TW201619850A - Question processing system and method thereof - Google Patents

Abstract

A question processing system and method thereof is provided. The question processing system comprises a question construction module with a wrong word processing unit and a mistaken word processing unit, and a question classification processing module. The wrong word processing unit detects and corrects wrong word or Martian language of natural language question to generate a correction question to conform to question intent of the natural language question. The mistaken word processing unit analyzes collocation relation of at least two phrases of the correction question, and amends mistaken word of the correction question to generate at least one candidate question to conform to the question intent. The question classification processing module analyzes the candidate question to generate question classification. Thereby, the disclosure can have a good fault-tolerant ability.

Description

Question processing system and method thereof

The disclosure relates to a question processing system and a method thereof, and in particular to a problem processing system and method thereof.

In the search engine or question answering system of the prior art, since the user does not have the ability to tolerate the question, when the user inputs a natural language question containing a typos, a Martian language, a misuse word or a missing word, It may cause the search engine or the question answering system to misjudge the intent of the natural language question and thus reply the wrong answer to the user.

Figs. 1A to 1C are diagrams showing a form of a natural language question containing a singular word of a Martian text, a sound like, and a typos of a similar word in the prior art. In the figure, the natural language questions contain many inappropriate terms (indicated by the bottom line _), such as the Martian text (such as phonetic text) shown in Figure 1A, or the typos of the sound shown in Figure 1B. Or the typos in the form shown in Figure 1C, these Martian or typos will significantly reduce the accuracy of the response to the natural language question by the search engine or question answering system.

2A to 2G are respectively shown in the prior art, which are provided with a keyword group 11, a typos 13 and a Mars in a search engine or an automatic question answering system. Page 14 or the page of the answer to the natural language question of the misuse of the word 15.

In the search engine (such as Google) in Figure 2A, the search engine can correct the keyword group 11 "Zhanling" to the keyword group 12 for the natural language question such as Chinese. And directly searched for "Japan has occupied Taiwan for a few years", but still can't find the correct answer about "years."

In the search engine (such as Google) in Figure 2B, the search engine for the natural language question such as Chinese "Yang Chuanguang is that one", and can not correct the typo 13 "that one foot", so that it can not be found The appropriate answer.

In the search engine (such as Google) in Figure 2C, the search engine can't use the Martian text 14 (such as the phonetic transcription) for the natural language question such as Chinese. The correction was made so that the correct answer could not be found.

Also in the search engine of the 2D map (such as Google), the search engine can correct the keyword group "rice plented" to "rice planted" for searching, for example, the natural language question "rice plented" in English. But still can't find the right answer.

In the search engine of FIG. 2E (such as Google), the search engine can search for the natural language phrase "who is taiwan president", for example, in English, but can directly search for keyword group 11 "whom" instead of "who". Still can't find the right answer.

In the automatic question answering system of Figure 2F (such as WorframAlpha), the automatic question answering system for natural language questions such as English "where does Rice live", and the misuse of the word "rice live" cannot be corrected, so that the wrong answer 16 is answered.

Also in the automatic question answering system of Fig. 2G (such as WorframAlpha), the automatic question answering system cannot correct the misuse word 15 "Where" for the natural language question "Where is the Taiwan President", for example, in English. Wrong answer 16.

Therefore, how to overcome the problems of the above-mentioned prior art has become a problem that is currently being solved.

The disclosure provides a question processing system and a method thereof, which can have good fault tolerance, so as to improve the correct rate of answering answers to natural language questions.

The problem processing system of the present disclosure is applied to an electronic device having a processor, a memory and an operating system, and the question processing system includes a question construction module having a typo processing unit and a misuse word processing unit, and a question Sentence category processing module. The typo processing unit detects and corrects the typos or natural language of the natural language question to generate a corrective question that matches the intent of the natural language question. The misused word processing unit analyzes the collocation relationship of at least two phrases of the correction question, and corrects the misuse word of the correction question according to the collocation relationship to generate at least one candidate question that meets the intention of the question. The question category processing module analyzes the candidate question to generate a question category of the candidate question.

The problem processing method of the present disclosure is applied to an electronic device having a processor, a memory, and an operating system, and the question processing method includes: detecting And correcting the typos or natural language of the natural language question to generate a correcting question that matches the intention of the natural language question; analyzing the collocation relationship of at least two phrases of the correcting question, and correcting according to the collocation relationship The mistype word of the correction question is used to generate at least one candidate question that meets the intent of the question; and the candidate question is analyzed to generate the question category of the candidate question.

In the above question processing system and method thereof, the missing word processing unit may analyze the missing words of the candidate question and extract at least one collocation from the corpus or the synonym/sense vocabulary to fill the missing words of the corrected question. The word produces the candidate question.

It can be seen from the above that in the problem processing system and method of the present disclosure, the typo processing unit, the misuse word processing unit and the missing word processing unit of the question construction module are mainly used to correct the typos of the natural language question respectively. , Martian text, misuse of words and missing words, and analyze the question type of the natural language question through the question category processing module.

In this way, the disclosure can have good fault tolerance, to tolerate the typos, the Martian, the misuse and the missing words of the natural language question, and to reduce the analysis error of the question of the natural language question. The impact, and thus the correct answer to the answer to the natural language question.

11, 12‧‧ ‧ keyword group

13‧‧‧ typos

14‧‧‧Martian

15‧‧‧ Misuse of words

16‧‧‧ answers

2‧‧‧Question Processing System

20‧‧‧User interface

21‧‧‧ Question Building Module

211‧‧‧ typo processing unit

211a‧‧‧ translation model

211b‧‧‧ language model

212‧‧‧Miscellaneous word processing unit

213‧‧‧ Missing word processing unit

214‧‧‧Keyword group capture unit

22‧‧‧Question class processing module

23‧‧‧ Corpus

24‧‧‧Synonyms/synonyms

25‧‧ ‧ knowledge base

26‧‧‧ paragraph search module

261‧‧ ‧ documents

262‧‧‧ paragraph

27‧‧‧Answer Processing Module

271‧‧‧ answers

41‧‧‧Natural language questions

42‧‧‧ Candidate Questions

43, 44‧‧‧ keyword group

S31 to S36‧‧‧ steps

Figures 1A to 1C are diagrams showing the forms of natural language questions containing singular words and similar typos in the prior art; and Figs. 2A to 2G are respectively shown in the drawings. Technology with search engine search or automated question answering system with keyword groups, typos, Mars a webpage of the answer to the natural language question of the word or misuse word; FIG. 3 is a block diagram showing the problem processing system of the present disclosure; and FIG. 4 is a schematic flow chart showing the processing method of the question of the present disclosure; FIG. 5 is a schematic diagram showing an embodiment of a question processing system and method thereof.

The embodiments of the present disclosure are described in the following specific embodiments, and those skilled in the art can easily understand other advantages and functions of the disclosure by the contents disclosed in the specification, and can also be implemented by other different embodiments. Or application.

FIG. 3 is a block diagram showing the problem processing system 2 of the present disclosure. As shown in the figure, the question processing system 2 can be applied to an electronic device having a processor, a memory, and an operating system, and the question processing system 2 mainly includes a user interface (UI) 20, one having The typo processing unit 211 and the question construction module 21 of the misuse word processing unit 212 and a question category processing module 22. The electronic device can be a personal computer, a tablet computer, a notebook computer, a web server, a cloud server, a mobile phone, or a smart phone.

The user interface 20 is for the user to input a natural language question, and the natural language question can be Chinese, English or various languages. The typo processing unit 211 detects and corrects the typos or the Martian text of the natural language question to generate a corrective question that matches the intent of the natural language question.

Specifically, the typo processing unit 211 can be based on the typo detection method. Or Martian translation method to detect and correct the typos or Martian text of the natural language question, and the typo processing unit 211 can have a translation model 211a and a language model 211b.

The translation model 211a provides the typos of the natural language question or the correction data of the Martian text, for example: (1) the erroneous word of the sound like or the like, and (2) the eigenvalue of the vocabulary of the phrase (like the radical, the same prefix) , the difference between the radicals, the radical strokes, the deviation of the radicals, the difference of the phonetic or the difference between the notes, and (3) the confusing vocabulary (such as 躁 vs. dry).

Moreover, the language model 211b corrects the typos or the Martian text of the natural language question according to the correction data to generate the correction question, and the language model 211b can be a statistical language model based on the n-joint word (Statistical) Language Model, SLM), or Neural Network-based Language Modeling (NNLM). The language model 211b may have a decoder to convert a Chinese character, a phonetic symbol or a complete string of phonetic symbols to the original word, the sound like/shape like word or the corresponding word of the phonetic.

The above-mentioned typos can be, for example, a typo of the sound shown in FIG. 1B of the prior art, or a typos of a similar shape as shown in FIG. 1C, and the typos can be a question word of the natural language question. Or function words, etc., the function words can be qualifiers (such as this, one, my), pronouns (such as you, me, him), prepositions / prepositions / postpositions (above, below, For), or connect words (such as and, or, if) and so on. The Martian text may be, for example, a phonetic transcription shown in FIG. 1A of the prior art, or may be an emoji (eg, *, #, !).

For example, in a natural language question such as Chinese, the typo processing unit 211 can detect the typos "she" and "Martian" of the natural language question "rice living in the house" input by the user. According to the question of the natural language question, it is intended to correct "she" and "ㄉ" to the correct words "sh" and "land" respectively. It is also about the natural language question "The rice lives in the house" Corrected to the correction question "Where does the rice live?"

The misuse word processing unit 212 analyzes the collocation relationship of at least two phrases of the correction question, and corrects the misuse word of the correction question according to the collocation relationship to generate at least one candidate question that meets the intention of the question.

In detail, the misuse word processing unit 212 analyzes whether the question intent conflicts with the context of at least two phrases of the correction question, and according to the context and from the corpus 23 or synonym/ At least one first collocation word is extracted from the thesaurus 24 to use the first collocation to correct the misuse words of the correction question to generate the candidate question, so that the context of the phrase of the candidate question does not conflict and In line with the intent of the question.

For example, the misuse word processing unit 212 analyzes the correction question phrase "Where does the rice live?", and the three phrases "rice", "residence" and "land" have a poor relationship and are conflicting because of "rice". Usually, it is not shared with "stay" in the same question, and the phrase "live" should be a misuse word, so the misuse word processing unit 212 can learn at least one first collocation according to the collocation relationship of the three phrases. "planting" or "planting" to correct the phrase "residence", and the correction question "Where does the rice live" is amended to be a candidate question in line with the intent of the question: "Where is rice planted" or "rice planting" In what place...and so on.

The question category processing module 22 analyzes the candidate question to generate a question category of the candidate question, and the question category can be person, event, time, place, object, quantity, quality, speed, altitude, Various types or types such as size.

The question construction module 21 may further include a missing word processing unit 213, which analyzes the missing words of the candidate question and extracts at least one second collocation from the corpus 23 or the synonym/sense vocabulary 24 to utilize The second collocation complements the missing word of the correction question to generate the candidate question, so that the phrase of the candidate question phrase is complete and conforms to the question intention.

For example, the missing word processing unit 213 analyzes the candidate question "Where is the rice planted" or "Where is the rice planted", and the "land" should be the meaning of "place" or "region", then the missing word The processing unit 213 extracts at least one second collocation "place" or "region" to correct the "ground" and adds a question mark "?" to supplement the missing words of the correction question to generate a complete candidate question. That is, the candidate question "Where is the rice planted" or "Where is the rice planted?" is amended to "Where is the rice planted?", "Where is the rice planted?", "Where is the rice planted?" or " Where is the rice planted?"...etc.

The at least one candidate question may be a plurality of first-priority candidate questions, and the question category processing module 22 may generate the points according to the question classification model and the knowledge base 25 to analyze the top priority candidate questions. The question category of the highest priority candidate question.

The question classification model may include hybrid approaches, regular expression rules, classifier for machine learning, and support vector machines. (support vector machine, SVM), Maximum Entropy classifier, or decision tree classifier.

The knowledge base 25 can provide information on the types of questions corresponding to the most preferred candidate questions. For example, if the highest priority candidate question is "Which of the Holy Stories is?", the knowledge base 25 provides the question. The sentence category is "person". The knowledge base 25 can also provide rules for the question categories of the top priority candidate questions, for example, if "there" is followed by "who", "who" or "what", Then, the knowledge base 25 provides the question type as "person"; or, if the word "to" is followed by "how long", the knowledge base 25 provides the question type as "hour".

The question category processing module 22 may reorder the top priority candidate questions according to the confidence scores of the top priority candidate questions, and extract more than a predetermined confidence score from the highest priority candidate questions. The threshold value and the highest confidence score are the first priority candidate questions.

For example, the question category processing module 22 can reorder the candidate questions "Where is the rice planted?", "Where is the rice planted?", "Where is the rice planted?" and "What is rice planted?" "Region?" and "Where is the rice planted?" is the first priority candidate.

The question construction module 21 may have a keyword group extraction unit 214, which generates at least one keyword group or a question construction result according to the first priority candidate question. For example, the keyword group capturing unit 214 may generate three keys from the first priority candidate question "Where is the rice planted?" The phrase "rice", "planting" and "region", or the result of a question construction "rice planting area".

The question processing system 2 can include a paragraph search module 26 and an answer processing module 27, and the paragraph search module 26 extracts a keyword group or a question from the at least one file 261 that matches the first priority candidate question. Constructing a paragraph 262 of the result, and the answer processing module 27 retrieves from the paragraph 262 an answer 271 that matches the question category of the first priority candidate question to the answer 271 (or includes the paragraph 262) Displayed in the user interface 20.

4 is a schematic flow chart showing a method for processing a question in the present disclosure, and FIG. 5 is a schematic view showing an embodiment of a question processing system 2 and a method thereof. Please refer to the question in the third figure. Processing system 2.

The problem processing method of the present disclosure can be applied to an electronic device having a processor, a memory, and an operating system, and the electronic device can be a personal computer, a tablet computer, a notebook computer, a network server, a cloud server, and an action. Phone or smart phone, etc. At the same time, the method for processing the question of the disclosure mainly includes the following steps:

(1) As shown in steps S31 and 5 of FIG. 4, in the primary school knowledge question answering system, the user first inputs the natural language question 41 "rice living in the house" from the user interface 20, and The natural language question 41 is received by the question processing system 2. Next, the process proceeds to step S32.

(2) As shown in step S32 of FIG. 4, the typo processing unit 211 of the question construction module 21 detects and corrects the typos or the Martian text of the natural language question 41 to generate a natural language question. The correction question of the intent of the sentence of sentence 41.

Specifically, the typo processing unit 211 can detect and correct the typos or Martian text of the natural language question 41 according to the typo detection method or the Martian translation method. In the meantime, the typo processing unit 211 can have a translation model 211a and a language model 211b, and the translation model 211a provides the typos of the natural language question 41 or the correction data of the Martian text, and the language model 211b is corrected according to the correction data. The natural language question 41 has a typos or a Martian text to generate the correction question.

For example, the typo processing unit 211 can detect the typos "she" and the Martian text "ㄉ" of the natural language question "rice living in the house", and according to the natural language question 41, the intent of the question will be " The "s" and "land" were corrected to the correct words "sh" and "land" respectively. The natural language question "rice living in the house" was corrected to the correct question "Where does the rice live?" Next, the process proceeds to step S33.

(3) As shown in step S33 of FIG. 4, the misuse word processing unit 212 of the question construction module 21 analyzes the collocation relationship of at least two phrases of the correction question, and corrects the correction question according to the collocation relationship. Misuse of words. At the same time, the missing word processing unit 213 of the question construction module 21 complements the missing words of the correction question to generate one or plural highest priority candidate questions.

In detail, the misuse word processing unit 212 analyzes whether the question intent is in conflict with the context of at least two phrases of the correction question, and in the event of a conflict, according to the context from the corpus 23 or the synonym/synonym 24 Extracting at least one first collocation to modify the misuse word of the correction question according to the first collocation to generate the candidate question, so that the phrase of the phrase of the candidate question phrase There is no conflict and it is consistent with the intent of the question.

For example, the misuse word processing unit 212 analyzes the correction question phrase "Where does the rice live?", and the three phrases "rice", "residence" and "land" have a poor relationship and are conflicting because of "rice". Usually, it is not shared with "stay" in the same question, and the phrase "live" should be a misuse word, so the misuse word processing unit 212 can learn at least one first match according to the collocation relationship of the three phrases. The word "planting" or "planting" to correct the phrase "live", and also correcting the question "Where does the rice live" is amended to be a candidate question that matches the intent of the question: "Where is rice planted" or "rice planting" In what place...and so on.

The missing word processing unit 213 analyzes the missing words of the candidate question, and extracts at least one second collocation from the corpus 23 or the synonym/thesaurus 24 to supplement the correction question with the second collocation. The candidate question is generated by missing the word, so that the phrase of the candidate question phrase is complete and conforms to the question intention.

For example, the missing word processing unit 213 analyzes the candidate question "Where is the rice planted" or "Where is the rice planted", and the "land" should be the meaning of "place" or "region", then the missing word The processing unit 213 extracts at least one second collocation "place" or "region" to correct the "ground" and adds a question mark "?" to supplement the missing words of the correction question to generate a complete candidate question, that is, The candidate question "Where is rice planted" or "Where is rice planted?" is amended to "Where is rice planted?", "Where is rice planted?", "Where is rice planted?" or "Rice What area is planted?"...etc. Next, the process proceeds to step S34.

(4) As shown in step S34 of FIG. 4, the question category processing module 22 analyzes the top priority candidate questions according to the question classification model and the knowledge base 25 to generate the highest priority candidate questions. Question category.

In addition, the question category processing module 22 may reorder the top priority candidate questions according to the confidence scores of the highest priority candidate questions, and extract more than the predetermined ones from the highest priority candidate questions. The confidence score threshold and the highest confidence score are used as the first priority candidate questions.

For example, the question category processing module 22 can reorder the candidate questions "Where is the rice planted?", "Where is the rice planted?", "Where is the rice planted?" and "What is rice planted?" "Region?" and "Where is the rice planted?" is the first priority candidate question, as shown in Figure 5 "I guess you want to ask "Where is the rice planted?" Candidate question 42 "What area is rice planted?" Next, the process proceeds to step S35.

(5) As shown in step S35 of FIG. 4, the keyword group capturing unit 214 of the question construction module 21 generates at least one keyword group or one sentence construction result according to the first priority candidate question. For example, from the first priority candidate question "Where is the rice planted?", the keyword group 43 "rice" and the keyword group 44 "planting" as shown in Fig. 5 are generated, or a question construction is generated. The result is "rice planting area". Next, the process proceeds to step S36.

(6) As shown in step S36 of FIG. 4, the paragraph search module 26 retrieves from the at least one file 261 a paragraph 262 that matches the keyword group or question construction result of the first priority candidate question, and Answer processing module 27 The answer 271 of the question category corresponding to the first priority candidate question is drawn in the paragraph 262 to display the answer 271 (or include the paragraph 262) on the user interface 20.

For example, from the Wikipedia document shown in Figure 5, draw the paragraphs that match the keyword group 43 "Rice" and the keyword group 44 "planting", and the answer "The rice is widely planted in mainland China and gradually spread westward. In India, the Middle Ages introduced European cadres. Currently, half of the world's population consumes rice, mainly in Asia, European cadres and tropical Americas and parts of Africa, as shown in the user interface20. Moreover, the answer can be one of the above paragraphs, or it can be a simple answer, such as "China."

Similarly, in the natural language question such as English, the question processing system 2 and the question processing method of FIG. 4 can be used in the same manner. The following is a brief description of an example.

(1) The natural language question 41 "What does rice live?" is input from the user interface 20 by the user as shown in step S31 of the third and fourth figures, and is received by the question processing system 2. Natural language question 41.

(2) As shown in step S32 of the third and fourth figures above, the typo processing unit 211 of the question construction module 21 detects and corrects the typos of the natural language question 41 "What does rice live?" Or Martian to produce a corrective question that matches the intent of the natural language question 41.

Since the typo processing unit 211 does not detect that the natural language question 41 "What does rice live?" contains a typos or a Martian text, and also conforms to the intent of the natural language question 41, it can directly The natural language question 41 is used as the correction question "What does rice live?".

(3) The error word processing unit 212 of the question construction module 21 analyzes at least two phrases "Where", "does" and "" of the correction question as shown in step S33 of the third and fourth figures above. According to the collocation relationship, the misuse word "live" of the correction question is corrected to the correct word "grown" or "planted".

At the same time, the missing word processing unit 213 of the question construction module 21 complements the missing words of the correction question to generate one or plural highest priority candidate questions. Since the correction question "What does rice live?" does not contain a missing word, the missing word processing unit 213 can directly generate one or a plurality of highest priority candidate questions, for example, the candidate question is "where where rice Grown?" and "where is rice planted?".

(4) As shown in step S34 of FIG. 3 and FIG. 4 above, the question category processing module 22 analyzes the top priority candidate questions according to the question classification model and the knowledge base 25 to generate the most The question type of the preferred candidate question, for example, the question category is "where".

In addition, the question category processing module 22 may reorder the top priority candidate questions according to the confidence scores of the highest priority candidate questions, and extract more than the predetermined ones from the highest priority candidate questions. The confidence score threshold and the highest confidence score are used as the first priority candidate question, for example, the first priority candidate question is "where does rice grown?".

(5) As shown in step S35 of FIG. 3 and FIG. 4 above, the keyword group capturing unit 214 of the question construction module 21 generates at least one keyword group or one question according to the first priority candidate question. The sentence construction result, for example, the keyword group is "where", "rice" and "grown", or the question is built The result is "where rice grown".

(6) As shown in step S36 of the third and fourth figures above, the paragraph search module 26 retrieves from the at least one file 261 the keyword group or question construction result that matches the first priority candidate question. Paragraph 262, and the answer processing module 27 retrieves an answer 271 from the paragraph 262 that matches the question category of the first priority candidate question to display the answer 271 (or include the paragraph 262) to the user. Interface 20 on.

It can be seen from the above that in the problem processing system and method of the present disclosure, the typo processing unit, the misuse word processing unit and the missing word processing unit of the question construction module are mainly used to correct the typos of the natural language question respectively. , Martian text, misuse of words and missing words, and analyze the question type of the natural language question through the question category processing module.

In this way, the disclosure can have good fault tolerance, to tolerate the typos, the Martian, the misuse and the missing words of the natural language question, and to reduce the analysis error of the question of the natural language question. The impact, and thus the correct answer to the answer to the natural language question.

The above-described embodiments are merely illustrative of the principles, features, and functions of the present disclosure, and are not intended to limit the scope of the present disclosure. Any person skilled in the art can practice the above without departing from the spirit and scope of the disclosure. The embodiment is modified and changed. Any equivalent changes and modifications made by the disclosure of the present disclosure should still be covered by the following claims. Therefore, the scope of protection of this disclosure should be as set forth in the scope of the patent application.

2‧‧‧Question Processing System

20‧‧‧User interface

21‧‧‧ Question Building Module

211‧‧‧ typo processing unit

211a‧‧‧ translation model

211b‧‧‧ language model

212‧‧‧Miscellaneous word processing unit

213‧‧‧ Missing word processing unit

214‧‧‧Keyword group capture unit

22‧‧‧Question class processing module

23‧‧‧ Corpus

24‧‧‧Synonyms/synonyms

25‧‧ ‧ knowledge base

26‧‧‧ paragraph search module

261‧‧ ‧ documents

262‧‧‧ paragraph

27‧‧‧Answer Processing Module

271‧‧‧ answers

Claims (19)

A question processing system is applied to an electronic device having a processor, a memory and an operating system. The question processing system includes: a question construction module, which has a typo processing unit that detects and corrects natural language. a typos or a Martian text of the question to generate a corrective question that matches the intent of the natural language question; and a misuse term processing unit that analyzes the collocation relationship of at least two phrases of the corrective question, and The collocation relationship corrects the misused words of the correction question to generate at least one candidate question that meets the intent of the question; and the question category processing module analyzes the candidate question to generate the question category of the candidate question. The question processing system described in claim 1 of the patent application further includes a user interface for the user to input the natural language question. The question processing system of claim 1, wherein the typo processing unit is based on a typo detection method or a Martian translation method to detect and correct a typos or a Martian text of the natural language question. The question processing system according to claim 1, wherein the typo processing unit has a translation model and a language model, and the translation model provides a typos of the natural language question or an amendment data of the Martian text. The language model corrects the typos or the Martian text of the natural language question according to the revised data to generate the correction question. The question processing system of claim 1, wherein the misuse word processing unit analyzes the context of the question intent and the phrase of the correction question, and according to the context, the corpus or synonym/synonym The candidate question is generated by extracting at least one first collocation in the library to correct the misuse word of the correction question. The question processing system according to claim 1, wherein the question construction module further has a missing word processing unit, and analyzes the missing words of the candidate question, and is in a self-corpus or synonym/synonym database. The candidate question is generated by taking at least one second collocation to complement the missing word of the correction question. The question processing system according to claim 1, wherein the at least one candidate question is a plural-first candidate question, and the question-class processing module is configured by a question classification model and a knowledge base. Analyzing the top priority candidate questions to generate the question categories of the top priority candidate questions. The question processing system of claim 7, wherein the question category processing module reorders the highest priority candidate questions according to the confidence scores of the highest priority candidate questions, and Among the top priority candidate questions, the one with the highest confidence score is taken as the first priority candidate question. The question processing system of claim 8, wherein the question construction module further has a keyword group capturing unit, and generates at least one keyword group or a question according to the first priority candidate question. Construct the results. For example, the question processing system described in claim 9 further includes a paragraph search module and an answer processing module, wherein the paragraph search module extracts a keyword group from the file that meets the first priority candidate question. Or the paragraph of the construction result, the answer processing module extracts an answer from the paragraph in the paragraph that matches the question type of the first priority candidate question. A question processing method is applied to an electronic device having a processor, a memory and an operating system, the method for processing the question comprising: detecting and correcting a typos of a natural language question or a Martian text to generate a match A correction question of the intention of the question of the natural language question; analyzing the collocation relationship of at least two phrases of the correction question, and correcting the misuse word of the correction question according to the collocation relationship to generate at least one intent to conform to the question a candidate question; and analyzing the candidate question to generate a question category for the candidate question. For example, the method for processing a question as described in claim 11 includes detecting or correcting a typos or a Martian text of the natural language question according to a typo detection method or a Martian translation method. For example, the method for processing a question as described in claim 11 further includes providing a typos of the natural language question or an amendment of the Martian text, and correcting the typos or the Martian text of the natural language question according to the revised data. To generate the correction question. The method for processing a question as described in claim 11 further includes analyzing the context of the sentence intent and the phrase of the correcting question, and extracting at least one from the corpus or the synonym/sense vocabulary according to the context. First ride The conjunction generates the candidate question by correcting the misused words of the correction question. For example, the method for processing a question as described in claim 11 further includes analyzing the missing words of the candidate question, and extracting at least one second collocation from the corpus or the synonym/sense vocabulary to complement the correction question. The candidate question is generated by missing the word. The method for processing a question according to claim 11, wherein the at least one candidate question includes a plurality of top priority candidate questions to analyze the highest priority candidates by using a question classification model and a knowledge base. The question arises from the question categories of the top priority candidate questions. The method for processing a question as described in claim 16 further includes reordering the top priority candidate questions according to the confidence scores of the top priority candidate questions, and selecting the highest priority candidate questions from the top priority candidate questions. The candidate with the highest confidence score is the first priority candidate. The method for processing a question as described in claim 17 further includes generating at least one keyword group or a question construction result according to the first priority candidate question. The method for processing a question as described in claim 18 of the patent application further includes extracting from the document a paragraph of a keyword group or a question construction result that satisfies the candidate question of the first priority, and drawing a match from the paragraph The answer to the question category of the first priority candidate question.