CN114330294B - A method for extracting character speech based on text syntactic analysis - Google Patents

Abstract

The present invention provides a method for extracting character speech based on text syntactic analysis, which extracts character speech appearing in the text based on provided character information and text information. It includes constructing a trigger word dictionary, that is, for the initial speech trigger words, using synonym technology to expand the trigger words and construct a complete trigger word dictionary; text sentence segmentation, that is, segmenting the entire text into complete sentences; sentence filtering, screening candidate sentences based on whether the provided character information and trigger word information are hit in the sentence; speech determination, that is, judging the relationship between the character and the trigger word based on syntactic analysis, and judging whether the sentence is the character speech. The character speech extraction method of the present invention can extract character speech information in the text simply, efficiently and accurately.

Description

A method for extracting character speech based on text syntactic analysis

Technical Field

The present invention relates to the technical field of text information extraction, and in particular to a method for extracting character speech based on text syntactic analysis.

Background Art

With the evolution of the times, the economy, society, production, and life are increasingly dependent on the Internet, and obtaining information through the Internet has become an indispensable part of people's daily life and work. Faced with massive amounts of information and lengthy reports, how to quickly and effectively obtain the speeches of the main characters involved has become an indispensable requirement for improving reading efficiency. Therefore, research on character speech extraction is of great significance.

At present, there are mainly rule-based methods and machine learning and deep learning methods for character speech extraction. Among them, rule-based methods mostly use trigger words, but there are few in-depth studies on the construction of trigger word dictionaries. At the same time, most of these methods do not pay attention to the relationship between trigger words and character words in sentences, resulting in reduced extraction accuracy. For machine learning and deep learning methods, although they are not constrained by trigger dictionaries, the early training and annotation require a lot of human resources, and when the target text and training samples deviate greatly, the effect is often mediocre. At the same time, deep learning methods often have higher algorithm complexity, and in actual operation, they rely on better hardware resources as a basis.

Summary of the invention

Purpose of the invention: The main purpose of the present invention is to solve the defects of the prior art and to invent a rule-based method that can automatically expand trigger words and introduce syntactic analysis to judge the relationship between trigger words and characters as a supplement, ultimately providing an accurate and efficient method for extracting character speech.

The present invention provides a method for extracting character speech based on text syntactic analysis, comprising the following steps:

Step S1, constructing a speech trigger word dictionary: for the initial speech trigger words, using synonym technology, expand the trigger words and construct a trigger word dictionary;

Step S2, text sentence segmentation: segment the entire text into complete sentences;

Step S3, sentence filtering;

Step S4, speech determination.

In step S1, the trigger word dictionary is constructed by using the initial trigger word list L: [W ₁ ,W ₂ ,…,W _n-1 ,W _n ], where W ₁ ,W ₂ ,…,W _n-1 ,W _n correspond to the first, second, third,…, nth initial trigger words respectively; the initial trigger words are speech trigger words obtained by preliminary screening in the news public opinion data. The number of initial trigger words is within 20, including: "say", "express", "tell", "point out", "reveal", "frankly", "statement" and other organized words. The initial speech trigger words are expanded by a variety of synonyms.

In step S1, the expansion through multiple synonym expansion methods includes an expansion method based on the synonym dictionary search for synonyms and an expansion method based on the word vector word2vec search for synonyms. The training corpus used by word2vec is the self-annotated news business data in the field of public opinion. Therefore, it is more suitable for the expansion of synonyms corresponding to the trigger words of character speech.

For the first initial trigger word W ₁ , the following steps are specifically included:

Step a1, taking W ₁ as input, searching for synonyms of W ₁ through the synonym dictionary, and returning a synonym set L ₁ of W ₁ , where the calculation formula of L ₁ is:

L ₁ ={W ₁ ⁱ |sim(W ₁ ,W ₁ ⁱ )>0.6}

In order to unify the expression, a list is used instead of a set, and L ₁ is recorded as: [W ₁ ¹ , W ₁ ² , W ₁ ³ , W ₁ ⁴ , … , W ₁ ^k ], W ₁ ⁱ represents the i-th synonym of W ₁ searched through the synonym dictionary, where: the minimum value of k here is set to 20, and the maximum value corresponds to the number of synonyms of W ₁ with a similarity greater than 60% to W _1. In particular, when the number of synonyms of W ₁ with a similarity threshold greater than 60% to W ₁ is less than 20, the synonyms of W ₁ are selected from the highest to the lowest similarity to make up 20.

Step a2, taking W ₁ as input, searching for synonyms of W ₁ through word2vec, and returning the synonym set L' ₁ of W ₁ : where the calculation formula of L' ₁ is:

L' ₁ ={W ₁ ⁱ |sim _word2vec (W ₁ ,W ₁ ⁱ )>0.6}

For a unified representation, a list is used instead of a set, and L' ₁ is recorded as: [W' ₁ ¹ , W' ₁ ² , W' ₁ ³ , W' ₁ ⁴ , ..., W' ₁ ^k ], W' ₁ ⁱ represents the i-th synonym of W ₁ searched by word2vec, where: The minimum value of k is set to 20, and the maximum value corresponds to the number of synonyms of W ₁ searched by word2vec with a similarity greater than 60% to W _1. In particular, when the number of synonyms of W ₁ searched by word2vec with a similarity threshold greater than 60% is less than 20, the synonyms of W ₁ are selected from the highest to _the lowest similarity to make up 20.

Step a3, sequentially perform the operation of step a2 on each word in the list _L1 obtained in step a1, and obtain a synonym list _{L1_total} corresponding to all words in the list _L1 ;

Step a4, performing the operation of step a1 on the words in the list L' ₁ obtained in step a2, and obtaining a synonym list L' _{1_total} corresponding to all the words in the list L'₁;

Step a5, merge and remove duplicates from L ₁ , L' ₁ , L _{1_total} and L' _{1_total} to obtain the candidate word library of W ₁ , and further screen to finally obtain all the synonymous trigger words corresponding to the trigger word W ₁ ;

For the trigger words in the initial trigger word list L, perform the operations of step a1 to step a5 to obtain all the synonymous trigger words corresponding to W ₁ , W ₂ , …, W n _-1 , W _n. Finally, all the trigger words corresponding to W ₁ , W ₂ , …, W _n-1 , W _n are merged and deduplicated to construct a trigger word dictionary.

In step S2, the text is segmented by first locating the position of the sentence separator to obtain ordered separator position list information, and then performing preliminary sentence segmentation according to the position list information; wherein the sentence separators are located by punctuation marks, line breaks, and spaces.

In step S2, a double quotation mark recognition method is used to determine whether the delimiter can be used as an actual delimiter.

In step S2, a double quotation mark recognition method is used to determine whether the delimiter can be used as an actual delimiter, specifically comprising: preliminarily separating the text according to the delimiter to obtain an ordered delimiter position list information, recorded as P: [x ₁ ,x ₂ ,x ₃ ,…,x _m-1 ,x _m ], wherein x ₁ ,x 2 ,x _{3 ,} …,x _m-1 ,x _m respectively represent the positions of the delimiters in the sentence, [1,x ₁ ] correspond to the start and end positions of the first sentence, and the delimiter position list is traversed from left to right. For the first traversal, the position where the start L of the sentence is fixed to 1, and the position corresponding to the end R of the sentence is x ₁ . The start position L and the end position R constitute a pre-sentence S, and determine whether the double quotation marks in the pre-sentence S meet the following conditions:

1) Whether the number of double quotes is even;

2) Whether the first one is a left quotation mark and the last one is a right quotation mark;

If both of the above conditions are met, S is output as a complete sentence. In the next traversal, the end of this sentence x ₁ + 1 is taken as the beginning of the sentence L, and the next element x ₂ of x ₁ is taken as the end of the sentence R, forming a pre-sentence S, and then performing the same judgment operation.

Step S3 includes: segmenting each sentence, determining whether the segmented sentence contains known character information and trigger word information, and if so, taking the sentence as a candidate speech statement.

In step S3, the word segmentation of each sentence specifically includes: analyzing and extracting the part of speech and the number of occurrences of each term through a large-scale corpus to construct a word segmentation dictionary for scanning the word graph of sentence segmentation, and generating a directed acyclic graph consisting of possible word formation situations of all Chinese characters in the sentence; for the directed acyclic graph, the initial word segmentation probability of each word segmentation situation is obtained according to the probability of occurrence of each term, and then the word segmentation penalty factor C is introduced to obtain the final word segmentation probability, and the maximum value of the final word segmentation probability is selected as the word segmentation result.

In step S3, the initial word segmentation probability _Pini is calculated using the following formula:

P _{ini_i} =P( _{wi_1} )×P( _{wi_2} )×…×P( _{wi_n} )

Wherein, _{wi_n} represents the nth word in the i-th possible word segmentation case, and P( _{wi_n} ) represents the word segmentation probability of the nth word _{wi_n} ;

For example, the sentence to be segmented is "playing at Nanjing University", which has multiple paths, that is, multiple segmentation results, as follows:

Path 1: Playing at Nanjing University

Path 2: Studying and playing at Nanjing University

Path 3: Play at Nanjing University

Path 4: Play at /Nanjing University/

The initial word segmentation probabilities corresponding to paths 1-4 are:

_{Pini_1} = P(in)×P(Nanjing)×P(Beijing)×P(university)×P(school)×P(play)

_{Pini_2} = P(in)×P(Nanjing)×P(university)×P(school)×P(play)

_{Pini_3} = P(in)×P(Nanjing)×P(university)×P(play)

_{Pini_4} = P(at)×P(Nanjing University)×P(play)

The word segmentation dictionary includes a public word segmentation dictionary and a domain word segmentation dictionary; the public word segmentation dictionary meets the word segmentation requirements in most situations. The public word segmentation dictionary contains the word frequency and part of speech of nearly 400,000 words. This part of the words is mainly obtained after public prediction training composed of news corpus, WeChat corpus, forum corpus, etc., and then manually reviewed. The domain word segmentation dictionary focuses more on the relevant vocabulary of specific people, places, institutions and proprietary fields.

The final word segmentation probability P _final is calculated as follows:

P _{final_i} = C ⁿ × _{Pini_i}

Where P _final _ _i is the final word segmentation probability under the i-th possible word segmentation case; C takes a value between [0.5, 0.9], and is generally selected as 0.8; n is the number of domain words that have not been successfully segmented in the sentence; for example, for the sentence to be segmented "playing at Nanjing University", if "Nanjing University" is a domain word, the penalty factor C is selected as 0.5, then for the above four word segmentation paths, the final word segmentation probabilities are:

P _{final_1} = 0.5 ¹ × _{Pini_1}

P _{final_2} = 0.5 ¹ × _{Pini_2}

P _{final_3} = 0.5 ¹ × _{Pini_3}

P _{final_4} = 0.5 ⁰ × _{Pini_4}

Finally, the segmentation path corresponding to the largest P _{final_i} is selected as the final segmentation result.

Step S4 includes: performing syntactic analysis on the candidate speech sentence to determine whether there is a subject-predicate relationship between the character and the trigger word; if so, determining that the sentence is the character's speech.

Beneficial effect: The character speech extraction method of the present invention can extract character speech from multiple steps such as comprehensive trigger word construction, accurate sentence segmentation, fast sentence filtering and reasonable speech judgment, thereby providing an accurate and efficient character speech extraction method.

BRIEF DESCRIPTION OF THE DRAWINGS

The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and/or other advantages of the present invention will become more clear.

FIG1 is a block diagram of the overall process of character speech extraction of the present invention.

FIG. 2 is a flowchart of constructing a trigger word dictionary of the invention.

FIG. 3 is a flow chart of text sentence segmentation of the present invention.

FIG. 4 is a flow chart of sentence filtering of the present invention.

FIG5 is a flow chart of speech determination of the present invention.

DETAILED DESCRIPTION

As shown in FIG1 , it is a flowchart of the overall process of a method for extracting character speech based on text syntactic analysis according to an embodiment of the present invention, including building a trigger word dictionary, text sentence segmentation, sentence filtering, and speech judgment.

As shown in Figure 2, the steps of constructing a trigger word dictionary in an embodiment of the present invention are as follows. The trigger words are expanded by using two expansion methods, namely, synonym forest search and word2vec search. The specific steps are as follows: using the initial trigger words as input data for synonym forest search and word2vec search, respectively, to obtain two sets of candidate word sets, and then using the obtained candidate word sets as input for another expansion method, and finally obtaining more candidate word sets, and finally constructing a trigger word dictionary through manual judgment.

As shown in FIG. 3 , the text sentence segmentation step of the embodiment of the present invention uses punctuation marks, spaces, line breaks and other information as sentence separators, and matches to obtain separator position list information, which is recorded as:

P＝[x ₁ ,x ₂ ,x ₃ ,…,x _m-1 ,x _m ]

Traverse the separator position list P from left to right. For the first traversal, the position where the sentence starts L is fixed to 1, and the position where the sentence ends R corresponds to is x ₁ . The starting position L and the ending position R constitute the pre-sentence S.

Determine whether the double quotes in the pre-clause S satisfy:

1) Check whether the number of double quotes is even, ensuring that double quotes appear in pairs;

2) Whether the first one is a left quotation mark and the last one is a right quotation mark;

If all conditions are met, S is output as a complete sentence. In the next traversal, the next position x ₁ + ₁ of the end position x 1 of this sentence is used as the start position L of the sentence, and the next element x ₂ of x ₁ in the separator list P is used as the end position R of the sentence. The start position L and the end position R constitute a pre-sentence S, and then the same judgment operation is performed.

If any one of them is not satisfied, S cannot be output as a complete sentence. The next element _x2 of _x1 in the separator list P is selected as the end R of the sentence to form a pre-sentence S, and then the same judgment operation is performed.

Repeat the process until the traversal is completed.

As shown in Figure 4, this is the sentence filtering flow chart of the present invention. The sentence list obtained by text segmentation is traversed, each sentence is segmented, and it is determined whether the segmented sentences contain known character information and trigger word information. If both are contained, they are used as candidate speech sentences.

As shown in Figure 5, this is the speech judgment flow chart of the present invention. The sentence list obtained after filtering the text sentences is traversed, each sentence is segmented and the segmentation results are subjected to syntactic analysis to determine whether there is a subject-predicate relationship between the characters and the trigger words. If so, the sentence is determined to be a speech sentence.

Example

This embodiment provides a method for extracting character speech based on text syntactic analysis, comprising the following steps:

Step S1, constructing a speech trigger word dictionary: for the initial speech trigger words, using synonym technology, expand the trigger words and construct a trigger word dictionary;

Step S2, text sentence segmentation: segment the entire text into complete sentences;

Step S3, sentence filtering;

Step S4, speech determination.

In step S1, the trigger word dictionary is constructed by adopting an initial trigger word list L: [ _W1 , _W2 , ..., Wn _-1 , _Wn ], which is expanded by a variety of synonyms, wherein _W1 , _W2 , ..., Wn _-1 , _Wn are initial speech trigger words manually obtained from news public opinion data, such as "say", "express", "tell", "point out", "reveal", "frankly", "statement" and other words.

In step S1, the expansion by multiple synonym expansion methods includes an expansion method based on the synonym search Cilin and an expansion method based on the word vector word2vec. For the initial trigger word W ₁ , the following steps are specifically included:

Step a1, taking W ₁ as input, searching for synonyms of W ₁ through the synonym dictionary, and returning a synonym list L ₁ of W ₁ : [W ₁ ¹ , W ₁ ² , W ₁ ³ , W ₁ ⁴ , … , W ₁ ^k ];

Step a2, taking W ₁ as input, searching for synonyms of W ₁ through word2vec, and returning a synonym list L' ₁ of W ₁ : [W' ₁ ¹ , W' ₁ ² , W' ₁ ³ , W' ₁ ⁴ , ..., W' ₁ ^k ], wherein the training corpus used by word2vec is the self-annotated news business data in the field of public opinion, and therefore, it is more suitable for the expansion of synonyms corresponding to trigger words of character speech;

Step a3, sequentially perform the operation of step a2 on each word in the list _L1 obtained in step a1, and obtain a synonym list _{L1_total} corresponding to all words in the list _L1 ;

Step a4, performing the operation of step a1 on the words in the list L' ₁ obtained in step a2, and obtaining a synonym list L' _{1_total} corresponding to all the words in the list L'₁;

Step a5, merge and remove duplicates from L ₁ , L' ₁ , L _{1_total} and L' _{1_total} to obtain the candidate word library of W ₁ , and further perform manual screening to finally obtain all the synonymous trigger words corresponding to the trigger word W ₁ ;

For the trigger words in the initial trigger word list L, perform the operations of step a1 to step a5 to obtain all the synonymous trigger words corresponding to W ₁ , W ₂ , …, W n _-1 , W _n. Finally, all the trigger words corresponding to W ₁ , W ₂ , …, W _n-1 , W _n are merged and deduplicated to construct a trigger word dictionary.

In step S2, the text sentence segmentation adopts the method of first locating the position of the sentence separator, obtaining the ordered separator position list information, and then performing preliminary sentence segmentation according to the position list information; wherein the sentence separator is located by punctuation, line break, and space, and its set is as follows: {. ? ; ! ! .? \n\r\r\n......space～}. As shown in the following paragraph.

"As far as I know, the national coverage rate should reach 40% by the beginning of July and 80% by the end of this year. Based on the 70% vaccine protection rate, China's COVID-19 vaccine coverage rate needs to reach nearly 80% to form herd immunity," Zhong Nanshan told China News Service. Based on the principle of voluntariness, all Chinese citizens aged 18 to 60 who meet the physical conditions can receive the COVID-19 vaccine free of charge. Residents A and B are preparing to receive the vaccine. There are two large hospitals and two community health service centers near their residence and workplace where they can receive the vaccine free of charge.

After the above paragraph is separated into sentences by delimiters, the resulting sentence list is:

(1) “As far as I know, by early July, the national coverage rate should reach 40%, and by the end of this year it should reach 80%.

(2) Based on the calculation that the vaccine protection rate reaches 70%, China’s new crown vaccine coverage rate needs to reach nearly 80% before herd immunity can be achieved.

(3)” Zhong Nanshan told China News Service.

(4) Based on the principle of voluntariness, all Chinese citizens aged 18 to 60 who meet the physical conditions can receive the COVID-19 vaccine free of charge. Residents A and B are preparing to receive the vaccine. There are two large hospitals and two community health service centers near their residence and workplace where they can receive the vaccine free of charge.

Double quotes are used to represent the speech of a person, but the delimiter in the double quotes cannot represent the end of the sentence. Therefore, a double quote recognition method is used to determine whether it is a delimiter in the double quotes. If it is, the sentence is not separated at the delimiter. For example, in the above example sentence, the first and second periods appear inside the double quotes, so they are not used as delimiters. The third period appears outside the double quotes, so it can be used as a delimiter.

In step S2, a double quote recognition method is used to determine whether the delimiter can be used as an actual delimiter. The specific implementation is: the text is initially separated according to the delimiter to obtain an ordered list of delimiter positions, recorded as P: [x ₁ ,x ₂ ,x ₃ ,…,x _m-1 ,x _m ], wherein x ₁ ,x ₂ ,x ₃ ,…,x _m-1 ,x _m respectively represent the positions where the delimiters appear in the sentence. Therefore, [1,x ₁ ] corresponds to the start and end positions of the first candidate sentence, and [x ₁ +1,x ₂ ] corresponds to the start and end positions of the second candidate sentence. Traverse the delimiter position list from left to right. For the first traversal, select 1 as the start position L of the sentence, and select x ₁ in the delimiter position list information as the end R of the sentence to form a pre-sentence S, and determine whether the double quotes in the pre-sentence S satisfy:

1) Whether the number of double quotes is even;

2) Whether the first one is a left quotation mark and the last one is a right quotation mark;

If both of the above conditions are met, S is output as a complete sentence. In the next traversal, the next position x ₁ + 1 of the end position x ₁ of this sentence is used as the start position L of the sentence, and the next element x ₂ of x ₁ in the separator list P is used as the end position R of the sentence to form the next pre-sentence S, and then the same judgment operation is performed;

Take the following paragraph as an example.

"As far as I know, the national coverage rate should reach 40% by the beginning of July and 80% by the end of this year. Based on the 70% vaccine protection rate, China's COVID-19 vaccine coverage rate needs to reach nearly 80% to form herd immunity," Zhong Nanshan told China News Service. Based on the principle of voluntariness, all Chinese citizens aged 18 to 60 who meet the physical conditions can receive the COVID-19 vaccine free of charge. Residents A and B are preparing to receive the vaccine. There are two large hospitals and two community health service centers near their residence and workplace where they can receive the vaccine free of charge.

If the method of the present invention is not adopted, the sentence results are:

(1) “As far as I know, by early July, the national coverage rate should reach 40%, and by the end of this year it should reach 80%.

(2) Based on the calculation that the vaccine protection rate reaches 70%, China’s new crown vaccine coverage rate needs to reach nearly 80% before herd immunity can be achieved.

(3)” Zhong Nanshan told China News Service.

(4) Based on the principle of voluntariness, all Chinese citizens aged 18 to 60 who meet the physical conditions can receive the COVID-19 vaccine free of charge. Residents A and B are preparing to receive the vaccine. There are two large hospitals and two community health service centers near their residence and workplace where they can receive the vaccine free of charge.

If the method of the present invention is adopted, the sentence results are:

(1) "As far as I know, the national coverage rate should reach 40% by the beginning of July and 80% by the end of this year. Based on the 70% vaccine protection rate, China's COVID-19 vaccine coverage rate needs to reach nearly 80% to form herd immunity," Zhong Nanshan told China News Service.

(2) Based on the principle of voluntariness, all Chinese citizens aged 18 to 60 who meet the physical requirements can receive the COVID-19 vaccine free of charge. Residents A and B are preparing to receive the vaccine. There are two large hospitals and two community health service centers near their residence and workplace where they can receive the vaccine free of charge.

Step S3 includes: segmenting each sentence, determining whether the segmented sentence contains known character information and trigger word information, and if both are contained, taking the sentence as a candidate speech statement. The segmentation adopts Jieba segmentation, and on the basis of Jieba segmentation technology, introduces custom vocabulary information to improve the accuracy of segmentation in business scenarios. Custom vocabulary information includes trigger words, key figures in the public opinion field, keywords in the public opinion field and other vocabulary information.

Step S4 includes: using syntactic analysis on the candidate speech sentence to determine whether there is a subject-predicate relationship between the character and the trigger word. If there is a subject-predicate relationship, the sentence is determined to be a character speech. In the following sentence (1), "Zhong Nanshan" is used as a character word, and "Say" is used as a trigger word. The relationship between the two is a subject-predicate relationship in syntactic analysis, so the sentence is determined to be a speech sentence. In sentence (2), although the character "Zhong Nanshan" and the trigger word "Say" are also hit, the two are not in a subject-predicate relationship, so the sentence is determined to be a non-speech sentence.

(1) "As far as I know, by the beginning of July, the national coverage should reach 40%, and by the end of this year it should reach 80%. Based on the 70% vaccine protection rate, China's COVID-19 vaccine coverage needs to reach nearly 80% to achieve herd immunity," Zhong Nanshan told China News Service. (2) Recently, Academician Zhong Nanshan said something very lovely in an interview, which was also talked about by many people. This 84-year-old man also has a very lovely side.

The present invention provides a method for extracting character speech based on text syntactic analysis. There are many methods and ways to implement the technical solution. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the protection scope of the present invention. All components not specified in this embodiment can be implemented by existing technologies.

Claims (7)

1. A method for extracting character speech based on text syntactic analysis, characterized in that it comprises the following steps: Step S1, constructing a speech trigger word dictionary: for the initial speech trigger words, using synonym technology, expand the trigger words and construct a trigger word dictionary; Step S2, text sentence segmentation: segment the entire text into complete sentences; Step S3, sentence filtering; Step S4, speech determination; In step S1, the trigger word dictionary is constructed by using an initial trigger word list L: [W ₁ ,W ₂ ,…,W _n-1 ,W _n ], where W ₁ ,W ₂ ,…,W _n-1 ,W _n correspond to the first, second, third,…, nth initial trigger words respectively; the initial trigger words are speech trigger words obtained by preliminary screening in news public opinion data; In step S1, the synonyms are expanded by a variety of expansion methods, including an expansion method based on the synonym dictionary search for synonyms and an expansion method based on word vector word2vec search for synonyms; For the first initial trigger word W ₁ , the following steps are specifically included: Step a1, taking W ₁ as input, searching for synonyms of W ₁ through the synonym dictionary, and returning a synonym set L ₁ of W ₁ , where the calculation formula of L ₁ is: L ₁ ={W ₁ ⁱ |sim(W ₁ ,W ₁ ⁱ )>0.6} To unify the expression, a list is used instead of a set, and L ₁ is denoted as: [W ₁ ¹ , W ₁ ² , W ₁ ³ , W ₁ ⁴ , … , W ₁ ^k ], where W ₁ ⁱ represents the i-th synonym of W ₁ searched through the synonym dictionary; Step a2, taking W ₁ as input, searching for synonyms of W ₁ through word2vec, and returning the synonym set L' ₁ of W ₁ : where the calculation formula of L' ₁ is: L' ₁ ={W ₁ ⁱ |sim _word2vec (W ₁ ,W ₁ ⁱ )>0.6} To unify the expression, a list is used instead of a set, and L' ₁ is denoted as: [W' ₁ ¹ , W' ₁ ² , W' ₁ ³ , W' ₁ ⁴ , … , W' ₁ ^k ], where W' ₁ ⁱ represents the i-th synonym of W ₁ searched by word2vec; Step a3, sequentially perform the operation of step a2 on each word in the list _L1 obtained in step a1, and obtain a synonym list _{L1_total} corresponding to all words in the list _L1 ; Step a4, performing the operation of step a1 on the words in the list L' ₁ obtained in step a2 in turn, to obtain a synonym list L' _{1_total} corresponding to all the words in the list L'₁; Step a5, merge and remove duplicates from L ₁ , L' ₁ , L _{1_total} and L' _{1_total} to obtain the candidate word library of W ₁ , and further screen to finally obtain all the synonymous trigger words corresponding to the trigger word W ₁ ; For the trigger words in the initial trigger word list L, perform the operations of step a1 to step a5 to obtain all the synonymous trigger words corresponding to W ₁ , W ₂ , …, W _n-1 , W _n , and finally merge and remove duplicates of all the trigger words corresponding to W ₁ , W ₂ , …, W _n-1 , W _n to construct a trigger word dictionary; Step S4 includes: performing syntactic analysis on the candidate speech sentence to determine whether there is a subject-predicate relationship between the character and the trigger word; if so, determining that the sentence is the character's speech. 2. The method according to claim 1 is characterized in that in step S2, the text sentence segmentation adopts the method of first locating the position of the sentence delimiter to obtain ordered separator position list information, and then performing preliminary sentence segmentation according to the position list information; wherein the sentence delimiters are located by punctuation marks, line breaks, and spaces. 3. The method according to claim 2 is characterized in that, in step S2, a double quotation mark recognition method is adopted to determine whether the delimiter can be used as an actual delimiter. 4. The method according to claim 3 is characterized in that, in step S2, a double quotation mark recognition method is used to determine whether the delimiter can be used as an actual delimiter, specifically comprising: preliminarily separating the text according to the delimiter to obtain an ordered delimiter position list information, recorded as P: [x ₁ , x ₂ , x ₃ , …, x _m-1 , x _m ], wherein x ₁ , x 2 , x ₃ , …, x _{m-1 ,} x _m respectively represent the positions of the delimiters in the sentence, [1, x ₁ ] correspond to the start and end positions of the first sentence, and the delimiter position list is traversed from left to right. For the first traversal, the position where the start L of the sentence is located is fixed to 1, and the position corresponding to the end R of the sentence is x ₁ . The start position L and the end position R constitute a pre-sentence S, and determine whether the double quotation marks in the pre-sentence S meet the following conditions: Whether the number of double quotes is even; Whether the first one is a left quotation mark and the last one is a right quotation mark; If both of the above conditions are met, S is output as a complete sentence. In the next traversal, the end _x1 of this sentence is taken as the beginning L of the sentence, and the next element _x2 of _x1 is taken as the end R of the sentence, forming a pre-sentence S, and then the same judgment operation is performed. 5. The method according to claim 4 is characterized in that step S3 includes: segmenting each sentence, determining whether the segmented sentence contains known character information and trigger word information, and if so, taking the sentence as a candidate speech statement. 6. The method according to claim 5 is characterized in that in step S3, the word segmentation of each sentence specifically includes: analyzing and extracting the part of speech and the number of occurrences of each term through a large-scale corpus to construct a word segmentation dictionary for scanning the word graph of sentence segmentation, and generating a directed acyclic graph consisting of possible word formation situations of all Chinese characters in the sentence; for the directed acyclic graph, the initial word segmentation probability of each word segmentation situation is obtained according to the probability of occurrence of each term, and then the word segmentation penalty factor C is introduced to obtain the final word segmentation probability, and the maximum value of the final word segmentation probability is selected as the word segmentation result. 7. According to the method of claim 6, in step S3, the initial word segmentation probability P _ini is calculated using the following formula: P _{ini_i} =P( _{wi_1} )×P( _{wi_2} )×…×P( _{wi_n} ), Where w _{i_n} represents the nth word in the i-th possible word segmentation case, and P(w _{i_n} ) represents the word segmentation probability of the nth word w _{i_n} ; The final word segmentation probability P _final is calculated as follows: P _{final_i} = C ^v × _{Pini_i} Where P _{final_i} is the final word segmentation probability under the i-th possible word segmentation case; v is the number of domain words that have not been successfully segmented in the sentence.