语句，也可以将所述候选问题集中所述相似度最高且大于预设相似度阈值的问题语句确定为目标问题语句。
相应的，所述预设相似度条件为：相似度最高，或，相似度最高且大于预设相似度阈值。
S205：采用所述目标问题语句对应的答案信息，对所述用户问题语句进行回应。
本实施例提供的自动问答方法，在获取到用户问题语句后，先采用BM25算法，从预置问题库中检索出与所述用户问题语句相关联的至少一条问题语句，作为候选问题集，而并不是直接将预置问题库整个作为候选问题集，从而降低了候选问题集的数据量，提高了利用问题模型从候选问题集中匹配用户问题语句的处理速度，减少了整个自动问答过程的响应时间，能够充分满足自动问答领域的高并发以及快速响应的需求。
请参阅图3，图3为本申请实施例提供的问题模型训练过程的一种流程图。
本实施例提供的问题模型训练过程，是指采用问题语句训练样本对注意力神经网络进行训练的过程。
如图3所示，所述问题模型训练过程包括：
S301：获取训练样本。
其中，所述训练样本中包括样本问题语句，样本问题语句就是作为训练样本的问题语句。
一示例中，所述样本问题语句可以是预置问题库(FAQ)中的问题语句，也可以是通过其他渠道获取到的问题语句。
一示例中，所述获取训练样本的过程可包括：
a1、获取第一问题语句；
其中，所述第一问题语句可以是所述预置问题库中的任一问题语句。
a2、获取与所述第一问题语句语义相同的第二问题语句，以及，与所述第一问题语句语义不同的第三问题语句；
a3、将所述第二问题语句作为所述第一问题语句的正样本，将所述第三问题语句作为所述第一问题语句的负样本；其中，所述正样本与所述负样本为训练样本。
其中，所述第二问题语句与所述第三问题语句也可以是所述预置问题库中的问题语句；所述第二问题语句的数量可以是一个或多个；同样的，所述第三问题语句的数量也可以是一个或多个。
利用步骤a1～a3、可以针对预置问题库中的每个问题语句，将语义相同的问题语句作为正样本，将语义不同的问题语句作为负样本。
一示例中，还可以通过一些无监督的方法来增加容易混淆的语义不同的问题语句作为负样本。所述无监督的方法可以包括：TF-IDF(term frequency–inverse document frequency，词频-逆向文件频率)的向量空间模型、BM25与WMD(word mover’s distance，词移距离)算法中的至少一种。
S302：采用所述样本问题语句对注意力神经网络进行训练，得到所述问题模型。
一示例中，根据获取到的训练样本，可使用tensorflow来对注意力神经网络进行训练，训练的目标是使语义相近的问题语句具有相近的句向量(语义相同的问题语句应具有相同的句向量)，相似度可通过余弦相似度来度量。
其他示例中，还可以使用Theano、Keras、Torch等来对注意力神经网络 进行训练。TensorFlow、Theano、Keras、Torch均是机器学习框架。
本实施例提供的问题模型训练过程，在获取训练样本后，采用所述样本问题语句对注意力神经网络进行训练，使语义相近的问题语句可具有相近的句向量，从而使机器学习过程与学习结果均具有较强的可解释性，得到具备可解释性的问题模型。
请参阅图4，图4为本申请实施例提供的问题模型训练过程的另一种流程图。
如图4所示，所述问题模型训练过程包括：
S401：获取训练样本。
其中，所述训练样本中包括样本问题语句，样本问题语句就是作为训练样本的问题语句。
S402：获取所述样本问题语句中各个词对应的词向量。
在训练注意力神经网络时，需要将样本问题语句中各个词先转化为词向量的形式，来作为注意力神经网络的输入数据。
S403：根据所述词向量，利用双向门结构的循环神经网络，抽取所述样本问题语句中各个时刻对应的特征向量。
在基于时序对语句进行分析时，可以语句中的词看做是一个时间序列，语句中的每个词均对应一个时刻。
利用双向门结构的循环神经网络，可以充分结合词向量的上下文关系来对当前时刻的信息进行编码，从而抽取出具有更多语义信息的特征向量。双向门结构的循环神经网络可包括长短期记忆网络(LSTM，Long Short-Term Memory)、门控循环单元(GRU，Gated Recurrent Unit)等。
S404：利用注意力机制，确定所述样本问题语句中各个时刻对应的权重信息。
在一条语句中，不同的词具有不同的重要度，所以，根据不同词的重要度，利用注意力机制，可以学习到语句中不同时刻的权重信息。
S405：根据所述样本问题语句中各个时刻对应的特征向量与权重信息，确定所述样本问题语句对应的句向量。
利用所述样本问题语句中各个时刻的权重信息，针对所述各个时刻的特征 向量进行加权汇总，便可以得到所述样本问题语句对应的句向量；当然也可以采取其他计算方式来确定所述样本问题语句对应的句向量，在此不做赘述。
其中，步骤S403-S405的实现过程可如图5所示，针对任一样本问题语句中各个词对应的词向量进行双向门结构的特征抽取，并确定各个词向量所对应时刻的权重w1～wn，最后根据权重w1～wn与抽取出来的各个特征向量，确定出所述样本问题语句的句向量。
S406：根据所述句向量，确定所述注意力神经网络的网络参数，得到所述问题模型。
利用双向门结构的循环神经网络与注意力机制编码得到的句向量，充分包含了样本问题语句的语义信息，降低了无用信息的干扰。语义相同的样本问题语句应具有相同或相近的句向量，所以，可以根据各个样本问题语句的句向量之间的相似度，来不断调整注意力神经网络的网络参数，直至确定出一组可以实现语义相同的样本问题语句具有相同或相近的句向量这一目标的网络参数，进而得到训练好的问题模型。
本实施例提供的问题模型训练过程，根据样本问题语句中各个词对应的词向量，利用双向门结构的循环神经网络，抽取所述样本问题语句中各个时刻对应的特征向量；并利用注意力机制确定所述样本问题语句中各个时刻对应的权重信息；然后根据所述样本问题语句中各个时刻对应的特征向量与权重信息确定所述样本问题语句对应的句向量，最后根据所述句向量，确定所述注意力神经网络的网络参数，得到所述问题模型，在确定句向量过程中，充分挖掘了样本问题语句的语义信息，进一步提高了机器学习过程与学习结果的可解释性，得到可解释性更强的问题模型。
请参阅图6，图6为本申请实施例提供的问题模型训练过程的又一种流程图。
如图6所示，所述问题模型训练过程包括：
S501：获取目标领域的文本语句。
目标领域为自动问答装置的应用领域，为了提高自动问答装置的精准度，需要获取到样本问题语句中各个词对应的准确的词向量。以司法领域为例，为了获得更适用于司法领域的词向量，可以大量收集司法领域的文本语句， 作为司法领域的文本语句库。
S502：对所述文本语句进行分词处理，得到所述文本语句的分词结果。
分词处理可使用开源分词工具软件来进行，如利用哈工大的语言技术平台(LTP，Language Technology Plantform)，对所述文本语句库中的文本语句进行分词，得到分词结果。所述文本语句的分词结果包括所述文本语句中的各个词。
S503：对所述文本语句的分词结果进行词向量训练，得到词向量模型。
对于所述文本语句的分词结果，可以采用词向量训练工具来进行训练，以生成相应的词向量模型。词向量训练工具可采用Word2vec，也可以采取其他词向量训练工具。
S504：获取训练样本，所述训练样本中包括样本问题语句。
S505：对所述样本问题语句进行分词处理，得到所述样本问题语句的分词结果。
样本问题语句的分词结果包括所述样本问题语句中的各个词。
S506：根据所述词向量模型，获取所述样本问题语句中各个词对应的词向量。
利用词向量模型可以获取词对应的词向量，所以，针对样本问题语句中的各个词，可以从所述词向量模型中获取样本问题语句中各个词对应的词向量。
一示例中，当样本问题语句中某一词或某些词，从所述词向量模型中获取不到对应的词向量时，可以将这些词称为“集外词(out of vocabulary)”，并为这些集外词指定一个随机词向量(如UNK的词向量)，作为所述集外词对应的词向量。
S507：根据所述词向量，利用双向门结构的循环神经网络，抽取所述样本问题语句中各个时刻对应的特征向量。
S508：利用注意力机制，确定所述样本问题语句中各个时刻对应的权重信息。
S509：根据所述样本问题语句中各个时刻对应的特征向量与权重信息，确定所述样本问题语句对应的句向量；
S510：根据所述句向量，确定所述注意力神经网络的网络参数，得到所 述问题模型。
本实施例提供的问题模型训练过程，获取目标领域的文本语句并对所述文本语句进行分词处理，再对所述文本语句的分词结果进行词向量训练得到词向量模型，以用于提高所述样本问题语句中各个词对应的词向量的获取效率，进而提高问题模型训练效率。
对应于前述的自动问答方法，本发明实施例还提供了相应的自动问答装置。下文描述的自动问答装置的技术内容，可与上文描述的自动问答方法的技术内容与相互对应参照。
请参阅图7，图7为本申请实施例提供的自动问答装置的一种结构示意图。
本实施例的自动问答装置，用于实施前述实施例的自动问答方法，如图7所示，所述装置包括：
问题获取单元100，用于获取用户问题语句与候选问题集，其中，所述候选问题集中的问题语句均对应有预设的答案信息。
问题确定单元200，用于利用预置的问题模型，从所述候选问题集中确定出与所述用户问题语句相匹配的问题语句，作为目标问题语句；其中，所述问题模型为，将历史问题语句作为训练样本对注意力神经网络进行训练得到的；
问题回应单元300，用于采用所述目标问题语句对应的答案信息，对所述用户问题语句进行回应。
一示例中，所述问题获取单元100具体用于：
获取用户问题语句；
采用BM25算法，从预置问题库中检索出与所述用户问题语句相关联的至少一条问题语句，作为候选问题集；所述预置问题库中的问题语句均对应有预设的答案信息。
一示例中，所述问题确定单元200具体用于：
利用预置的问题模型，获取所述用户问题语句与所述候选问题集中的问题语句的相似度；
将所述候选问题集中所述相似度符合预设相似度条件的问题语句，确定为 与所述用户问题语句相匹配的问题语句，作为目标问题语句。
本实施例提供的自动问答装置，先获取用户问题语句与候选问题集，再利用通过问题语句训练样本对注意力神经网络进行训练得到的问题模型，从所述候选问题集中确定出与所述用户问题语句相匹配的问题语句，作为目标问题语句，最后采用所述目标问题语句对应的答案信息，对所述用户问题语句进行回应，利用注意力神经网络的特性，使利用所述问题模型实现的自动问答过程及结果更具有可解释性，提高了自动问答效果。
请参阅图8，图8为本申请实施例提供的自动问答装置的另一种结构示意图。
如图8所示，本实施例的自动问答装置，除了包括前述实施例中的问题获取单元100、问题确定单元200、问题回应单元300外，还包括：模型训练单元400与问题库单元500。
所述模型训练单元400，用于获取训练样本，所述训练样本中包括样本问题语句；采用所述样本问题语句对注意力神经网络进行训练，得到所述问题模型。
所述问题库单元500，用于收集历史问题语句，根据收集到的历史问题语句配置问题库，得到预置问题库。
一示例中，所述模型训练单元400获取训练样本的过程包括：
获取第一问题语句；
获取与所述第一问题语句语义相同的第二问题语句，以及，与所述第一问题语句语义不同的第三问题语句；
将所述第二问题语句作为所述第一问题语句的正样本，将所述第三问题语句作为所述第一问题语句的负样本；其中，所述正样本与所述负样本为训练样本。
一示例中，所述模型训练单元400采用所述样本问题语句对注意力神经网络进行训练，得到所述问题模型的过程包括：
获取所述样本问题语句中各个词对应的词向量；
根据所述词向量，利用双向门结构的循环神经网络，抽取所述样本问题语句中各个时刻对应的特征向量；
利用注意力机制，确定所述样本问题语句中各个时刻对应的权重信息；
根据所述样本问题语句中各个时刻对应的特征向量与权重信息，确定所述样本问题语句对应的句向量；
根据所述句向量，确定所述注意力神经网络的网络参数，得到所述问题模型。
一示例中，所述模型训练单元400在采用所述样本问题语句对注意力神经网络进行训练之前，还用于：
获取目标领域的文本语句；
对所述文本语句进行分词处理，得到所述文本语句的分词结果；
对所述文本语句的分词结果进行词向量训练，得到词向量模型。
相应的，
所述模型训练单元400获取所述样本问题语句中各个词对应的词向量的过程包括：
对所述样本问题语句进行分词处理，得到所述样本问题语句的分词结果；
根据所述词向量模型，获取所述样本问题语句中各个词对应的词向量。
本实施例提供的自动问答装置，根据样本问题语句中各个词对应的词向量，利用双向门结构的循环神经网络，抽取所述样本问题语句中各个时刻对应的特征向量；并利用注意力机制确定所述样本问题语句中各个时刻对应的权重信息；然后根据所述样本问题语句中各个时刻对应的特征向量与权重信息确定所述样本问题语句对应的句向量，最后根据所述句向量，确定所述注意力神经网络的网络参数，得到所述问题模型，在确定句向量过程中，充分挖掘了样本问题语句的语义信息，进一步提高了机器学习过程与学习结果的可解释性，得到可解释性更强的问题模型。
请参阅图9，图9为本申请实施例提供的自动问答业务流程示例图。
如图9所示，本实施例的自动问答业务流程包括：
首先，获取用户查询问题，并对用户查询问题进行预处理以及用户意图识别，得到用户问题语句；其中，预处理可以包括个性化分词与关键词提取。
然后，从知识库(即预置问题库)中初步检索出与所述用户问题语句相关联的至少一条问题语句，作为候选问题集。知识库中问题语句的获取方式可以 包括网络智能爬取、用户数据挖掘与用户自定义中至少一项。
最后，利用训练好的问题模型，对所述用户问题语句与所述候选问题集进行语义匹配，确定出目标问题语句及其答案信息，并采用所述答案信息对所述用户问题语句进行回应。
本发明实施例提供的自动问答装置，包括处理器和存储器，上述问题获取单元100、问题确定单元200、问题回应单元300、模型训练单元400与问题库单元500等均作为程序单元存储在存储器中，由处理器执行存储在存储器中的上述程序单元来实现相应的功能。
处理器中包含内核，由内核去存储器中调取相应的程序单元。内核可以设置一个或以上，通过调整内核参数来解决目前自动问答方案缺乏可解释性的技术问题。
存储器可能包括计算机可读介质中的非永久性存储器，随机存取存储器(RAM)和/或非易失性内存等形式，如只读存储器(ROM)或闪存(flash RAM)，存储器包括至少一个存储芯片。
本发明实施例提供了一种存储介质，其上存储有程序，该程序被处理器执行时实现所述自动问答方法。
本发明实施例提供了一种处理器，所述处理器用于运行程序，其中，所述程序运行时执行所述自动问答方法。
本发明实施例提供了一种设备，设备包括处理器、存储器及存储在存储器上并可在处理器上运行的程序，处理器执行程序时实现以下步骤：
获取用户问题语句与候选问题集，其中，所述候选问题集中的问题语句均对应有预设的答案信息；
利用预置的问题模型，从所述候选问题集中确定出与所述用户问题语句相匹配的问题语句，作为目标问题语句；其中，所述问题模型为，将历史问题语句作为训练样本对注意力神经网络进行训练得到的；
采用所述目标问题语句对应的答案信息，对所述用户问题语句进行回应。
优选的，所述获取用户问题语句与候选问题集包括：
获取用户问题语句；
采用BM25算法，从预置问题库中检索出与所述用户问题语句相关联的 至少一条问题语句，作为候选问题集；所述预置问题库中的问题语句均对应有预设的答案信息。
优选的，所述利用预置的问题模型，从所述候选问题集中确定出与所述用户问题语句相匹配的问题语句，作为目标问题语句包括：
利用预置的问题模型，获取所述用户问题语句与所述候选问题集中的问题语句的相似度；
将所述候选问题集中所述相似度符合预设相似度条件的问题语句，确定为与所述用户问题语句相匹配的问题语句，作为目标问题语句。
优选的，所述问题模型的训练过程包括：
获取训练样本，所述训练样本中包括样本问题语句；
采用所述样本问题语句对注意力神经网络进行训练，得到所述问题模型。
优选的，所述获取训练样本包括：
获取第一问题语句；
获取与所述第一问题语句语义相同的第二问题语句，以及，与所述第一问题语句语义不同的第三问题语句；
将所述第二问题语句作为所述第一问题语句的正样本，将所述第三问题语句作为所述第一问题语句的负样本；其中，所述正样本与所述负样本为训练样本。
优选的，所述采用所述样本问题语句对注意力神经网络进行训练，得到所述问题模型包括：
获取所述样本问题语句中各个词对应的词向量；
根据所述词向量，利用双向门结构的循环神经网络，抽取所述样本问题语句中各个时刻对应的特征向量；
利用注意力机制，确定所述样本问题语句中各个时刻对应的权重信息；
根据所述样本问题语句中各个时刻对应的特征向量与权重信息，确定所述样本问题语句对应的句向量；
根据所述句向量，确定所述注意力神经网络的网络参数，得到所述问题模型。
优选的，在所述采用所述样本问题语句对注意力神经网络进行训练之前， 所述方法还包括：
获取目标领域的文本语句；
对所述文本语句进行分词处理，得到所述文本语句的分词结果；
对所述文本语句的分词结果进行词向量训练，得到词向量模型；
相应的，
所述获取所述样本问题语句中各个词对应的词向量包括：
对所述样本问题语句进行分词处理，得到所述样本问题语句的分词结果；
根据所述词向量模型，获取所述样本问题语句中各个词对应的词向量。
本文中的设备可以是服务器、PC、PAD、手机等。
本申请还提供了一种计算机程序产品，当在数据处理设备上执行时，适于执行初始化有如下方法步骤的程序：
获取用户问题语句与候选问题集，其中，所述候选问题集中的问题语句均对应有预设的答案信息；
利用预置的问题模型，从所述候选问题集中确定出与所述用户问题语句相匹配的问题语句，作为目标问题语句；其中，所述问题模型为，将历史问题语句作为训练样本对注意力神经网络进行训练得到的；
采用所述目标问题语句对应的答案信息，对所述用户问题语句进行回应。
优选的，所述获取用户问题语句与候选问题集包括：
获取用户问题语句；
采用BM25算法，从预置问题库中检索出与所述用户问题语句相关联的至少一条问题语句，作为候选问题集；所述预置问题库中的问题语句均对应有预设的答案信息。
优选的，所述利用预置的问题模型，从所述候选问题集中确定出与所述用户问题语句相匹配的问题语句，作为目标问题语句包括：
利用预置的问题模型，获取所述用户问题语句与所述候选问题集中的问题语句的相似度；
将所述候选问题集中所述相似度符合预设相似度条件的问题语句，确定为与所述用户问题语句相匹配的问题语句，作为目标问题语句。
优选的，所述问题模型的训练过程包括：
获取训练样本，所述训练样本中包括样本问题语句；
采用所述样本问题语句对注意力神经网络进行训练，得到所述问题模型。
优选的，所述获取训练样本包括：
获取第一问题语句；
获取与所述第一问题语句语义相同的第二问题语句，以及，与所述第一问题语句语义不同的第三问题语句；
将所述第二问题语句作为所述第一问题语句的正样本，将所述第三问题语句作为所述第一问题语句的负样本；其中，所述正样本与所述负样本为训练样本。
优选的，所述采用所述样本问题语句对注意力神经网络进行训练，得到所述问题模型包括：
获取所述样本问题语句中各个词对应的词向量；
根据所述词向量，利用双向门结构的循环神经网络，抽取所述样本问题语句中各个时刻对应的特征向量；
利用注意力机制，确定所述样本问题语句中各个时刻对应的权重信息；
根据所述样本问题语句中各个时刻对应的特征向量与权重信息，确定所述样本问题语句对应的句向量；
根据所述句向量，确定所述注意力神经网络的网络参数，得到所述问题模型。
优选的，在所述采用所述样本问题语句对注意力神经网络进行训练之前，所述方法还包括：
获取目标领域的文本语句；
对所述文本语句进行分词处理，得到所述文本语句的分词结果；
对所述文本语句的分词结果进行词向量训练，得到词向量模型；
相应的，
所述获取所述样本问题语句中各个词对应的词向量包括：
对所述样本问题语句进行分词处理，得到所述样本问题语句的分词结果；
根据所述词向量模型，获取所述样本问题语句中各个词对应的词向量。
本领域内的技术人员应明白，本申请的实施例可提供为方法、系统、或计 算机程序产品。因此，本申请可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本申请可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本申请是参照根据本申请实施例的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上，使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
在一个典型的配置中，计算设备包括一个或多个处理器(CPU)、输入/输出接口、网络接口和内存。
存储器可能包括计算机可读介质中的非永久性存储器，随机存取存储器(RAM)和/或非易失性内存等形式，如只读存储器(ROM)或闪存(flash RAM)。存储器是计算机可读介质的示例。
计算机可读介质包括永久性和非永久性、可移动和非可移动媒体可以由任何方法或技术来实现信息存储。信息可以是计算机可读指令、数据结构、程序的模块或其他数据。计算机的存储介质的例子包括，但不限于相变内存(PRAM)、静态随机存取存储器(SRAM)、动态随机存取存储器(DRAM)、其他类型的随机存取存储器(RAM)、只读存储器(ROM)、电可擦除可编程只读 存储器(EEPROM)、快闪记忆体或其他内存技术、只读光盘只读存储器(CD-ROM)、数字多功能光盘(DVD)或其他光学存储、磁盒式磁带，磁带磁磁盘存储或其他磁性存储设备或任何其他非传输介质，可用于存储可以被计算设备访问的信息。按照本文中的界定，计算机可读介质不包括暂存电脑可读媒体(transitory media)，如调制的数据信号和载波。
还需要说明的是，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、商品或者设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、商品或者设备所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括要素的过程、方法、商品或者设备中还存在另外的相同要素。
本领域技术人员应明白，本申请的实施例可提供为方法、系统或计算机程序产品。因此，本申请可采用完全硬件实施例、完全软件实施例或结合软件和硬件方面的实施例的形式。而且，本申请可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
以上仅为本申请的实施例而已，并不用于限制本申请。对于本领域技术人员来说，本申请可以有各种更改和变化。凡在本申请的精神和原理之内所作的任何修改、等同替换、改进等，均应包含在本申请的权利要求范围之内。

Claims (10)

  1. 一种自动问答方法，其中，包括：
    获取用户问题语句与候选问题集，其中，所述候选问题集中的问题语句均对应有预设的答案信息；
    利用预置的问题模型，从所述候选问题集中确定出与所述用户问题语句相匹配的问题语句，作为目标问题语句；其中，所述问题模型为，将历史问题语句作为训练样本对注意力神经网络进行训练得到的；
    采用所述目标问题语句对应的答案信息，对所述用户问题语句进行回应。
  2. 如权利要求1所述的方法，其中，所述获取用户问题语句与候选问题集包括：
    获取用户问题语句；
    采用BM25算法，从预置问题库中检索出与所述用户问题语句相关联的至少一条问题语句，作为候选问题集；所述预置问题库中的问题语句均对应有预设的答案信息。
  3. 如权利要求1所述的方法，其中，所述利用预置的问题模型，从所述候选问题集中确定出与所述用户问题语句相匹配的问题语句，作为目标问题语句包括：
    利用预置的问题模型，获取所述用户问题语句与所述候选问题集中的问题语句的相似度；
    将所述候选问题集中所述相似度符合预设相似度条件的问题语句，确定为与所述用户问题语句相匹配的问题语句，作为目标问题语句。
  4. 如权利要求1所述的方法，其中，所述问题模型的训练过程包括：
    获取训练样本，所述训练样本中包括样本问题语句；
    采用所述样本问题语句对注意力神经网络进行训练，得到所述问题模型。
  5. 如权利要求4所述的方法，其中，所述获取训练样本包括：
    获取第一问题语句；
    获取与所述第一问题语句语义相同的第二问题语句，以及，与所述第一问题语句语义不同的第三问题语句；
    将所述第二问题语句作为所述第一问题语句的正样本，将所述第三问题语 句作为所述第一问题语句的负样本；其中，所述正样本与所述负样本为训练样本。
  6. 如权利要求4所述的方法，其中，所述采用所述样本问题语句对注意力神经网络进行训练，得到所述问题模型包括：
    获取所述样本问题语句中各个词对应的词向量；
    根据所述词向量，利用双向门结构的循环神经网络，抽取所述样本问题语句中各个时刻对应的特征向量；
    利用注意力机制，确定所述样本问题语句中各个时刻对应的权重信息；
    根据所述样本问题语句中各个时刻对应的特征向量与权重信息，确定所述样本问题语句对应的句向量；
    根据所述句向量，确定所述注意力神经网络的网络参数，得到所述问题模型。
  7. 如权利要求6所述的方法，其中，在所述采用所述样本问题语句对注意力神经网络进行训练之前，所述方法还包括：
    获取目标领域的文本语句；
    对所述文本语句进行分词处理，得到所述文本语句的分词结果；
    对所述文本语句的分词结果进行词向量训练，得到词向量模型；
    相应的，
    所述获取所述样本问题语句中各个词对应的词向量包括：
    对所述样本问题语句进行分词处理，得到所述样本问题语句的分词结果；
    根据所述词向量模型，获取所述样本问题语句中各个词对应的词向量。
  8. 一种自动问答装置，其中，包括：
    问题获取单元，用于获取用户问题语句与候选问题集，其中，所述候选问题集中的问题语句均对应有预设的答案信息；
    问题确定单元，用于利用预置的问题模型，从所述候选问题集中确定出与所述用户问题语句相匹配的问题语句，作为目标问题语句；其中，所述问题模型为，将历史问题语句作为训练样本对注意力神经网络进行训练得到的；
    问题回应单元，用于采用所述目标问题语句对应的答案信息，对所述用户问题语句进行回应。
  9. 一种存储介质，其中，所述存储介质包括存储的程序，其中，在所述 程序运行时控制所述存储介质所在设备执行如权利要求1-7中任一项所述的自动问答方法。
  10. 一种处理器，其中，所述处理器用于运行程序，其中，所述程序运行时执行如权利要求1-7中任一项所述的自动问答方法。
PCT/CN2019/073662 2018-05-23 2019-01-29 自动问答方法及装置 WO2019223362A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810502726.3 2018-05-23
CN201810502726.3A CN110597966A (zh) 2018-05-23 2018-05-23 自动问答方法及装置

Publications (1)

Publication Number Publication Date
WO2019223362A1 true WO2019223362A1 (zh) 2019-11-28

Family

ID=68616527

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/073662 WO2019223362A1 (zh) 2018-05-23 2019-01-29 自动问答方法及装置

Country Status (2)

Country Link
CN (1) CN110597966A (zh)
WO (1) WO2019223362A1 (zh)

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113268572A (zh) * 2020-02-14 2021-08-17 华为技术有限公司 问答方法及装置
CN111737424A (zh) * 2020-02-21 2020-10-02 北京沃东天骏信息技术有限公司 一种问题的匹配方法、装置、设备及存储介质
CN111666376B (zh) * 2020-05-21 2023-07-18 武汉大学 一种基于段落边界扫描预测与词移距离聚类匹配的答案生成方法及装置
CN111666770B (zh) * 2020-06-02 2023-07-18 泰康保险集团股份有限公司 一种语义匹配方法及装置
CN111984763B (zh) * 2020-08-28 2023-09-19 海信电子科技(武汉)有限公司 一种答问处理方法及智能设备
CN112380328B (zh) * 2020-11-11 2024-02-06 广州知图科技有限公司 一种安全应急响应机器人交互方法及系统
CN114691815A (zh) * 2020-12-25 2022-07-01 科沃斯商用机器人有限公司 模型训练方法、装置、电子设备和存储介质
CN113553412B (zh) * 2021-06-30 2023-07-25 北京百度网讯科技有限公司 问答处理方法、装置、电子设备和存储介质
CN113486165A (zh) * 2021-07-08 2021-10-08 山东新一代信息产业技术研究院有限公司 一种用于云化机器人的faq自动问答方法、设备及介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106649868A (zh) * 2016-12-30 2017-05-10 首都师范大学 问答匹配方法及装置
CN106649612A (zh) * 2016-11-29 2017-05-10 中国银联股份有限公司 一种自动问答模板匹配的方法及装置
CN107358948A (zh) * 2017-06-27 2017-11-17 上海交通大学 基于注意力模型的语言输入关联性检测方法
CN107545003A (zh) * 2016-06-28 2018-01-05 中兴通讯股份有限公司 自动问答方法及系统
CN107967302A (zh) * 2017-11-08 2018-04-27 江苏名通信息科技有限公司 基于深度神经网络的游戏客服对话系统

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7392185B2 (en) * 1999-11-12 2008-06-24 Phoenix Solutions, Inc. Speech based learning/training system using semantic decoding
US10628521B2 (en) * 2015-08-03 2020-04-21 International Business Machines Corporation Scoring automatically generated language patterns for questions using synthetic events
CN106484664B (zh) * 2016-10-21 2019-03-01 竹间智能科技（上海）有限公司 一种短文本间相似度计算方法
CN107980130A (zh) * 2017-11-02 2018-05-01 深圳前海达闼云端智能科技有限公司 自动回答方法、装置、存储介质及电子设备
CN108052588B (zh) * 2017-12-11 2021-03-26 浙江大学城市学院 一种基于卷积神经网络的文档自动问答系统的构建方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107545003A (zh) * 2016-06-28 2018-01-05 中兴通讯股份有限公司 自动问答方法及系统
CN106649612A (zh) * 2016-11-29 2017-05-10 中国银联股份有限公司 一种自动问答模板匹配的方法及装置
CN106649868A (zh) * 2016-12-30 2017-05-10 首都师范大学 问答匹配方法及装置
CN107358948A (zh) * 2017-06-27 2017-11-17 上海交通大学 基于注意力模型的语言输入关联性检测方法
CN107967302A (zh) * 2017-11-08 2018-04-27 江苏名通信息科技有限公司 基于深度神经网络的游戏客服对话系统

Also Published As

Publication number Publication date
CN110597966A (zh) 2019-12-20

Similar Documents

Publication Publication Date Title
WO2019223362A1 (zh) 自动问答方法及装置
US11003856B2 (en) Processing text using neural networks
Nagarhalli et al. A review of current trends in the development of chatbot systems
WO2018157805A1 (zh) 一种自动问答处理方法及自动问答系统
CN109460459B (zh) 一种基于日志学习的对话系统自动优化方法
WO2022227162A1 (zh) 问答数据处理方法、装置、计算机设备及存储介质
CN106202153A (zh) 一种es搜索引擎的拼写纠错方法及系统
JP2017073137A (ja) 記述的なトピックラベルの生成
CN109359178A (zh) 一种检索方法、装置、存储介质及设备
US11526512B1 (en) Rewriting queries
CN111291177A (zh) 一种信息处理方法、装置和计算机存储介质
US11417339B1 (en) Detection of plagiarized spoken responses using machine learning
US20230244934A1 (en) Augmenting machine learning language models using search engine results
CN109522920B (zh) 基于结合语义特征的同义判别模型的训练方法及设备
CN116150306A (zh) 问答机器人的训练方法、问答方法及装置
CN114722176A (zh) 一种智能答疑的方法、装置、介质及电子设备
Hakimov et al. Evaluating architectural choices for deep learning approaches for question answering over knowledge bases
Ye et al. A sentiment based non-factoid question-answering framework
Karpagam et al. Deep learning approaches for answer selection in question answering system for conversation agents
CN116028626A (zh) 文本匹配方法、装置、存储介质以及电子设备
Khan et al. Making Personal Digital Assistants Aware of What They Do Not Know.
WO2021063060A1 (zh) 文本信息提取方法、装置、存储介质及设备
CN112765328A (zh) 一种文本相似度确定方法、系统、存储介质以及设备
Harichandana et al. Adaptive Beam Search to Enhance On-device Abstractive Summarization
Zhang Machine learning techniques for holistic computational paralinguistics

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: 19806634

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19806634

Country of ref document: EP

Kind code of ref document: A1