Dialogue system, dialogue processing method, translating apparatus and method of translation

A semantic analysis model using an N-best algorithm enhances conversational and translation systems by accurately generating system responses and translations through improved semantic analysis of user utterances, addressing the accuracy issues in existing technologies.

KR103047027B1Active Publication Date: 2026-11-11HYUNDAI MOTOR CO LTD +1
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
KR1020190118290
Authority / Receiving Office
KR · KR
Patent Type
Patents
Current Assignee / Owner
Filing Date
2019-09-25
Publication Date
2026-11-11
Estimated Expiration
2039-09-25

AI Technical Summary

Technical Problem

Conversational systems and translation devices struggle with accuracy in generating system responses and translating sentences due to the lack of effective semantic analysis of user utterances, leading to suboptimal user experience and translation quality.

Method used

Implementing a semantic analysis model that utilizes an N-best algorithm to rank and re-rank output sentences based on similarity and confidence scores, ensuring accurate generation of system responses and translations by analyzing the meaning of user utterances.

Benefits of technology

Improves the accuracy of system responses and translation quality by aligning the semantic representation of input and output sentences, enhancing user convenience and translation precision.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure R1020190118290_ABST
    Figure R1020190118290_ABST
Patent Text Reader

Abstract

The disclosed invention provides a dialogue system and a dialogue processing method that can improve the accuracy of system responses and user convenience by generating system responses using a semantic analysis model used to analyze the meaning of user utterances. In addition, the disclosed invention provides a translation device and a translation method that can improve translation accuracy by using a semantic analysis model used to analyze the meaning of user utterances when selecting a final output sentence from a plurality of output sentences that have been translated from an input sentence into another language. A dialogue system according to one embodiment includes: a speech recognizer that converts user utterance into text to generate an input sentence; a dialogue manager that generates a semantic expression for the input sentence; and a result processor that generates a plurality of output sentences corresponding to the semantic expression. The dialogue manager generates a semantic expression for each of the plurality of output sentences, and the result processor can generate a system response based on the semantic expression for the input sentence and the semantic expression for each of the plurality of output sentences.
Need to check novelty before this filing date? Find Prior Art

Description

Technology Field

[0001] The disclosed invention relates to a conversation system capable of exchanging conversations with a user and a conversation processing method, and a translation device and translation method for translating an input sentence into another language and outputting it. Background Technology

[0002] A conversational system can recognize a user's utterance and provide a service corresponding to the recognized utterance.

[0003] The conversational system may output a system response to provide a service corresponding to a recognized utterance. The system response may include an answer to a user's question, a question to verify the user's utterance, or guidance on the service to be provided.

[0004] The system responses output by the conversational system allow users to feel as if they are actually having a conversation, enabling them to speak naturally as if conversing with someone. Therefore, when evaluating the performance of a conversational system, the extent to which the system responses match user utterances can serve as an important factor. The problem to be solved

[0005] The disclosed invention provides a dialogue system and a dialogue processing method that can improve the accuracy of system responses and user convenience by generating system responses using a semantic analysis model used to analyze the meaning of user utterances.

[0006] In addition, the disclosed invention provides a translation device and a translation method that can improve translation accuracy by using a semantic analysis model used to analyze the meaning of user utterances when selecting a final output sentence from a plurality of output sentences that have been translated from an input sentence into another language. means of solving the problem

[0007] A dialogue system according to one embodiment includes: a speech recognizer that converts user utterance into text to generate an input sentence; a dialogue manager that generates a semantic expression for the input sentence; and a result processor that generates a plurality of output sentences corresponding to the semantic expression. The dialogue manager generates a semantic expression for each of the plurality of output sentences, and the result processor can generate a system response based on the semantic expression for the input sentence and the semantic expression for each of the plurality of output sentences.

[0008] The above result processor can determine the ranking of the plurality of output sentences using an N-best algorithm.

[0009] The above result processor can re-rank the plurality of output sentences based on the similarity between the semantic representation of the input sentence and the semantic representation of each of the plurality of output sentences.

[0010] The above result processor can assign a confidence score to each of the plurality of output sentences using the N-best algorithm.

[0011] The above result processor can assign a similarity score to each of the plurality of output sentences based on the similarity between the semantic representation of the input sentence and the semantic representation of each of the plurality of output sentences.

[0012] The above result processor can generate the system response based on the sum of the reliability score and the similarity score.

[0013] A translation device according to one embodiment may include: a translator that translates an input sentence of a first language into a second language to generate a plurality of output sentences; a first dialogue manager that generates a semantic expression for the input sentence of the first language; a second dialogue manager that generates a semantic expression for each of the plurality of output sentences of the second language; and a decision unit that determines a final output sentence among the plurality of output sentences of the second language based on the semantic expression for the input sentence of the first language and the semantic expression for the plurality of output sentences of the second language.

[0014] The above translator can determine the ranking of multiple output sentences of the above second language using an N-best algorithm.

[0015] The above decision unit can re-rank the plurality of output sentences of the second language based on the similarity between the semantic expression of the input sentence of the first language and the semantic expression of each of the plurality of output sentences of the second language.

[0016] The above translator can assign a confidence score to each of multiple output sentences of the above second language using an N-best algorithm.

[0017] The above-described judgment device can assign a similarity score to each of the plurality of output sentences of the second language based on the similarity between the semantic expression of the input sentence of the first language and the semantic expression of each of the plurality of output sentences of the second language.

[0018] The above decision unit can determine the final output sentence based on the sum of the confidence score and the similarity score.

[0019] A dialogue processing method according to one embodiment comprises: converting a user utterance into text to generate an input sentence; generating a semantic expression for the input sentence; generating a plurality of output sentences corresponding to the semantic expression; generating a semantic expression for each of the plurality of output sentences; and generating a system response based on the semantic expression for the input sentence and the semantic expression for each of the plurality of output sentences.

[0020] Generating multiple output sentences corresponding to the above semantic expression may include determining the ranking of the multiple output sentences using an N-best algorithm.

[0021] Generating the above system response may include re-ranking the plurality of output sentences based on the similarity between the semantic representation of the input sentence and the semantic representation of each of the plurality of output sentences.

[0022] Generating a plurality of output sentences corresponding to the above semantic expression may include assigning a confidence score to each of the plurality of output sentences using an N-best algorithm.

[0023] Generating the above system response may include assigning a similarity score to each of the plurality of output sentences based on the similarity between the semantic representation of the input sentence and the semantic representation of each of the plurality of output sentences.

[0024] Generating the above system response may further include determining the system response among the plurality of output sentences based on a sum of the reliability score and the similarity score.

[0025] A translation method according to one embodiment comprises: translating an input sentence of a first language into a second language to generate a plurality of output sentences; generating a semantic expression for the input sentence of the first language; generating a semantic expression for each of the plurality of output sentences of the second language; and determining a final output sentence among the plurality of output sentences of the second language based on the semantic expression for the input sentence of the first language and the semantic expression for the plurality of output sentences of the second language.

[0026] Generating multiple output sentences of the second language may include determining the ranking of multiple output sentences of the second language using an N-best algorithm.

[0027] Determining the final output sentence above may include re-ranking the plurality of output sentences of the second language based on the similarity between the semantic expression of the input sentence of the first language and the semantic expression of each of the plurality of output sentences of the second language.

[0028] Generating multiple output sentences of the second language may include assigning a confidence score to each of the multiple output sentences of the second language using an N-best algorithm.

[0029] Determining the above final output sentence may include assigning a similarity score to each of the plurality of output sentences of the second language based on the similarity between the semantic expression of the input sentence of the first language and the semantic expression of each of the plurality of output sentences of the second language.

[0030] Determining the above final output sentence may further include determining the above final output sentence based on a sum of the reliability score and the above similarity score. Effects of the invention

[0031] According to a conversational system and conversational processing method according to one aspect, the accuracy of the system response and user convenience can be improved by generating a system response using a semantic analysis model used to analyze the meaning of user utterances.

[0032] According to one aspect, a translation device and translation method can improve the accuracy of translation by using a semantic analysis model used to analyze the meaning of a user's utterance when selecting a final output sentence from among multiple output sentences that have translated an input sentence into another language. Brief explanation of the drawing

[0033] FIG. 1 is a control block diagram of a conversation system according to one embodiment. FIG. 2 is a control block diagram illustrating a result processor in a conversational system according to one embodiment. FIGS. 3 to 6 are diagrams illustrating a specific process in which a system response is output when a user utterance is input into a conversational system according to one embodiment. FIG. 7 is a control block diagram of a translation device according to one embodiment. FIGS. 8 to 10 are drawings illustrating a specific process in which a translated sentence is output when an input sentence is received by a translation device (200) according to one embodiment. FIG. 11 is a flowchart of a dialogue processing method according to one embodiment. FIG. 12 is a flowchart of a translation method according to one embodiment. Specific details for implementing the invention

[0034] The embodiments described in this specification and the configurations illustrated in the drawings are preferred examples of the disclosed invention, and various modifications that may replace the embodiments and drawings of this specification may exist at the time of filing this application.

[0035] Furthermore, the terms used in this specification are for describing embodiments and are not intended to limit or / or restrict the disclosed invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this specification, terms such as “comprising,” “comprising,” or “having” are intended to indicate the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and do not preclude the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0036] In addition, terms such as "~part," "~unit," "~block," "~part," and "~module" may refer to a unit that processes at least one function or operation. For example, the above terms may refer to at least one piece of hardware such as an FPGA (field-programmable gate array) or an ASIC (application specific integrated circuit), at least one piece of software stored in memory, or at least one process processed by a processor.

[0037] Furthermore, ordinal numbers such as "First~" and "Second~" used before the components described in this specification are used solely to distinguish the components from one another and do not have any other meanings, such as the order of connection, order of use, or priority among these components.

[0038] The symbols attached to each step are used to identify each step and do not indicate the order of the steps relative to one another; the steps may be performed differently from the specified order unless a specific order is clearly indicated in the context.

[0039] Meanwhile, the disclosed embodiments may be implemented in the form of a recording medium that stores instructions executable by a computer. The instructions may be stored in the form of program code and, when executed by a processor, can perform the operation of the disclosed embodiments. The recording medium may be implemented as a computer-readable recording medium.

[0040] Computer-readable recording media include all types of recording media that store instructions that can be decoded by a computer. Examples include ROM (Read Only Memory), RAM (Random Access Memory), magnetic tape, magnetic disk, flash memory, optical data storage devices, etc.

[0041] Hereinafter, embodiments relating to a dialogue system, a dialogue processing method, a translation device, and a translation method according to one aspect will be described in detail with reference to the attached drawings.

[0042] A conversational system according to one embodiment is a device that analyzes a user's utterance to identify the user's intent and provides a service suitable for the user's intent. By outputting a system response in providing a service suitable for the user's intent, the conversational system can give the user the feeling of having a conversation with the conversational system. The system response may include an answer to the user's question, a question to confirm the user's intent, and guidance on the service to be provided.

[0043] FIG. 1 is a control block diagram of a conversation system according to one embodiment.

[0044] Referring to FIG. 1, a conversation system (100) according to one embodiment includes a speech recognizer (110) that converts user utterance into text to generate an input sentence, a conversation manager (120) that generates a semantic expression for the input sentence, and a result processor (130) that generates a plurality of output sentences corresponding to the semantic expression.

[0045] A conversation system (100) according to one embodiment may include at least one memory that stores a program and related data that perform the operation described below, and at least one processor that executes the stored program.

[0046] The components in the present embodiment are distinguished based on operation rather than physical location or configuration, and it is possible for the voice recognizer (110), dialogue manager (120), and result processor (130) to each use separate memory and processor, and it is also possible for some or all of them to share memory and processor.

[0047] A conversation system (100) according to one embodiment may be implemented as a server, in which case the components (110, 120, 130) of the conversation system (100) may be provided on the server. However, it is also possible for some of the components (110, 120, 130) of the conversation system (100) to be provided on a user terminal that connects the user with the conversation system (100).

[0048] For example, if the user terminal is a vehicle, some of the components of the conversation system (100) may be provided in the vehicle, and if the user terminal is a mobile device such as a smartphone, an AI speaker, or a PC, some of the components of the conversation system (100) may be provided in the mobile device, an AI speaker, or a PC.

[0049] Alternatively, it is also possible for the user to download and use a program that performs some of the functions of the components of the conversation system (100) to a user terminal.

[0050] A microphone for inputting user speech and a speaker for outputting system responses may be provided in a user terminal such as a vehicle, mobile device, or PC, and the user terminal may be connected to a conversation system (100) via wireless communication. Although not shown in the control block diagram of FIG. 1, the conversation system (100) may further include a wireless communication module capable of exchanging data with the user terminal via wireless communication.

[0051] User speech input through the microphone can be transmitted to the wireless communication module of the conversation system (100).

[0052] The speech recognizer (110) can convert transmitted user utterances into text by applying a speech recognition algorithm or a STT (Speech to Text) algorithm. For example, the speech recognizer (110) can extract a feature vector of a voice signal corresponding to a user utterance by applying feature vector extraction techniques such as Cepstrum, Linear Predictive Coefficient (LPC), Mel Frequency Cepstral Coefficient (MFCC), or Filter Bank Energy.

[0053] In addition, recognition results can be obtained by comparing the extracted feature vectors with trained reference patterns. To this end, an acoustic model that models and compares the signal characteristics of speech and a language model that models the linguistic sequential relationships, such as words or syllables corresponding to the vocabulary of recognition, may be used.

[0054] Acoustic models can be further divided into a direct comparison method, which sets the recognition target as a feature vector model and compares it with the feature vectors of speech data, and a statistical method, which utilizes the feature vectors of the recognition target by statistically processing them.

[0055] The direct comparison method sets units such as words or phonemes to be recognized as feature vector models and compares how similar the input speech is to them; a representative example is Vector Quantization. According to Vector Quantization, the feature vectors of the input speech data are mapped to a reference model, a codebook, and encoded into representative values, which are then compared with each other.

[0056] Statistical model methods are methods that organize units of the recognition target into a state sequence and utilize the relationships between state sequences. A state sequence can be composed of multiple nodes. Methods that utilize the relationships between state sequences include Dynamic Time Warping (DTW), Hidden Markov Models (HMM), and methods using neural networks.

[0057] Dynamic time warping is a technique that compensates for differences in the time axis when compared with a reference model by considering the dynamic characteristics of speech, where the length of the signal changes over time even if the same person pronounces the same sound. The Hidden Markov Model is a recognition technique that assumes speech as a Markov process having state transition probabilities and observation probabilities of nodes (output symbols) in each state, estimates state transition probabilities and node observation probabilities through training data, and calculates the probability of the input speech occurring in the estimated model.

[0058] Meanwhile, language models that model linguistic sequential relationships, such as words or syllables, can reduce acoustic ambiguity and recognition errors by applying the sequential relationships between units that constitute language to units obtained from speech recognition. Language models include statistical language models and models based on finite state networks (Finite State Automata: FSA), and statistical language models utilize word sequence probabilities such as Unigram, Bigram, and Trigram.

[0059] The speech recognizer (110) may use any of the methods described above to recognize a speech signal corresponding to a user's speech and convert it into text. For example, an acoustic model with a hidden Markov model applied may be used, or a model that integrates an acoustic model and a speech model may be used.

[0060] The speech recognition unit (110) converts user utterance into text and inputs it into the conversation manager (120). In the embodiment described below, the user utterance converted into text is referred to as an input sentence.

[0061] The conversation manager (120) can identify the user's intent included in the user's utterance by applying natural language understanding technology to the input sentence.

[0062] Specifically, the conversation manager (120) performs morphological analysis on user utterances in text form to convert the input string into a sequence of morphemes. Additionally, the conversation manager (120) can recognize named entities from user utterances. Named entities are proper nouns such as personal names, place names, organization names, times, dates, and currency. Named entity recognition is the process of identifying named entities in a sentence and determining the type of the identified named entities. Through named entity recognition, important keywords can be extracted from the sentence to understand the meaning of the sentence.

[0063] The conversation manager (120) can extract a domain from a user's utterance. The domain is capable of identifying the subject of the language spoken by the user, and various domains such as text, navigation, schedule, weather, traffic information, vehicle control, and home appliance control may be predetermined, and the conversation manager (120) can determine which domain of the user's utterance is among the predetermined domains.

[0064] The conversation manager (120) can analyze the speech acts of a user's utterance. Speech act analysis is a task of analyzing the intent of a user's utterance, and is used to determine the intent of the utterance, such as whether the user is asking a question, making a request, responding, or simply expressing an emotion.

[0065] The conversation manager (120) can identify the user's intent based on information such as the domain, entity name, and speech act corresponding to the user's utterance, and extract an action corresponding to the user's intent. The action can be defined by an object and an operator.

[0066] For example, if the identified user intent is [Run Air Conditioner], the corresponding action can be defined as 『Air Conditioner(Object)_ON(Operator)』, and if the identified user intent is [Send Text Message], the corresponding action can be defined as 『Text(Object)_Send(Operator)』.

[0067] Additionally, the conversation manager (120) can also extract factors related to action execution. Factors related to action execution may be valid factors directly required to perform the action, or they may be invalid factors used to extract such valid factors.

[0068] For example, if the input sentence generated by the speech recognition unit (110) is "Send a text message to Gildong," the conversation manager (120) can determine the domain corresponding to the user utterance as "text" and determine the action as "text_send". The speech act corresponds to "request".

[0069] The named entity "Gildongi" corresponds to {Argument 1: Recipient} related to the action execution, and {Argument 2: Specific message content} is additionally required for the actual text transmission.

[0070] The dialogue manager (120) can generate a meaning representation used to generate a system response corresponding to the user's intention or to provide a service corresponding to the user's intention in the result processor (130), based on the analysis results of the text (input sentence) described above. The meaning representation in dialogue processing may be the result of natural language understanding or the input of natural language generation. For example, the dialogue manager (120) can generate a meaning representation that expresses the user's intention by analyzing the user's utterance, and can generate a meaning representation corresponding to the next system utterance by considering the flow and situation of the conversation.

[0071] The result processor (130) can generate a sentence to be output as a system response (hereinafter referred to as the output sentence) based on the semantic expression output from the dialogue manager (120), and the generated sentence can be synthesized into a voice signal by a Text-to-Speech (TTS) engine provided in the result processor (130) and output through a speaker provided in the user terminal.

[0072] Meanwhile, the result processor (130) can generate output sentences by applying natural language generation technology, and can generate multiple output sentences by applying an N-best algorithm.

[0073] When the result processor (130) generates three output sentences (output sentence 1, output sentence 2, output sentence 3) based on the semantic expression of the input sentence, the three output sentences are input to the dialogue manager (120), and the dialogue manager (120) can generate a semantic expression for each of the multiple output sentences.

[0074] As described above, the dialogue manager (120) can analyze the meaning of an input sentence by applying natural language understanding technology. The dialogue manager (120) can also generate a semantic expression by applying the same technology to an output sentence generated by a result processor (130).

[0075] The conversation manager (120) can generate semantic expression 1 for output sentence 1, semantic expression 2 for output sentence 2, and semantic expression 3 for output sentence 3, respectively, and send them back to the result processor (130).

[0076] The result processor (130) can generate a system response based on the semantic representation of the input sentence and the semantic representation of each of the multiple output sentences.

[0077] Hereinafter, the process by which a result processor generates a system response will be explained in more detail with reference to FIG. 2.

[0078] FIG. 2 is a control block diagram illustrating a result processor in a conversational system according to one embodiment.

[0079] Referring to FIG. 2, the result processor (130) of the conversational system (100) according to one embodiment may include a sentence generator (131) and a decision generator (132). These components are also classified based on the operation they perform, and may share a memory in which the program and related data are stored and a processor that executes the program.

[0080] The sentence generator (131) can generate multiple output sentences by applying natural language generation technology to the semantic representation of the input sentence generated by the dialogue manager (120).

[0081] The sentence generator (131) can calculate a confidence score for each of multiple output sentences using the N-best algorithm and output N output sentences in order of highest confidence score. That is, the ranking of multiple output sentences can be determined using the N-best algorithm. The confidence score is a measure indicating how reliable the result is for the natural language generation result, and can be expressed as a value between 0 and 1.

[0082] As shown in the example of FIG. 2, the sentence generator (131) can output three output sentences in order of highest confidence score, and the three output sentences are input back into the dialogue manager (120).

[0083] The dialogue manager (120) can generate semantic expressions for each of the three output sentences by applying natural language understanding technology. Semantic expression 1 for output sentence 1, semantic expression 2 for output sentence 2, and semantic expression 3 for output sentence 3 can be generated and output, and the outputted semantic expressions are input back into the result processor (130).

[0084] The decision unit (132) of the result processor (130) can re-rank the multiple output sentences based on the similarity between the semantic representation of the input sentence and the semantic representation of each of the multiple output sentences.

[0085] The more similar the semantic expression of the input sentence and the semantic expression of the output sentence are, the more the accuracy of the output sentence is improved. Accordingly, the decision unit (132) can assign a similarity score to each of the multiple output sentences based on the similarity between the semantic expression of the input sentence and the output sentence. A higher similarity score is assigned as the similarity increases.

[0086] The decision unit (132) calculates a sum score by summing the reliability score and the similarity score for each of the multiple output sentences, and can re-rank the multiple output sentences based on the sum score. The ranking can be re-ranked in order of highest sum score, and finally, the output sentence with the highest sum score can be selected as the system response.

[0087] Meanwhile, although only a sentence generator (131) and a decision unit (132) are shown in the result processor (130), it is of course possible to include a Text-to-Speech (TTS) engine that synthesizes the final output sentence output from the decision unit (132) into a speech signal, and a control signal generator that generates various control signals to provide a service corresponding to the user's intention.

[0088] The system response output from the result processor (130) according to the above-described operation can be output as voice through a speaker provided in the user terminal.

[0089] Hereinafter, specific examples of system responses output from a conversation system (100) according to one embodiment will be described with reference to FIGS. 3 to 6.

[0090] FIGS. 3 to 6 are diagrams illustrating a specific process in which a system response is output when a user utterance is input into a conversational system according to one embodiment.

[0091] Referring to the example in Fig. 3, when a user utterance "Send a text message to Hong Gil-dong" is input into the conversation system (100), the speech recognizer (110) converts the user utterance into text to generate an input sentence.

[0092] The input sentence is input into the conversation manager (120), and natural language understanding technology is applied to analyze the meaning of the input sentence "Send a text message to Hong Gil-dong" and generate a meaning expression based on the analysis result.

[0093] The conversation manager (120) can determine that the action of the input sentence "Send a text message to Hong Gil-dong" is "Text_Send", the speech act is "Request", and the argument_Recipient is "Hong Gil-dong", and can generate a semantic expression containing this information. The generated semantic expression is input to the result processor (130).

[0094] The result processor (130) can generate multiple output sentences based on the semantic representation of the input sentence. As described above, a reliability score can be calculated for the multiple output sentences, and N output sentences can be ranked in order of highest reliability score. Referring to the example in FIG. 4, the result processor (130) can generate output sentence 1: "Is this text message being sent to Hong Gil-dong correct?", output sentence 2: "Please tell me the content of the text message to be sent to Hong Gil-dong", and output sentence 3: "Please tell me the content of the text message being sent" based on the semantic representation of the input sentence, and can rank the three sentences based on the reliability score. In this example, output sentence 1 (reliability score: 0.7) Output sentence 2 (Confidence score: 0.6) The ranking is determined in the order of output sentence 3 (confidence score: 0.3).

[0095] Multiple output sentences generated based on the semantic expression of the input sentence are input into the dialogue manager (120).

[0096] The conversation manager (120) can analyze the meaning of each of the multiple output sentences by applying natural language understanding technology in the same way that the meaning of the input sentence was analyzed, and can generate a semantic expression.

[0097] Referring to the example in Fig. 5, as a result of semantic analysis of output sentence 1, it can be determined that the action is "text_sending," the speech act is "confirm," and the argument_recipient is "Hong Gil-dong," and a semantic expression 1 containing this information can be generated.

[0098] Based on the semantic analysis of output sentence 2, it can be determined that the action is "text_send", the speech act is "request", and the argument_recipient is "Hong Gil-dong", and a semantic expression 2 containing this information can be generated.

[0099] Based on the semantic analysis of output sentence 3, it can be determined that the action is "text_sending" and the speech act is "request", and semantic expression 3 containing this information can be generated.

[0100] The semantic expression for each of the multiple output sentences generated by the dialogue manager (120) is input back into the result processor (130).

[0101] The result processor (130) can assign a similarity score to each of the multiple output sentences by comparing the semantic expression for each of the multiple output sentences with the semantic expression for the input sentence. As described above, a higher similarity score can be assigned as the semantic expressions are more similar.

[0102] Referring to the example in Figure 6, the similarity score of output sentence 1 based on the comparison result of semantic expressions is 0.66, the similarity score of output sentence 2 is 1.0, and the similarity score of output sentence 3 is 0.66.

[0103] The result processor (130) can calculate a sum score by summing the confidence score and the similarity score for each of the multiple output sentences. The sum score for output sentence 1 is 0.7+0.66=1.36, the sum score for output sentence 2 is 0.6+1.0=1.6, and the sum score for output sentence 3 is 0.3+0.66=0.96.

[0104] The result processor (130) reorders the order among multiple output sentences based on the sum score. When the order is reordered in order of highest sum score, output sentence 2 becomes 1st rank, output sentence 1 becomes 2nd rank, and output sentence 3 becomes 3rd rank.

[0105] Therefore, the result processor (130) can convert the first-priority output sentence 2 "Please tell me the content of the text message to send to Hong Gil-dong" into a voice signal and output it as a system response.

[0106] As mentioned above, the higher the similarity between the semantic expression of the input sentence and the semantic expression of the output sentence, the higher the accuracy of the output sentence can be. Since the same principle can be applied to machine translation to improve the accuracy of the translation result, a translation device according to one embodiment will be described below.

[0107] FIG. 7 is a control block diagram of a translation device according to one embodiment.

[0108] Referring to FIG. 7, a translation device (200) according to one embodiment includes a translator (210) that translates an input sentence of a first language into a second language to generate a plurality of output sentences, a first semantic analyzer (220) that generates a semantic expression for an input sentence of the first language, a second semantic analyzer (230) that generates a semantic expression for each of a plurality of output sentences of the second language, and a judgment unit (240) that determines a final output sentence among a plurality of output sentences of the second language based on the semantic expression for an input sentence of the first language and the semantic expression for a plurality of output sentences of the second language.

[0109] A translation device (200) according to one embodiment may be a device that performs machine translation to translate an input sentence into another language. A user may input the input sentence in the form of text or input it in the form of a voice signal through speech. In the case of inputting in the form of a voice signal, a voice recognition device may convert the voice signal into text and input it to the translation device (200), and it is also possible for the translation device (200) to include a voice recognition device. In either case, the input sentence input to the translator (210) is considered to be text.

[0110] A translation device (200) according to one embodiment may include at least one memory that stores a program and related data that perform the operation described below, and at least one processor that executes the stored program.

[0111] The components in the present embodiment are distinguished based on operation rather than physical location or configuration, and it is possible for the translator (210), the first semantic analyzer (220), the second semantic analyzer (230), and the judgment unit (240) to each use separate memory and processors, and it is also possible for some or all of them to share memory and processors.

[0112] A translation device (200) according to one embodiment may be implemented as a server, and in this case, the components (210, 220, 230, 240) of the translation device (200) may be provided on the server. However, it is also possible for some of the components (210, 220, 230, 240) of the translation device (200) to be provided on a user terminal that connects the user and the translation device (200).

[0113] Alternatively, it is also possible for the user to download and use a program that performs some of the functions of the components of the conversation system (100) to a user terminal.

[0114] Input devices such as a microphone, touch screen, and keyboard for inputting sentences of the first language and output devices such as a display and speaker for outputting sentences translated into the second language may be provided in a user terminal such as a vehicle, mobile device, or PC, and the user terminal may be connected to the translation device (200) via wireless communication. Although not shown in the control block diagram of FIG. 7, the translation device (200) may further include a wireless communication module capable of exchanging data with the user terminal via wireless communication.

[0115] The translator (210) can generate multiple output sentences by translating an input sentence of a first language into a second language using one of various machine translation models. As long as the first language and the second language are different from each other, there is no restriction on the types of the first language and the second language, such as Korean, English, Chinese, Japanese, Spanish, Portuguese, German, French, Indian, etc.

[0116] When the translator (210) generates multiple output sentences, an N-best algorithm may be used. As described in the embodiment of the conversational system (100) above, the translator (210) may rank multiple output sentences of the second language, and the ranking of multiple output sentences of the second language may be determined in order of highest reliability score.

[0117] The first semantic analyzer (220) can apply natural language understanding technology to the input sentence of the first language to analyze the meaning of the input sentence and generate a semantic expression for the input sentence. The explanation of semantic analysis and semantic expression is as described in the embodiment of the conversational system (100) above.

[0118] The second semantic analyzer (220) can apply natural language understanding technology to multiple output sentences of the second language to analyze the meaning of each of the multiple output sentences and generate semantic expressions.

[0119] As shown in the example of FIG. 7, when the translator (210) generates output sentence 1, output sentence 2, and output sentence 3, the second semantic analyzer (230) can generate semantic expression 1 for output sentence 1, semantic expression 2 for output sentence 2, and semantic expression 3 for output sentence 3 and input them to the decision unit (240).

[0120] The first semantic analyzer (210) and the second semantic analyzer (230) differ only in the language of the sentence to be analyzed, but they are identical in that they apply natural language understanding technology to analyze the meaning.

[0121] The decision unit (240) can re-rank multiple output sentences of the second language based on the similarity between the semantic expression of the input sentence of the first language and the semantic expression of each of the multiple output sentences of the second language.

[0122] Specifically, the decision unit (240) assigns a higher similarity score as the similarity between semantic expressions increases, and can calculate a sum score by summing the reliability score and the similarity score for each of the multiple output sentences. The decision unit (240) can rank the multiple output sentences in order of highest sum score.

[0123] The judgment unit (240) can determine the output sentence with the highest sum score as the final output sentence, and the translation device (200) can output the final output sentence as a translation sentence for the input sentence. The final output sentence may be output as text, or it may be synthesized into a voice signal and output as a system response.

[0124] Hereinafter, an exemplary operation of a translation device (200) according to one embodiment will be specifically described with reference to FIGS. 8 to 10.

[0125] FIGS. 8 to 10 are drawings illustrating a specific process in which a translated sentence is output when an input sentence is received by a translation device (200) according to one embodiment.

[0126] In the example, the first language is English and the second language is Korean. As illustrated in FIG. 8, when the input sentence "Trump met Moon" in the first language is input into the translation device (200), the translator (210) can translate the input sentence into the second language, Korean, to generate multiple output sentences.

[0127] For example, you can generate output sentence 1 "Trump met the moon", output sentence 2 "Trump met President Moon", and output sentence 3 "Trump met Moon Jae-in".

[0128] The translator (210) can assign a reliability score to each of multiple output sentences by applying an N-best algorithm, for example, a reliability score of 0.7 to output sentence 1, a reliability score of 0.6 to output sentence 2, and a reliability score of 0.4 to output sentence 3.

[0129] Therefore, according to the N-best algorithm, the ranking of multiple output sentences is output sentence 1 Output statement 2 It is the order of output sentence 3.

[0130] Output sentence 1, output sentence 2, and output sentence 3 can be input into a second semantic analyzer (230). The second semantic analyzer (230) can generate semantic expressions by analyzing the semantics of the second language, Korean.

[0131] Referring to FIG. 9, the second semantic analyzer (230) can analyze the semantics of output sentence 1 to determine that the speech act is "information" and the entity name is "Trump / person name", and can generate a semantic expression 1 containing this information.

[0132] The second semantic analyzer (230) can analyze the semantics of the output sentence 2 to determine that the speech act is "information" and the entity names are "Trump / person name" and "President Moon / person name," and can generate a semantic expression 2 that includes this information.

[0133] The second semantic analyzer (230) can analyze the semantics of the output sentence 3 to determine that the speech act is "information" and the entity names are "Trump / person name" and "Moon Jae-in / person name," and can generate a semantic expression 3 that includes this information.

[0134] The semantic expression for the input sentence and the semantic expression for each of the multiple output sentences are input to the decision unit (240), and the decision unit (240) can determine the final output sentence by comparing the semantic expression for the input sentence and the semantic expression for each of the multiple output sentences.

[0135] Referring to the example of FIG. 10, the decision unit (240) can calculate a similarity score of 0.66 by comparing a semantic expression for an input sentence with a semantic expression 1 for an output sentence 1, calculate a similarity score of 1.0 by comparing a semantic expression for an input sentence with a semantic expression 2 for an output sentence 2, and calculate a similarity score of 1.0 by comparing a semantic expression for an input sentence with a semantic expression 3 for an output sentence 3.

[0136] The judgment unit (240) calculates a sum score by summing the reliability score and the similarity score for each of the multiple output sentences, and can re-rank based on the sum score.

[0137] According to the sum score, the ranking is output sentence 2 Output statement 3 The output sentences change in order, and the decision unit (240) can determine and output the output sentence 2 with the highest sum score as the final output sentence.

[0138] In this way, by considering the similarity between the semantic expression of the input sentence and the semantic expression of the output sentence translated from the input sentence, a translated sentence having the most similar meaning to the input sentence can be output to the user.

[0139] Hereinafter, a dialogue processing method and a translation method according to one embodiment will be described. In implementing the dialogue processing method according to one embodiment, the aforementioned dialogue system (100) may be used, and in implementing the translation method according to one embodiment, the aforementioned translation device (200) may be used. Accordingly, the previously described contents regarding the dialogue system (100) and the translation device (200) may be applied in the same way to the dialogue processing method and the translation method without separate mention.

[0140] FIG. 11 is a flowchart of a dialogue processing method according to one embodiment.

[0141] According to the conversation processing method illustrated in FIG. 11, user utterance is converted into text to generate an input sentence (310), and a semantic expression for the input sentence is generated (320). A speech recognizer (110) can convert user utterance into text using a speech recognition algorithm, and a conversation manager (120) can generate a semantic expression for the input sentence using natural language understanding technology.

[0142] A plurality of output sentences corresponding to the semantic expression are generated (330). When the semantic expression for an input sentence is input to the result processor (130), the result processor (130) uses an N-best algorithm to generate a plurality of output sentences corresponding to the semantic expression of the input sentence and can rank them based on the confidence score of each output sentence.

[0143] A semantic expression is generated for each of the multiple output sentences (340). The multiple output sentences generated by the result processor (130) are input into the dialogue manager (120), and the dialogue manager (120) can analyze the meaning of each of the multiple output sentences and generate a semantic expression by applying natural language understanding technology in the same way as it was done for the input sentences.

[0144] A system response is generated based on a semantic representation of an input sentence and a semantic representation of each of a plurality of output sentences (350). The semantic representation of each of the plurality of output sentences is input again to a result processor (130), and the result processor (130) can assign a similarity score to each of the plurality of output sentences based on the similarity between the semantic representation of the input sentence and the semantic representation of each of the plurality of output sentences. The higher the similarity between the semantic representations, the higher the similarity score can be assigned.

[0145] The result processor (130) calculates a sum score by summing the reliability score and the similarity score for each of the multiple output sentences, and can re-rank them according to the sum score. The result processor (130) can determine the output sentence with the highest sum score as the system response.

[0146] FIG. 12 is a flowchart of a translation method according to one embodiment.

[0147] According to the translation method illustrated in FIG. 12, an input sentence of a first language is translated into a second language to generate a plurality of output sentences (410). When an input sentence of the first language is received by the translator (210), the translator (210) generates a plurality of output sentences of the second language using an N-best algorithm, calculates a reliability score for each output sentence, and can rank the plurality of output sentences based on the reliability score.

[0148] A semantic expression for an input sentence of the first language can be generated (420), and a semantic expression for each of multiple output sentences of the second language can be generated (430). The first semantic analyzer (220) can analyze the meaning by applying natural language understanding technology performed by the conversational system (100) to the input sentence of the first language and generate a semantic expression. The second semantic analyzer (230) can analyze the meaning by applying natural language understanding technology performed by the conversational system (100) to multiple output sentences of the second language and generate a semantic expression.

[0149] Based on the semantic expression of the input sentence of the first language and the semantic expression of the multiple output sentences of the second language, the final output sentence is determined among the multiple output sentences of the second language (440). The semantic expression of the input sentence of the first language and the semantic expressions for each of the multiple output sentences of the second language are input to the decision unit (240). The decision unit (240) can re-rank the multiple output sentences of the second language based on the similarity between the semantic expression of the input sentence of the first language and the semantic expression of each of the multiple output sentences of the second language. The higher the similarity with the semantic expression of the input sentence, the higher the similarity score assigned to the output sentence of the second language, and a sum score can be calculated by summing the reliability score and the similarity score for each of the multiple output sentences. The decision unit (240) can re-rank the multiple output sentences based on the sum score, and the higher the sum score, the higher the rank. The decision unit (240) can determine the output sentence with the highest rank as the final output sentence according to the determined rank.

[0150] According to the above-described embodiment, when generating an output sentence for an input sentence in a conversational system or translation device, the accuracy of the output sentence can be improved by considering the similarity between the semantic expression of the input sentence and the semantic expression of the output sentence.

[0151] As described above, the disclosed embodiments have been explained with reference to the attached drawings. Those skilled in the art will understand that the present invention may be practiced in forms different from the disclosed embodiments without changing the technical spirit or essential features of the invention. The disclosed embodiments are illustrative and should not be interpreted restrictively. Explanation of the symbols

[0153] 100: Dialogue System 110: Voice Recognizer 120: Conversation Manager 130: Result handler 200: Translation Device 210: Translator 220: First Semantic Analyzer 230: Secondary Semantic Analyzer 240: Judgment

Claims

Claim 1 A conversational system comprising: a speech recognition unit that converts user utterance into text to generate an input sentence; a dialogue manager that generates a semantic expression for the input sentence; and a result processor that generates a plurality of output sentences corresponding to the semantic expression; wherein the dialogue manager generates a semantic expression for each of the plurality of output sentences, and the result processor generates a system response based on the semantic expression for the input sentence and the semantic expression for each of the plurality of output sentences, assigns a reliability score to each of the plurality of output sentences, assigns a similarity score to each of the plurality of output sentences based on the similarity between the user intent represented by the semantic expression for the input sentence and the user intent represented by the semantic expression for each of the plurality of output sentences, and generates the system response based on a sum of the reliability score and the similarity score. Claim 2 In claim 1, the result processor is a dialogue system that determines the ranking of the plurality of output sentences using an N-best algorithm. Claim 3 In claim 2, the result processor is a conversational system that re-ranks the plurality of output sentences based on the similarity between the semantic representation of the input sentence and the semantic representation of each of the plurality of output sentences. Claim 4 In claim 1, the result processor is a dialogue system that assigns a reliability score to each of the plurality of output sentences using an N-best algorithm. Claim 5 delete Claim 6 delete Claim 7 A translation device comprising: a translator that translates an input sentence of a first language into a second language to generate a plurality of output sentences; a first dialogue manager that generates a semantic expression for the input sentence of the first language; a second dialogue manager that generates a semantic expression for each of the plurality of output sentences of the second language; and a decision unit that determines a final output sentence among the plurality of output sentences of the second language based on the semantic expression for the input sentence of the first language and the semantic expression for the plurality of output sentences of the second language; wherein the translator assigns a reliability score to each of the plurality of output sentences, and the decision unit assigns a similarity score to each of the plurality of output sentences of the second language based on the similarity between the user intent represented by the semantic expression for the input sentence of the first language and the user intent represented by the semantic expression for each of the plurality of output sentences of the second language, and determines the final output sentence based on a sum of the reliability score and the similarity score. Claim 8 In claim 7, the translator is a translation device that ranks multiple output sentences of the second language using an N-best algorithm. Claim 9 In claim 8, the judgment device is a translation device that re-ranks a plurality of output sentences of the second language based on the similarity between the semantic expression of an input sentence of the first language and the semantic expression of each of a plurality of output sentences of the second language. Claim 10 In claim 7, the translator is a translation device that assigns a reliability score to each of a plurality of output sentences of the second language using an N-best algorithm. Claim 11 delete Claim 12 delete Claim 13 A conversation processing method performed by a conversation system, comprising: a speech recognition unit converting user utterance into text to generate an input sentence; a conversation manager generating a semantic expression for the input sentence; a result processor generating a plurality of output sentences corresponding to the semantic expression; the conversation manager generating a semantic expression for each of the plurality of output sentences; the result processor generating a system response based on the semantic expression for the input sentence and the semantic expression for each of the plurality of output sentences, assigning a reliability score to each of the plurality of output sentences, assigning a similarity score to each of the plurality of output sentences based on the similarity between the user intent represented by the semantic expression for the input sentence and the user intent represented by the semantic expression for each of the plurality of output sentences, and generating the system response based on a sum of the reliability score and the similarity score. Claim 14 A dialogue processing method according to claim 13, wherein generating a plurality of output sentences corresponding to the semantic expression comprises the result processor determining the ranking of the plurality of output sentences using an N-best algorithm. Claim 15 A dialogue processing method according to claim 14, wherein generating the system response comprises the result processor re-ranking the plurality of output sentences based on the similarity between the semantic representation of the input sentence and the semantic representation of each of the plurality of output sentences. Claim 16 A dialogue processing method according to claim 13, wherein generating a plurality of output sentences corresponding to the semantic expression comprises the result processor assigning a reliability score to each of the plurality of output sentences using an N-best algorithm. Claim 17 delete Claim 18 delete Claim 19 A translation method performed by a translation device, comprising: a translator translating an input sentence of a first language into a second language to generate a plurality of output sentences; a first dialogue manager generating a semantic expression for the input sentence of the first language; a second dialogue manager generating a semantic expression for each of the plurality of output sentences of the second language; the translator assigning a reliability score to each of the plurality of output sentences, and a judgment unit assigning a similarity score to each of the plurality of output sentences of the second language based on the similarity between the user intent represented by the semantic expression for the input sentence of the first language and the user intent represented by the semantic expression for each of the plurality of output sentences of the second language, and determining a final output sentence based on a sum of the reliability score and the similarity score. Claim 20 A translation method according to claim 19, wherein generating a plurality of output sentences of the second language comprises the translator determining the ranking of the plurality of output sentences of the second language using an N-best algorithm. Claim 21 A translation method according to claim 20, wherein determining the final output sentence comprises the judgment device re-ranking the plurality of output sentences of the second language based on the similarity between the semantic expression for the input sentence of the first language and the semantic expression for each of the plurality of output sentences of the second language. Claim 22 A translation method according to claim 19, wherein generating a plurality of output sentences of the second language comprises the translator assigning a reliability score to each of the plurality of output sentences of the second language using an N-best algorithm. Claim 23 delete Claim 24 delete

Citation Information

Patent Citations

  • apparatus for translation and method thereof

    KR1020130112654A

  • Voice dialogue method and its system

    JP3667614B2

  • Discriminating ambiguous expressions to enhance user experience

    US20160188565A1

  • Bootstrap call router to other languages using selected N-best translations

    US8838434B1