CN110516215A - An automatic writing method of sports news - Google Patents

Abstract

The invention relates to an automatic writing method for sports news. First, a difference-time function is constructed according to live text data, and the data is modeled, and then the data is merged according to the characteristics of the difference-time function, and the live text segment information is extracted, Then, according to the key point information of the live text of the battle report data, extract important live clips, generate a live clip data set, extract the trigger conditions in the live clip data set, and compare them with the trigger conditions of the constructed template to select the optimal template. Fill in the factual data of the game into the template, generate battle report sentences, and finally generate battle report articles. The sports news articles generated by the sports news automatic writing method provided by the present invention have achieved excellent results in terms of the false-to-true rate, the true rate and the vivid rate, and can assist people to complete the sports news writing, thereby saving a lot of manpower and material resources, greatly Changing people's daily writing style can well meet the needs of practical applications.

Description

An automatic writing method of sports news

technical field

The invention belongs to the technical field of computer automatic writing, and in particular relates to a sports news automatic writing method.

Background technique

Chinese automatic writing is an important branch of artificial intelligence. It carries out intelligent writing work, and uses computers to completely automatically organize, extract, filter, filter, assemble, and write collected documents, and automatically generate a related article to realize automatic machine writing. The idea of opening up a new writing space, greatly changing people's daily writing methods, assisting people to complete writing, thus saving a lot of manpower and material resources. The defects of the sports news automatic writing method in the prior art include: high error rate in writing, untimely time, and strong subjectivity. Now it is urgent to develop an automatic writing method of sports news that can overcome the shortcomings of the existing technology.

SUMMARY OF THE INVENTION

Aiming at the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide a sports news automatic writing method that can avoid the above-mentioned technical defects.

In order to realize the above-mentioned purpose of the invention, the technical scheme provided by the present invention is as follows:

The sports news automatic writing method provided by the present invention achieves excellent results in terms of the false-to-true rate, the true rate and the vivid rate of the generated sports news articles, which can assist people to complete the sports news writing, thereby saving a lot of manpower and material resources, extremely Greatly changing people's daily writing methods can well meet the needs of practical applications.

Description of drawings

Figure 1 is a difference-time function diagram;

Figure 2 is a diagram of the data segmentation algorithm.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present invention.

An automatic writing method of sports news, firstly constructs a difference-time function according to the live text data, and models the data, and secondly combines the data according to the characteristics of the difference-time function, extracts the live text fragment information, and then according to the battle report The key point information of the data live text, extract important live clips, generate live clip data sets, extract the trigger conditions in the live clip data set, and compare them with the trigger conditions of the constructed template, select the optimal template, and match the Fill in the template with the factual data, generate the battle report sentence, and finally generate the battle report article.

Through the analysis of the data of the news report, it was found that the score, personal performance of the players, continuous scoring, anxiety, chasing points, go-ahead, last minute, buzzer-beater, etc. are the focus of the report. The present invention will model the data according to the score transformation of the two teams, and incorporate player performance, team performance, and game trends into it. Divide and extract the live broadcast data, match the extracted key points with the template library that has been constructed, and then assemble the extracted live broadcast data into the template library, and finally complete the automatic writing.

The construction method of the template library for sports news automatic writing includes: firstly, calculating the similarity of template data according to the classified categories to find the same template; secondly, using CRF to identify the trigger condition of the template, and finally forming the trigger condition-template.

Writing template calculation includes writing template calculation based on cosine similarity. Writing template calculation based on cosine similarity includes: using cosine similarity to calculate the similarity of two sentences, segmenting two sentences, listing all words, calculating Word frequency, write the word frequency vector; imagine the word frequency vector as two line segments in the space, both starting from the origin and pointing in different directions, forming an angle between the two line segments, and calculating the cosine similarity of the angle.

The writing template calculation includes the Word2Vec-based writing template calculation, and the Word2Vec-based writing template calculation includes: using the Word2Vec tool to formally represent the words in the background corpus with vectors, simplifying the processing of the text into vector operations in the vector space, through The similarity in the vector space is calculated to represent the semantic similarity of the text, and the related word expansion is realized.

CRF-based writing template trigger condition construction includes:

For a given game description sentence Text _i , the score difference between the away team and the home team is _Diffsore , the writing template is Y, and the trigger condition is Xi;

Diffsore _i =Text _i ·Score ₁ -Text _i ·Score ₂ ;

Y=Diffsore(∑ _i=1 X _i );

Calculate the score difference of each text and sort the diffsore

List = dis(diffsore);

List means that based on the text set after the difference, the data with the same difference is combined to form a difference data set, and the trigger condition is extracted for the data in the difference data set.

The construction of trigger conditions for writing templates based on CRF includes: role tagging; feature template selection. The role tags include: defining the trigger condition as: an NBA game in a certain period of time, describing the condition of the fact, denoted as CS; defining the trigger word as a word used to describe CS, denoted as CSword; each type of Trigger conditions include many trigger words. The construction of triggering conditions for writing templates based on CRF includes: firstly perform word segmentation and part-of-speech tagging on each annotated sentence; secondly, annotate role data, and finally select words, parts of speech, and roles as features, and use CRF to identify trigger words. The selection of feature templates includes: selecting words, parts of speech, and roles as features; using B, I, E, and O as the labeling symbols of trigger words, where B represents the first word of the trigger word, I represents the middle word, and E represents the end of the basic trigger word word, O represents a non-trigger word, and a single feature template and a compound feature template are used to identify trigger words.

The steps of data modeling based on the difference-time function include:

The score and time are crucial factors for the game, and the present invention uses the difference between the two teams to describe the actual process of the game.

For a given piece of live data T _i , the home team's score is Zsore _i , the away team's score is Ksore _i , and the difference between the two teams is Diffsore _i

Diffsore _i = Zsore _i -Ksore _i (1.1)

Using the difference function to obtain the trend of Diffsore _i with time, the present invention connects all the Diffsore _i points, and finally constructs the difference-time function:

Diffsore=F(time) (1.2)

As can be seen from Figure 1, if F(time) is a linear function in the time _j -time _i time period, the main states have the following three situations, and each trend represents a different state:

1. If F(time) _j - F(time) _i > 0, it means that the home team is ahead in the time _j - time _i time period

2. If F(time) _j - F(time) _i = 0, it means that the two teams are tied within the time _j - time _i time period

3. If F(time) _j - F(time) _i <0, it means that the home team is behind in the time _j - time _i time period

If F(time) is a multiple function in the time _j -time _i time period, it is a mixed combination of the above cases, which can be expressed as shown in Table 1.1 below

Table 1.1 Difference-time function

lead draw behind lead - draw draw - lead trailing - draw lead - lag draw - behind lag - lead lead - draw - lead draw - lead - behind behind - draw - behind lead-draw-behind draw - lead - draw behind-draw-lead lead-lag-lead draw-behind-draw lag - lead - lag lead-lag-draw draw-behind-lead trailing-leading-drawing .... ... ...

The steps of data segmentation based on the difference-time function include:

Since it is impossible to report every detail during the reporting process, key details need to be reported. Therefore, the present invention needs to segment the live text to determine the key data segments of the game. The present invention makes judgments based on the properties of the differential-time function.

If F(time) _j -F(time) _i =0, and _timej - _timei >3 minutes, keep the data, otherwise delete the data. Secondly, if F(time) _j -F(time) _i ＞=0, merge the leading data into fragments, and if F(time) _j -F(time) _i ＜=0, merge the backward data into fragments .

The specific steps of the data segmentation algorithm are as follows:

Algorithm: Slice Algorithm

1.Input: Live text Text=Text ₁ Text ₂ Text ₃ ....Text _n

2. Preprocess: For each Text _i , calculate the live score of each text by formula 1.1, and determine the symbolic relationship between F(time) _j -F(time) _i

3. Shard position poi=1, slice set Slict=NULL, number of slices, num=1,

4. for K=1 to n do

5.if F(Text _k .Time) is the opposite of F(Text _k+1 .Time)

6. begin _num = Diffsore _poi

7. end _num = Diffsore _k

8.time _num1 = Text _poi · Time

9.time _num2 = Text _k · Time

10. slice _num = {begin _num , end _num , time _num1 , time _num2 }

11.slice _num is added to the Slict collection

12.poi←k, num=num+1

13.endif

14.Output: Sliced collection Slict

As shown in Figure 2, the present invention can divide data into different segments through the data segmentation algorithm, which is convenient for subsequent extraction and provides services for data generation.

Analysis of key points of the game includes:

The determination of key points is very important for the report of the game. Through the analysis of the game report, the present invention finds that the key parts of the game report roughly have the following six situations: the first type of game turning point, the second type of game focus, and the third type Player performance, the fourth type of team performance, the fifth type of coaching ability, and the sixth type of buzzer-beater.

Definition 1.1 The turning point of the game: It refers to the process in which the score of the two teams gradually expands to reach a maximum value within a certain period of time, or the process in which the score of the two teams gradually shrinks to a minimum value. The process of this extreme value change often reflects the climax of the game and needs to be reported emphatically.

Definition 1.2 Game focus: Refers to two teams in a certain period of time, the two teams lead alternately, and the game is in a tense and anxious atmosphere. It usually appears in the second half of the fourth quarter and should be highlighted.

Definition 1.3 Player performance: refers to the excellent or bad performance of a player within a specific period of time, such as consecutive points, critical goals, critical assists, consecutive strikes, etc., usually with continuity and criticality per unit time. In general, stars are reported more often than other players. It will also be reported if the bench player can help the team take the lead.

Definition 1.4 Team performance: refers to the excellent or bad performance of the team in a specific time, such as the team scoring consecutive points in a specific time to play the climax of the game, or the team scoring no points or only a few points in a unit time. Team performance is usually dominated by negative performances. If a team does not score for a few minutes.

Definition 1.5 Coach's commanding ability: The coach is the brain of the team and can usually turn the situation around in the game. The coach's active suspension or substitution often results in unexpected results, which are usually reported.

Definition 1.6 buzzer-beater: Refers to the ball thrown by the player near the end of each quarter (usually within 10s), which has elements of luck and ability. If the ball is scored, it will improve the morale of the team and should be reported.

The steps to automatically generate sports news based on trigger conditions include:

According to the key points of the game report and the data segmentation algorithm based on the difference-time function, find the corresponding key data, extract the data, build the report triggering conditions of the live text, and search for it in the template library that has been constructed. According to the corresponding template, fill in the fact data, regenerate the sentence, and finally complete the NBA automatic writing.

Use Key to represent the set of key points of the game, keyi to represent the key points of the game, Slict to represent the split live text data, Sent to represent the writing template, Condition to represent the trigger condition, Cword to represent the trigger word, Condition _i = {Cword ₁ , Cword ₂ .. Cword _n }, Sent _i = {condition ₁ , conditong ₂ , ... conditon _n }. Slict={Slict1, Slict2, Slict3...Slictn}. If Slict _i contains a key, keep the data fragment Slicti, otherwise discard it.

The present invention re-extracts Slict _i with ReportSenti = ₁ to form a report data set Report. That is, Report=Σ _i=1 Slict _i and Slict _i ∈Key.

The present invention extracts the trigger condition RCondition in Report _i , and searches in the SCondition in the existing Sent template set, if every trigger condition extracted in RCondition _i can be found in SCondition _i , then the SCondition _i template will be activated . For RCondition _i , there may be multiple SCondition _i corresponding to it, and the present invention randomly selects a template for data filling to form a report sentence.

Experimental results and analysis:

The invention crawls 867 news corpus construction templates of Sina, and crawls all the text live broadcast data of Hupu 2017 season, a total of 1305 pieces.

Evaluation indicators:

The present invention adopts three kinds of evaluation indexes commonly used in the prior art (please refer to: Chen Yujing, Lv Xueqiang, Zhou Jianshe, Li Ning. Research on Automatic Writing of NBA Event News [J]. Journal of Peking University (Natural Science Edition), 2017, (02): 1+6.201), then use the method of manual evaluation, invite 3 experts to score, and use 3 indicators including: index 1, whether it is written by computer; index 2, whether it conforms to the authenticity of the competition; index 3 , describing whether the language is vivid. The present invention has formulated two evaluation methods, one is the strict accuracy rate, if all three experts believe that it is not written by computer, it is considered that it is not written by computer. Similarly, the strict standard of the latter two indicators also adopts this method. The other is the loose accuracy rate. As long as a majority of the three experts (that is, two or more people) think that it is not written by a computer, it is considered that it is not written by a computer. The loose standard of the latter two indicators also adopts this method. The representation method is as follows:

Among them, C _strict , R _strict and L _strict represent strict accuracy, respectively, and C _lenient , R _lenient and L _lenient represent loose accuracy. C _i (i=1, 2, 3) means that the i-th expert thinks it is not written by the computer, Ri ( _i =1, 2, 3) means that the i-th expert thinks it is true to the game, Li ( _i = 1, 2, 3) indicates that the i-th expert thinks the description language is vivid, and N represents the total number of articles.

The present invention randomly selects 100 live texts for experiments, the method proposed by the present invention and the method commonly used in the prior art (for this method, please refer to the literature: Chen Yujing, Lv Xueqiang, Zhou Jianshe, Li Ning. Research on automatic writing of NBA event news [J] .Journal of Peking University (Natural Science Edition), 2017, (02): 1+6.201) made a comparison, the experimental results show that the method proposed by the present invention is better than the comparison method, and the specific conditions are shown in Tables 1.2, 1.3 and 1.4:

Table 1.2 The experimental result table of false and true

Table 1.3 Authenticity test result table

Table 1.4 Vividness test result table

It can be seen from Table 1.2, Table 1.3 and Table 1.4 that the method proposed by the present invention is better than the comparative experimental method, because the solution proposed by the present invention is superior to the comparative method in template construction and selection, for example, the template selection method is fixed There are more than 50, and the template of the present invention is dynamic and can be continuously updated through news reports. Therefore, the present invention is superior to the contrast method in terms of confounding the real, authenticity and vividness, especially in terms of vividness.

The method of the present invention is based on the false-to-true rate, wherein the strict false-to-true rate reaches 90%, and the loose false-to-true rate can reach 95%, indicating that the automatic writing of the computer can be considered as manual writing. The strict truth rate of 94% and the loose truth rate of 96% show the effectiveness of the method of the present invention, which meets the real live broadcast requirements of NBA event news. The strict vividness rate reaches 85%, and the loose vividness rate can reach 93%, which shows that the vividness of the method of the present invention also achieves a good result.

The above-mentioned embodiment only expresses the embodiment of the present invention, and its description is more specific and detailed, but should not therefore be construed as a limitation to the scope of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the appended claims.

Claims (8)

1. an automatic writing method of sports news, it is characterized in that, comprising: carry out data modeling based on difference-time function; carry out data segmentation based on difference-time function; analyze key points of game; automatically generate sports news based on trigger condition . 2. sports news writing method according to claim 1, is characterized in that, described sports news automatic writing method, at first according to live text data constructs difference-time function, and data is modeled, secondly according to difference- The characteristics of the time function combine the data, extract the information of the live text segment, and then extract the important live segment according to the key point information of the live text of the battle report data, generate the live segment data set, and extract the trigger conditions in the live segment data set. The constructed template trigger conditions are compared, the optimal template is selected, and the factual data of the competition is filled in the template, the battle report sentence is generated, and finally the battle report article is generated. 3. sports news writing method according to claim 1, is characterized in that, the step of carrying out data modeling based on difference-time function comprises: For a given piece of live data T _i , the home team's score is Zsore _i , the away team's score is Ksore _i , and the difference between the two teams is Diffsore _i =Zsore _i -Ksore _i ; Use the difference function to get the trend of Diffsore _i with time, connect all the Diffsore _i points, and finally construct the difference-time function: Diffsore=F(time); If F(time) is a function in the time _j -time _i time period, the state has the following three cases, each trend represents a different state: If F(time) _j - F(time) _i > 0, it means that the home team is ahead in the time _j - time _i time period; If F(time) _j - F(time) _i = 0, it means that the two teams are in a draw within the time _j - time _i time period; If F(time) _j - F(time) _i <0, it means that the home team is behind in the time _j - time _i period; If F(time) is a multiple function in the time _j -time _i time period, it is a mixed combination of the above three cases. 4. sports news writing method according to claim 1, is characterized in that, the step of carrying out data segmentation based on difference-time function comprises: Judging according to the nature of the difference-time function: If F(time) _j -F(time) _i =0, and _timej - _timei >3 minutes, keep the data, otherwise remove the data; If F(time) _j -F(time) _i >= 0, the leading class data is merged into segments; If F(time) _j - F(time) _i <= 0, the lag class data is segmented. 5. sports news writing method according to claim 1 is characterized in that, the analysis of key points of the game comprises: the key part of the reported game has the following six situations: the first type of game turning point, the second type of game focus, the third Player performance, team performance fourth, coaching ability fifth, buzzer-beater sixth. 6. sports news automatic writing method according to claim 1-5 is characterized in that, the definition of described six kinds of situations is respectively: The turning point of the game refers to the process in which the score of the two teams gradually expands to reach a maximum value within a certain period of time, or the process in which the score between the two teams gradually shrinks to a minimum value; this process of extreme value changes reflects the climax of the game. , to focus on reporting; The hot spot of the game refers to the two teams taking the lead alternately within a certain period of time, and the game is in a tense and anxious atmosphere; if it occurs in the second half of the fourth quarter, it should be reported intensively; Player performance refers to the excellent or bad performance of a player within a certain period of time, which is continuous and critical within a unit of time; the number of times a star is reported is higher than that of other players; if a substitute player can help the team take the lead, it must also be reported. report; Team performance refers to the excellent or bad performance of the team within a certain period of time, or the team has no points or only few points in a unit time; the performance of the team is mainly negative; Coach's commanding ability: the coach takes the initiative to suspend or make a substitution, and report it; The buzzer-beater is the ball thrown by the player towards the end of each quarter and is reported. 7. sports news writing method according to claim 1 is characterized in that, the step of automatically generating sports news based on trigger condition comprises: according to the key point of game report and the data segmentation algorithm based on difference-time function, find corresponding Key data, extract the data, construct the report trigger conditions of the live text, search in the template library that has been constructed, find the corresponding template, fill in the fact data, regenerate the sentence, and finally complete the automatic writing of sports news. 8. sports news automatic writing method according to claim 1-7 is characterized in that, the step of automatically generating sports news based on trigger condition comprises: Use Key to represent the set of key points of the game, keyi to represent the key points of the game, Slict to represent the split live text data, Sent to represent the writing template, Condition to represent the trigger condition, Cword to represent the trigger word, Condition _i = {Cword ₁ , Cword ₂ .. Cword _n }, Sent _i ={condition ₁ , conditong ₂ ,...conditon _n }, Slict={Slict1, Slict2, Slict3...Slictn}, if Slict _i contains key, keep the data segment Slicti, otherwise give up, Re-extract Slict _i with ReportSenti = 1 to form a report data set Report, Report=∑ _{i =1} Slict _i and Slict _i ∈Key; Extract the trigger condition RCondition in Report _i , and look for it in the SCondition in the existing Sent template set. If every trigger condition extracted in RCondition _i can be found in SCondition _i , the SCondition _i template will be activated; randomly Select a SCondition _i template for data filling to form report sentences.