Abstract— This report deals with continuous conversation - based quality improvement system in daily conversation. In this study, daily conversation scenarios in various fields that do not depend on specific areas are collected. Based on this scenario, we propose a continuous dialogue method in which the dialogue reflecting the current trend is met according to user 's time - varying needs. In this continuous dialogue, we describe a method and system implementation for improving conversation quality.
I. INTRODUCTION
A. Research Background
Recently, various chatbots are on the market, and most chatbots work with specific purpose-based chat. In the case of purpose-based dialogues, the scenario dialogue proceeds according to a specific intent, so it appears to be a continuous dialogue. In everyday conversations without a specific purpose, it is difficult to communicate continuously, and most conversation systems can only conduct straightforward conversations. Therefore, we aim to continue conversation on a variety of subjects as if talking to a friend, and we will explain how to make continuous conversations, and how to improve conversation quality in continuous conversations.
B. Related Studies
Recently, a method of collecting daily conversation scenarios from Twitter has been studied[1]. The three-dimensional dialog scenario authoring tool that manages the collected dialog scenarios and visualizes them in three-dimensional space was also studied[2]. In this paper, we propose a dialogue method for continuously conducting daily conversation by adjusting similar sentence weights based on user response and selecting the user 's preferred scenario. In this continuous dialogue, we will introduce a method for improving conversation quality and a system for implementing this method.
II. Main subject
A. Overall system configuration diagram
The whole system consists of a continuous conversation system, a database server, and a machine learning server as shown below.
The continuous dialog system searches sentences and scenarios similar to the input sentences and adjusts sentence similarity weights based on user responses. The database server provides input sentences and word and sentence semantic vectors for similarity search, and provides a dialog scenario DB and a similar sentence ranking DB. The machine learning server provides a semantic vector model of a word or sentence and deduces positive and negative responses to the input sentence of the user. Then, the continuous conversation system adjusts the weights of similar sentences in the input sentence and stores the information in the database server.
Through this system, we first extract similar sentence list and ranking in the order most similar to the input sentence. In the second step, the dialogue scenario is searched in order of similar sentence, and the following sentence in the similar sentence is extracted as the system response. In the third step, positive and negative responses are deduced from the user's input sentences, and the sentence weights are adjusted to be similar to the user's input sentences so that the sentence is selected in the user's preferred scenario. In this way, the conversation continues in succession.
B. Continuous conversation progress
The continuous dialog system consists of an input / output management system, a sentence similarity retrieval system, a scenario retrieval system, and a sentence similarity weight adjustment system. The input / output management system accepts the user's utterance sentence as an input and outputs the response sentence of the system. The sentence similarity retrieval system retrieves the user 's utterance sentence, searches for the sentence most similar to the uttered sentence, and ranks it by the score.
The scenario retrieval system first searches the scenarios to which the similar sentence belongs in the similar sentence ranking order. If the similar sentence is the end of the scenario, it searches the scenario to which the similar sentence of the next rank belongs. If the similar sentence is not the end sentence of the scenario, it plays the role of extracting the next sentence following the similar sentence in response. The sentence similarity weighting system deduces whether the input sentence of the user is positive or negative or neutral. If positive, add a positive number or keep your current score. If the sentence is negative, the weight is lowered and the similar sentence rank of the input sentence is lowered, thereby guiding the user to select the preferred sentence.
The database server is composed of a word / sentence semantic vector DB, a dialog scenario DB, and a similar sentence ranking DB. The word / sentence semantic vector DB is used to convert the input sentence into semantic vector and to compute sentences similar to the input sentence based on semantic vector. The input vector may derive its own sentence semantic vector from the machine learning server or derive the semantic vector of the input sentence by summing the pre-constructed word semantic vectors. The dialogue scenario DB is composed of a dialogue scenario list composed of daily conversation and a sentence list included in the scenario. A dialogue scenario is a sequential list of conversations between two or more people on a specific topic. The similar sentence ranking DB is the DB which lists the sentences most similar to the input sentence according to the similarity value.
The machine learning server consists of a user response inference model and a three - dimensional vector model. The user response inference model assumes that the input sentence of the user is positive, negative, or neutral when the input sentence is transmitted in the continuous dialogue system, and transmits it to the continuous dialogue system. The three-dimensional vector model can derive the sentence semantic vector from the input sentence in real time or derive the semantic vector of the input sentence by summing the pre-constructed word semantic vectors for the word.
As shown in the figure below, the sentence of the user's speech or the sentence directly input is sent to the machine learning server to obtain the semantic vector of the input sentence, and the sentence most similar to the semantic vector is searched and ranked and stored in the database server. In addition, it plays a role of adjusting the weight from the response of the user.
The following figure is a description of a continuous conversation system.
When an input sentence comes in, similar sentence list and rank are extracted by semantic similarity search. First, the first sentence (sentence 23-4) is extracted and the next sentence is extracted from the scenario to which the sentence list belongs (scenario 23). If the extracted sentence is the end sentence of the scenario, the next sentence (sentence 108-3) on the scenario (scenario 108) to which the sentence (sentence 108-2) derived in the second order belongs is fetched Select the sentence. Then, the following sentence following the sentence is taken and outputted as a response sentence. In addition, positive and negative responses are deduced from the user's input sentences. If the user responds positively, we maintain the positive weight or the current similarity score, and if negative, assign the negative weight to the existing similarity score to lower the overall similarity score. Then the sentence will not be selected as a reply sentence.
C. Quality improvement performance
We checked the user response to the 56,000 Set (196,000 sentences) daily conversation scenario with a vector similarity based system for user questions. We confirmed whether the user reaction was positive or negative and weighed it so that the scenario with high preference was selected.
In case of vector response based on the vector similarity based on 100 randomly prepared questions, the conversation success rate is 35% and when the user preference is applied, the conversation success rate is increased to 69%.
III. Conclusion
Through this study, we developed a continuous conversation - based quality improvement system. Through this system, conversation scenarios in various fields that do not depend on a specific domain are collected and based on this scenario, conversations reflecting current trends according to user's time varying needs can be performed. In addition, it was possible to improve the overall conversation quality by allowing the user to prefer the conversation in the continuous conversation.
인공지능 음성, 언어, 영상 분석/처리 전문기업 bory.io/
최근댓글