Event Trigger Recognition Based on Positive and Negative Weight Computing and its Application

Event trigger recognition is a sub-task of event extraction, which is important for text classiﬁcation, topic tracking and so on. In order to improve the effectiveness of using word features as a benchmark, a new event trigger recognition method based on positive and negative weight computing is proposed. Firstly, the associated word feature, the part-of-speech feature and the dependency feature are combined. Then, the combination of these three features with positive and negative weight computing is used to identify triggers. Finally, the text classiﬁcation is carried out based on the event triggers. Findings from our experiments show that the application of our method achieves ideal results.


INTRODUCTION
With the rapid development of the Internet and mobile Internet, the volume of data is increasing rapidly, making big data a hot research topic. At the same time, various emergency situations occur frequently, often reported on the Internet. Hence, the research on events in the big data environment is attracting a great deal of attention.
An event refers to something that occurs at a specific time and in a particular environment, involves several actors, and exhibits several behavioral characteristics (Liu et al. 2009). The events in texts have important implications for the fields of automatic summarization, text classification, topic tracking, and information retrieval and so on. However, a computer cannot directly recognize the events in the texts that will affect subsequent applications, so it is especially important to automatically extract such events. Event extraction (EE) is a * tliao@aust.edu.cn † 1425548176@qq.com ‡ Email of corresponding author: sxzhang@aust.edu.cn § ztliu@shu.edu.cn sub-task of information extraction whereby event information is extracted from unstructured text in a structured form for subsequent use. Event extraction generally comprises two steps: event detection (ED) and event argument extraction (AE). The focus of this paper is on event trigger recognition, which is the main task of ED. Event triggers, which can clearly indicate the occurrence of an event, are mostly verbs and nouns. The purpose of event trigger recognition is to find event triggers in texts such as news texts and social media texts, which have a great impact on the effect of EE.
In the event trigger recognition process, we can use the word feature to identify the event trigger. This method takes the event trigger that appears in the training set as a feature, and when it also appears in the test set, we identify it as an event trigger. Although this method is simple and the recall rate is high, it has poor precision, resulting in poor final event recognition. However, we can use the word feature to identify the event trigger as a benchmark and add other features to improve event trigger recognition. Nevertheless, there are two challenges: 1) There are few features that greatly improve the result of the benchmark application.  Trigger table  construction   Trigger table  Matching, computing  positive and negative  weights   Event  trigger  recognition The results of event trigger recognition 2) The result achieved by using other features to improve event trigger recognition is still not ideal.
In order to solve these two problems, we first define a feature called the 'associated word', which can greatly improve the result of benchmark. After that, the single feature will be categorised as either a positive or negative feature which can be used for weight computing. The positive and negative weight computing method is used to identify the event triggers, and the result of the benchmark combined with a single feature is improved. Finally, based on the benchmark, we combine multiple features with positive and negative weight computing, and obtain the final and best event trigger recognition result.
The architecture of the event trigger recognition system in this paper comprises four parts: XML files preprocessing, feature set construction, trigger table construction and event trigger recognition, as shown in Figure 1. First, XML files preprocessing and feature set construction are conducted for the entire corpus. Then, by processing the training set of the corpus, the trigger table is constructed. Finally, the candidate triggers and their feature sets in the test set are matched in the trigger table, and the positive and negative weight computing is performed to complete the event trigger recognition process.
The rest of this paper is organized as follows: Section 2 describes related works. Section 3 introduces an event trigger recognition method based on positive and negative weight computing and the corresponding experiments. Section 4 proposes a text classification method based on event triggers and the experiment results. Conclusions are drawn in Section 5.

RELATED WORKS
Existing event trigger recognition methods mainly include rulebased methods and machine learning methods.

Rule-Based Methods
In the field of rule-based methods, Zhao

Machine Learning Methods
In the field of machine learning, Ahn (2006)

Segmentation
[" ", " ", " ", "8", " ", " ", " "] Trigger index [6] Part-of-speech tagging ["nt", "d", "v", "m", "n", "v", " From the above researches at home and abroad, it can be seen that, whether they are rule-based or machine learningbased event trigger recognition methods, most of them need to obtain features and then carry out subsequent research work. Therefore, we consider adding different features to assist event trigger recognition based on word features. At the same time, the positive and negative weight computing method is introduced to further improve the performance on event trigger recognition.

XML Files Preprocessing
For XML files in a corpus, the content of the Denoter tag is marked first (the tag is used to describe the event trigger). Then, the XML file is parsed into uncommented raw news text. After that, the NLP (Natural Language Processing) tool is used to segment sentences and words of the original news text. By using the trigger mark made previously, the index of the trigger in the sentence is recorded, so as to provide the basis for the subsequent judgment of the trigger. Finally, part-of-speech tagging and dependency parsing are performed through the NLP tool. Table 1 shows an example of a preprocessed XML file fragment.
In Table 1, the trigger index is used to record the index value of the trigger. In the example, " " (death) is the trigger word, and its index value is 6. The result of the part-of-speech tagging is one-to-one with the result of the word segmentation. For example, the part of the word " " (death) in the example is "v", which means the verb. The result of the dependency parsing also corresponds to the result of the word segmentation, but the content corresponding to each word includes two parts: the previous number represents the index of the parent node, and the following English abbreviation represents the dependency between the current word and the parent node. Taking the first "3: ADV" in the instance as an example, the "3" in front of the colon means that the parent node of the word " " (currently) is " " (caused) with an index value of 3, and the "ADV" after the colon indicates that the dependency relationship between " " (currently) and " " (caused) is adverbial.

Constructing Feature Set
After the XML files have been preprocessed, the feature set can be constructed. The feature set construction includes three parts: feature set construction of the associated words, the determination of the part-of-speech feature and the determination of the dependency feature.
Definition 1: event-triggered associated words, abbreviated as 'associated words', include the parent node (the parent node of the head word is empty) and the child nodes of the word in dependency parsing, as well as the left or right positions of the word in the sentence, but exclude punctuation.
First of all, an associated word feature set is constructed from all the words associated with a particular word within a sentence, thereby forming a set of words. If S rel represents the feature set of associated words, W par represents the parent node in dependency parsing, W chi represents the child nodes, W le f represents the left position word, W rig represents the right position word, and W wp represents the punctuation, then the feature set of associated words can be expressed by formula (1).
Then, the determination of the part-of-speech feature means identifying the part-of-speech of a word as the feature of the word, represented by Phere.
After that, the determination of the dependency feature means finding the dependency relationship as a feature when a word is found as a child node in dependency parsing, which is represented by R as_chi . Each word can have only one parent node in the dependency parsing of a sentence, that is to say, each word can be a child node only once in the dependency parsing of a sentence. The result of determining the dependency parsing relationship is unique.
Finally, S rel , P and R as_chi are combined to form a feature set S f ea , thereby completing the construction of the feature set, as shown in formula (2).

Construction of Trigger Table
The preprocessing of XML files and the construction of feature sets are carried out on the entire corpus, while the trigger table is constructed by processing the training set of the corpus. The trigger table is equivalent to a dictionary in which there are keyvalue pairs. The key here is the potential event trigger, and the value includes two parts: positive feature set and negative feature set. The contents of these two parts of the value comprise keyvalue pairs. They are also key-value pairs in both feature sets.
The key here is the feature of potential trigger, whose value is the number of times the feature appears in the training set. In the positive feature set, it is the statistics when potential triggers are used as triggers. And in the negative feature set, it is the statistics when potential triggers are used as non-triggers. This is the difference between the positive and negative feature sets. When the trigger table is constructed, it is processed in sentences. Firstly, each word in a sentence is traversed to determine whether or not it is a trigger according to the index of triggers. Following this determination, the feature in the feature set of the word is added to the corresponding position in the trigger table, and the construction process is completed. If the index of w == trigger index 8. Add S f ea to the positive feature set part of the keyword w as the secondary keyword 9. If the secondary keyword already exists 10. Increase its value by 1 11. Else create the secondary keyword first and initialize its value to 1 12. End If 13. Else similar to steps 8 to 12, the difference is that the feature is added to the negative feature set part of the keyword w 14. End If 15. End For 16. For keyword kw in TTC 17. If the positive feature set part of kw is empty 18. Delete kw and its value form TTC 19.

End If 20. End For
The constructed trigger table contains a number of key-value pairs. The key of each key-value pair is a potential trigger, and the value is the number of times the positive and negative features of the potential triggers appear. Thus, the trigger table can be considered as a statistic for the triggers in the training set and their positive and negative features.

Event Trigger Recognition
After the trigger table has been constructed, the event trigger recognition test can be carried out on the test set of the corpus. The candidate triggers and their feature sets obtained from the preprocessed XML files and the constructed feature sets are matched in the trigger table. According to the matching results and weight computing results, the triggers are judged in terms of event trigger recognition. The positive weight is obtained by formula (3), the negative weight is obtained by formula (4), and the final weight is obtained by formula (5).
In formula (3) and formula (4), pos c and neg c respectively represent the positive and negative weights of candidate trigger Set c as the candidate word for event trigger recognition 3. If c in TT If i in the negative feature set part of v 12. the value of the secondary keyword i is assigned to t n (c i ) 13.
If w c >= 0 18. c is judged as a trigger 19. Else c is judged as a non-trigger 20. End If 21. Else c is judged as a non-trigger 22. End For c. n c represents the number of features in the feature set of the candidate trigger c. t p (c i ) and t n (c i ) respectively represent the statistical values when the i-th feature of the candidate trigger c is successfully matched in the positive feature part and negative feature part of the trigger table. In formula (5), w c is the difference between pos c and neg c , that is, the final weight of candidate trigger c, which is used to judge whether c is a trigger. Algorithm 2 is the trigger recognition algorithm.
After the event trigger recognition, we need to compare the recognition result with the trigger index obtained from the XML files preprocessing stage. The result of this comparison indicates the level of accuracy of the recognition result, and prepares for the evaluation of the effectiveness of the event trigger recognition.

Setting of Experiments
The programming language used in the experiments is Python 3.6, the corpus is CEC 2.0, and the NLP tool is PYLTP (Python version of LTP). For the experiment, the corpus is divided into a training set and a test set in the ratio of 3:1. The result of event trigger recognition is evaluated by precision P, recall R and F 1 measures.

Comparison Experiment 1
In order to determine the effectiveness of self-defined associated word features on event trigger recognition, we set up the comparison experiment 1. In the experiment, we use the word feature to identify the event trigger as a benchmark. Then different features are added to the trigger table to assist trigger recognition. This completes comparison experiment 1.
The results of comparison experiment 1 are shown in Table 2. As indicated in Table 2, the introduction of dependency feature reduces the recall but improves the precision of event trigger recognition. The final F 1 measure is 4.54% higher than the benchmark, indicating that the dependency feature improves event trigger recognition. Compared with the dependency feature, the self-defined associated word feature reduces the recall R, but improves the precision P, making the final F 1 measure increase by 4.81% compared with the benchmark, reaching the highest 68.85% in Table 2. It shows that the associated word feature and dependency feature are similar and both have a good effect on event trigger recognition.

Comparison Experiment 2
Comparison experiment 1 shows that the F 1 measure is 68.85% by using the feature of associated word to assist trigger recognition, but obviously this result is not ideal. In order to strengthen the role of a single feature in event trigger recognition, this paper proposes a positive and negative weight computing method. To verify the effectiveness of this method, different features are added to the benchmark method, and the trigger recognition is carried out by combining the positive and negative weight computing. We compared this method with the method using the same feature but not combining the positive and   Table 3. It can be seen from Table 3 that after combining the positive and negative weight computing method, the result of all the event trigger recognition methods using a single feature has been greatly improved. The reason is that the method of positive and negative weight computing can take into account the positive and negative aspects and give full play to the role of a single feature in event trigger recognition. Among them, the F 1 measure of event trigger recognition method using positive and negative dependency features is the highest, reaching 77.80%. Therefore, the positive and negative weight computing method can effectively improve the results of event trigger recognition after adding different features to the benchmark, and the final experimental results also confirm this point.

Comparison Experiment 3
To further improve the effectiveness of event trigger recognition, we combine the three features of the associated word feature, the part-of-speech feature and the dependency feature. The combination of these three features with positive and negative weight computing is used to recognize event triggers. This is also the method introduced in section 3 of this paper. In order to determine the effectiveness of this method, we compare it with other existing event trigger recognition methods. The results of comparison experiment 3 are shown in Table 4. Because of the selection of different corpora, different preprocessing methods, and even different selections of documents from the same corpus for the test set and training set, event trigger recognition results will be greatly affected. At present, there is not a standardized, open and unified evaluation system in the field of event trigger recognition, so it is impossible to compare the advantages and disadvantages of each method objectively and fairly, and the effectiveness of only one method can be determined to a certain extent. From the comparison experiment 3, we can see that our method is the best of three evaluation indexes, which is some indication that the result of event trigger recognition based on positive and negative weight computing is ideal.

APPLICATION OF TEXT CLASSIFICATION BASED ON EVENT TRIGGERS
The research on event trigger recognition based on positive and negative weight computing was introduced earlier.

Classification Method
After calculating the feature weights of all trigger features in a text and each text category, they are accumulated by text category. The accumulation is used as the membership degree of the text and each text category. The text category with the maximum membership degree is regarded as the result of the text classification. The membership degree md(d,c) of text d and text category c can be calculated by formula (7). Where n represents the number of trigger features in text d. weight(t i , c) represents the feature weight of the i-th feature t in text d and the text category c, which can be obtained from the previous feature weight calculation stage. For the trigger feature t i that does not appear in the training set, set the weight(t i , c)to 0.
The classification result r (d)of text d can be calculated by formula (8) . When md(d,c) is the maximum, the corresponding text category c is the classification result of text d.

Text classification Experiment Based on Event Triggers
The text classification experiment based on event triggers was performed on CEC 2.0. In order to avoid the cascade error generated by the event trigger recognition method, we directly use the event triggers marked in CEC 2.0 for the experiment. In the experiment, 60% to 80% of the texts in each text category are randomly extracted to form the training set for each text category, and the remaining texts of each category constitute the test set for this category. The average precision P, recall R and F 1 measures are used to evaluate the overall text classification result. Figure 2. shows the result of text classification when the corpus is divided into a training set and a test set according to different proportions. The horizontal axis represents the proportion of the training set in the corpus, and the vertical axis represents the percentage of evaluation index. It can be seen that the result of text classification is the best when the training set and test set are divided into 76% and 24% respectively.
In order to evaluate the proposed text classification method more objectively, we compare it with naive Bayes method. Both the methods in this paper and naive Bayes use CEC 2.0 as the corpus, and the proportion of training set to test set is 76%:24%. The results of the comparison are shown in Table 5. According to Table 5, our method has less precision than the naive Bayes method only for the earthquake category, and other evaluation indexes are higher than naive Bayes method, especially the average F 1 measure which is 6.94% higher than the naive Bayes result.

CONCLUSIONS
In order to improve the effectiveness of using word features as a benchmark in event trigger recognition, an event trigger recognition method based on positive and negative weight computing is proposed. On this basis, we study the application of text classification based on event triggers. The experimental results show that our event trigger recognition method and its application have achieved excellent results. However, when using multiple features, the event trigger recognition method based on positive and negative weight computing is only a simple combination of methods using a single feature. Therefore, the next step is to combine deep learning with the method using multiple features.