TWI688969B - Dialogue system for medical product recommendation - Google Patents

Abstract

A dialogue system for medical produce recommendation includes a preprocessing module, a dialog understanding module and a dialog management module. The preprocessing module used for processing an input sentence, and the intent detection unit of the dialog understanding module detects the intention of the input sentence, the slot filling unit of the dialog understanding module fills the filling slot. Finally, dialog management module determines the user's disease and gives the recommended medicine, thereby achieving disease diagnosis by artificial intelligence.

Description

Drug Selection Dialogue System

The invention relates to a dialogue system, in particular to a dialogue system for drug selection.

In recent years, Chinese people have paid more and more attention to physical health and quality of life. A large number of new Chinese and Western medicines or health foods have appeared on the market, which makes it difficult for users to choose medicines and health foods, which makes users take more medicines. Taking healthy foods to hospitals or clinics to ask medical staff, the problem of the shortage of medical staff is even more serious. Therefore, how to use AI artificial intelligence to correctly and quickly help users know what their symptoms may be corresponding to the disease, and give the user a preliminary choice of drugs or healthy food suggestions is also a very important application for AI artificial intelligence.

Please refer to Taiwan Patent Application No. 097136480 "Chronic Disease Self-Health Care Consultation System". The chronic disease self-health care consultation system gradually guides users to conduct health consultation through patterns and guide words, but due to the lack of the system In addition to making the consultation more monotonous, the dialogue between them also makes it difficult for the health consultation to quickly get close to the actual problems of the user due to the limitation of the guiding words, which causes the traditional artificial intelligence health consultation to be more time-consuming.

The main purpose of the present invention is to understand the meaning of the sentence entered by the user through the intention detection and slot filling of the dialogue understanding module, and to provide the diagnosis result and drug selection according to the meaning detected by the dialogue understanding module through the dialogue management module Suggestions.

The medicine selection dialogue system of the present invention includes a preprocessing module, a dialogue understanding module and a dialogue management module. The preprocessing module has a sentence processing unit and a word embedding unit, and the sentence processing unit has a word breaker Element and a stop word element, the word breaker element receives an input sentence, and the word breaker element is used to break the input sentence into plural words, and the stop word element is coupled to the word breaker element to receive the words Vocabulary, the stop word component is used to delete the vocabulary belonging to at least one stop word to output a processing sentence, the word embedding unit is coupled to the sentence processing unit, and the word embedding unit is used to convert the processing sentence into a sentence Vector, the dialogue understanding module is coupled to the pre-processing module, the dialogue understanding module has an intention detection unit and a slot filling unit, the intention detection unit receives the sentence vector, and the intention detection unit is used for An intention database detects an intention of the sentence vector, the slot filling unit receives the sentence vector and the intention, and the slot filling unit is used to fill the sentence vector in a filling slot according to the intention, the dialog management mode The group is coupled to the dialogue understanding module. The dialogue management module has a diagnosis unit and a drug selection unit. The diagnosis unit is coupled to the slot filling unit to diagnose the sentence vector based on the filling slot and a diagnosis symptom weight database. Corresponding to a disease, the drug selection unit is coupled to the diagnosis unit to output a recommended drug based on the disease and a drug selection database.

In the present invention, the input sentence is processed by the preprocessing module, and the intent detection unit of the dialogue understanding module can detect the intention of the input sentence and fill the slot of the dialogue understanding module The unit fills it into the filling tank, and finally, through the dialog management module, the user can judge the disease and recommend the medicine to be used, so as to achieve the effect of disease diagnosis in the manner of artificial intelligence.

Please refer to FIG. 1, which is a functional block diagram of a medicine selection dialogue system 100 according to an embodiment of the present invention. The medicine selection dialogue system 100 includes a preprocessing module 110, a dialogue understanding module 120, and a dialogue management Module 130, wherein the dialogue understanding module 120 is coupled to the preprocessing module 110, and the dialogue management module 130 is coupled to the dialogue understanding module 120.

Referring to FIG. 2, the preprocessing module 110 has a sentence processing unit 111 and a word embedding unit 112. The sentence processing unit 111 receives an input sentence. The input sentence can be selected to receive the user's voice through a microphone After conversion, or the user directly enters in text. Among them, the sentence processing unit 111 has a word breaking element 111a and a stop word element 111b, because in a sentence in Chinese, there is no space between words and words as in the foreign language grammar contains spaces, and it is difficult to distinguish words with meanings. A distinction is made, for example, "I am today", "Tian Zao", "Qi You" in "I have a headache early in the morning" does not have meaning, but the meaning of the input sentence must be extracted by the word-breaking component 111a Word. In this embodiment, the word breaker 111a receives the input sentence, and the word breaker 111a breaks the input sentence into a plurality of words. In this embodiment, the word breaker 111a is broken in a Jieba Chinese The word system 111c performs word segmentation on the input sentence, for example, "I got a bit of a headache early today". The word segmentation is "I/today/wake up early/a little/headache", which makes every word meaningful. However, other embodiments may use other word breaking systems to break the input sentence, and the invention is not limited thereto.

In addition, because these words contain many stop words that can not represent the semantic meaning of the input sentence, such as the "me", "today" and "today" and "today" in the word breaking result "me/today/early rise/a bit/headache" "Some points" are stop words that cannot represent the semantics of the input sentence, so these stop words must be deleted by the stop word element 111b, where the stop word element 111b is coupled to the word breaker element 111a Receiving the vocabulary, the stop word component 111b is used to delete the vocabulary belonging to the stop word and output a processing sentence. In this embodiment, the stop word element 111b has a stop word database 111d, the stop word database 111d stores a plurality of the stop words, and the stop word element 111b is used to compare the input The vocabulary of the sentence and the stop words are used to delete the words belonging to the stop words.

Please refer to Figure 2. Because computer systems are not as capable of recognizing the meaning of these words as the human brain, comparing words with words one by one is quite time-consuming, and the meaning of each word may be expressed in different words, resulting in direct The comparison by text may face the problem of non-convergence in calculating the similarity. Therefore, in this embodiment, the word embedding unit 112 is coupled to the sentence processing unit 111 to convert the processed sentence through the word embedding unit 112 Is a sentence vector. In this embodiment, the word embedding unit 112 has a word2vec module 112a, and the word2vec module 112a is first trained with a corpus, which allows words with similar meaning to be represented by similar vectors. The word embedding unit 112 can convert the processed sentence into the sentence vector by the word2vec module 112a.

Referring to FIGS. 1 and 3, the dialogue understanding module 120 has an intent detection unit 121 and a slot filling unit 122, the intent detection unit 121 receives the sentence vector, and the intent detection unit 121 is used according to an intent The database 121a detects an intent of the sentence vector. In this embodiment, the intent database 121a stores a plurality of intent vocabularies corresponding to a symptom diagnosis intent and a drug selection intent. The measuring unit 121 calculates a similarity between the sentence vector and each of the intention vocabulary by Cosine similarity to measure the intention of the sentence vector as the symptom diagnosis intention or the drug selection intention.

Referring to FIG. 3, the slot filling unit 122 receives the statement vector and the intention, and the slot filling unit 122 is used to fill the statement vector into a filling slot 122a according to the intention of the statement vector. In this embodiment In this, the filling slot 122a includes a symptom slot 122b and a medicine slot 122c. If the intention of the sentence vector is the symptom diagnosis intention, which means that the sentence vectors belong to the vocabulary of the symptom, the slot filling unit 122 uses the sentence vector Filled in the symptom slot 122b, if the intention of the sentence vector is the drug selection intention, which means that the sentence vectors belong to the vocabulary of drug selection, the slot filling unit 122 fills the sentence vector in the medicine slot 122c to improve The subsequent calculation speed of the disease diagnosis and drug selection of the dialog management module 130.

Please refer to FIGS. 1 and 4, the dialogue management module 130 has a diagnosis unit 131 and a drug selection unit 132, the diagnosis unit 131 is coupled to the slot filling unit 122, in this embodiment, the diagnosis unit 131 has a The first judgment element 131a, a score calculation element 131b, a second judgment element 131c, an interrogation element 131d, and a diagnosis symptom weight database 131e.

Please refer to FIG. 4, the first determining element 131a is coupled to the slot filling unit 122 to determine whether the number of filled slots in the symptom slot 122b is greater than a slot threshold. In this embodiment, the first determination Element 131a determines whether the number of filled slots in the symptom slot 122b is greater than 2, that is, determines whether there are enough symptoms in the symptom slot 122b for disease analysis, and if so, the score calculation element 131b calculates the disease score If it is otherwise to the questioning element 131d, let the questioning element 131d ask the user more about the symptoms.

The score calculation element 131b is coupled to the first judgment element 131a and the diagnosis symptom weight database 131e to calculate a disease score, wherein the diagnosis symptom weight database 131e stores a plurality of diseases, each of which corresponds to multiple Symptoms, therefore, the score calculation component 131b can calculate the disease score for each disease based on the sentence vector filled in the symptom slot 122b. The second judging element 131c is coupled to the score calculating element 131b for judging whether each disease score is greater than a score threshold, and if one of the disease scores is greater than the score threshold, it means that the symptoms of the disease appear in use It is enough for the patient, so the second judgment element 131c can judge that the diagnosis of the disease is completed, and output the disease greater than the score threshold to the drug selection unit 132, otherwise, to the interrogation element 131d, let the interrogation element 131d Ask the user more about the symptoms.

The interrogation element 131d is coupled to the first judgment element 131a, the score calculation element 131b, and the second judgment element 131c. The first judgment element 131a judges that the number of filled slots is less than or equal to the slot threshold and the first When the judging component 131c judges that the disease score is less than or equal to the score threshold, the interrogation component 131d outputs an interrogation signal to ask the user for more descriptions of the symptoms of the disease by voice or text, and reply to the user After the symptom is vectorized, it is filled in the symptom slot 122b, and the disease score is recalculated by the score calculation component 131b.

Referring to FIG. 4, the drug selection unit 132 is coupled to the second judgment unit 131c of the diagnosis unit 131 to receive the disease, and to output a recommended drug based on the disease and a drug selection database 132a.

In the present invention, the input sentence is processed by the preprocessing module 110, and the intention detection unit 121 of the dialogue understanding module 120 can detect the intention of the input sentence and the dialogue understanding module 120 The tank filling unit 122 fills it into the filling tank 122a, and finally can determine the user's disease through the dialogue management module 130 and give the recommended medicine, so as to achieve the effect of disease diagnosis in the manner of artificial intelligence .

The scope of protection of the present invention shall be subject to the scope defined in the attached patent application. Any changes and modifications made by those who are familiar with this skill without departing from the spirit and scope of the present invention shall fall within the scope of protection of the present invention. .

100: Drug selection dialogue system

110: pre-processing module

111: sentence processing unit

111a: word breaking component

111b: stop word component

111c: Jieba Chinese word breaker system

111d: Stopword database

112: word embedding unit

112a: word2vec module

120: Dialogue Understanding Module

121: Intent detection unit

121a: Intent database

122: slot filling unit

122a: filling tank

122b: symptom slot

122c: Medicine tank

130: Dialogue management module

131: Diagnostic unit

131a: the first judgment element

131b: Score calculation component

131c: Second judgment element

131d: Interrogation element

131e: Diagnosis symptom weight database

132: Drug selection unit

132a: Drug selection database

Figure 1: According to one embodiment of the present invention, a functional block diagram of a drug dialogue system. Figure 2: A functional block diagram of a pre-processing module according to an embodiment of the invention. Figure 3: According to an embodiment of the present invention, a functional block diagram of a dialogue understanding module. Figure 4: According to an embodiment of the present invention, a functional block diagram of a dialog management module.

100: Drug selection dialogue system

110: pre-processing module

120: Dialogue Understanding Module

130: Dialogue management module

Claims (8)

A medicine selection dialogue system, including: a preprocessing module with a sentence processing unit and a word embedding unit, the sentence processing unit has a word breaking component and a stop word component, the word breaking component receives an input sentence And the word breaker element is used to break the input sentence into a plurality of words, the stopword element is coupled to the wordbreaker element to receive the words, and the stopword element is used to belong to at least one stopword The vocabulary is deleted and a processing sentence is output. The word embedding unit is coupled to the sentence processing unit. The word embedding unit is used to convert the processing sentence into a sentence vector; a dialogue understanding module is coupled to the preprocessing module, The dialogue understanding module has an intent detection unit and a slot filling unit. The intent detection unit receives the sentence vector. The intent detection unit is used to detect an intent of the sentence vector according to an intent database. The slot The filling unit receives the statement vector and the intention, and the slot filling unit is used to fill the statement vector in a filling slot according to the intention; and a dialogue management module, coupled to the dialogue understanding module, the dialogue management module The group has a diagnosis unit and a medicine selection unit. The diagnosis unit is coupled to the slot filling unit to diagnose a disease corresponding to the sentence vector according to the filling slot and a diagnosis symptom weight database. The medicine selection unit is coupled to the diagnosis A unit for outputting a suggested medicine according to the disease and a medicine selection database, wherein the diagnosis unit has a first judgment element, a score calculation element, a second judgment element and an interrogation element, the first judgment element is coupled The slot filling unit is used to determine whether the number of filled slots in the filling slot is greater than a slot threshold, the score calculation element is coupled to the first determination element and the diagnostic symptom weight database to calculate a disease score, The second judgment element is coupled to the score calculation element for judging whether the disease score is greater than a score threshold, and the second judgment element outputs the disease, and the questioning element is coupled to the first judgment element and the score calculation element And the second judgment element, the interrogation element is used to output an interrogation signal. According to the drug selection dialogue system described in item 1 of the patent application scope, wherein the word segmentation component is to use a Jieba Chinese word segmentation system to segment the input sentence. The drug selection dialogue system according to item 1 of the patent application scope, wherein the stop word component has a stop word database, the stop word database stores a plurality of the stop words, and the stop word component It is used to compare the vocabulary of the input sentence with each stop word to delete the word belonging to each stop word. The drug selection dialogue system according to item 1 of the patent application scope, wherein the word embedding unit is trained with a corpus by a word2vec module, and the word embedding unit converts the processing sentence into the sentence by the word2vec module vector. According to the drug selection dialogue system described in item 1 of the patent application scope, wherein the intent database stores a plurality of intent vocabularies, the intent detection unit calculates a similarity between the sentence vector and each of the intent vocabularies to measure Get the intention of the statement vector. According to the drug selection dialogue system described in item 5 of the patent application scope, wherein the intent vocabulary corresponds to a symptom diagnosis intent and a drug selection intent, the filled slot contains a symptom slot and a drug slot, if the intent of the sentence vector For the symptom diagnosis intention, the slot filling unit fills the sentence vector in the symptom slot, and if the intention of the sentence vector is the drug selection intention, the slot filling unit fills the sentence vector in the medicine slot. According to the drug selection dialogue system described in item 5 or 6 of the patent application scope, wherein the intent detection unit calculates the similarity between the sentence vector and each intent vocabulary by Cosine similarity. According to the medicine selection dialogue system described in item 1 of the patent application scope, the medicine selection unit is coupled to the second judgment unit to receive the disease.

Images