An event trigger word extraction method and device

Abstract

The application discloses an event trigger word extraction method, which comprises the following steps: S1, inputting a text; S2, segmenting the text and calculating a segment feature vector; S3, calculating a trigger degree attenuation coefficient of different paragraphs; S4, calculating a value coefficient of a sentence n at different positions in the same paragraph; S5, calculating a sentence score according to the paragraph attenuation coefficient and the sentence value coefficient, taking the sentence score as a word position gain, and obtaining an event trigger sentence candidate set; S6, segmenting the sentences in the event trigger sentence candidate set and marking the parts of speech and the word sentiment polarity; and S7, calculating an event trigger degree according to the sentence position gain, the parts of speech and the sentiment polarity, and extracting the event trigger word. The application can be applied to long text level event trigger word extraction, and the purpose of accuracy, rapidity and wide range can be achieved.

Description

Technical Field

[0001] This invention relates to the field of natural language processing, and in particular to a method and apparatus for extracting event-triggered words. Background Technology

[0002] Event extraction aims to present unstructured text in a structured form. Event trigger word extraction is a subtask of event extraction, referring to the core words in a sentence that enable an event to occur, often verbs or nouns. The category corresponding to the trigger word is the event category contained in the sentence. Event trigger word extraction is the foundation of trigger word classification and event argument extraction in event detection, and an indispensable step in completing event extraction.

[0003] A method and apparatus for extracting Chinese event trigger words, published under CN113468884A, focuses on capturing the features of neighboring nodes of the same type, i.e., considering the hierarchical relationship of sentences, but does not measure the vector distance between sentences in the same paragraph and the first and last sentences of the paragraph. Furthermore, this comparative document considers the semantic information units and sentiment polarity of words, but does not extract event trigger words based on the combined sentiment polarity of words. Another comparative document, CN113255321A, on a text-level event extraction method in the financial field based on article entity word dependencies, uses the entity word dependencies in the article as input data for an event extraction deep learning model, without considering the influence of text paragraphs on event triggering, nor does it involve paragraph attenuation coefficients to aid in trigger word extraction.

[0004] The main problems exposed in current research on event trigger word extraction are: the extraction of event trigger words is concentrated at the single sentence level, and there is a lack of research on event trigger words in long texts; the positional information of the occurrence of trigger words is not considered, and different paragraphs and different sentences in the same paragraph contribute differently to the occurrence of the event; and the relationship between parts of speech and the emotional color of words and event trigger words is not considered. Summary of the Invention

[0005] The purpose of this invention is to overcome the problem that existing technologies for extracting event trigger words focus on the single sentence level and lack methods for extracting event trigger words from long texts, and to provide a method for extracting event trigger words. To achieve the above objective, this invention provides the following technical solution:

[0006] A method for extracting event trigger words, characterized by comprising the following steps:

[0007] S1. Enter text;

[0008] S2. Segment the text into sentences and calculate the feature vectors of the segmented sentences;

[0009] S3. Calculate the trigger attenuation coefficient for different paragraphs;

[0010] S4. Calculate the value coefficient of sentence n in different positions within the same paragraph;

[0011] S5. Calculate the sentence score based on the paragraph attenuation coefficient and the sentence value coefficient, and use the sentence score as the word position gain to obtain the candidate set of event-triggered sentences;

[0012] S6. Segment the sentences in the candidate set of event-triggered sentences and mark their parts of speech and sentiment polarity;

[0013] S7. Calculate the event trigger degree based on sentence position gain, part of speech, and sentiment polarity, and extract event trigger words. Preferably, in step S2, the text is segmented into sentences using periods, exclamation marks, and question marks as segmentation criteria. The calculation of the segmented sentence feature vectors is performed by using the Doc2Vec unsupervised algorithm to generate fixed-length feature vectors to represent sentences ending with periods.

[0014] Preferably, in step S3, the formula for calculating the trigger attenuation coefficient of different paragraphs is as follows:

[0015] d i =d0e -αi

[0016] Where d0 is the initial paragraph decay coefficient of the text, and α>0 is the exponential decay constant.

[0017] Preferably, in step S4, the value coefficient of sentences n at different positions in the same paragraph is calculated based on the directional distance between sentence n and the first sentence s and the last sentence e of the paragraph, and the calculation formula is as follows:

[0018]

[0019] Where Vs and Ve are the value coefficients of the first sentence s and the last sentence e of the paragraph, respectively, and d(n,s) and d(n,e) are the vector distances between the current sentence n and the first sentence s and the last sentence e of the paragraph, respectively.

[0020] Preferably, the calculation formula for the sentence score based on the paragraph attenuation coefficient and the sentence value coefficient is as follows:

[0021] Score(n|i)=d i V(n)

[0022] Preferably, the step of segmenting sentences in the candidate set of event trigger sentences and marking them with parts of speech and sentiment polarity is to use a bidirectional maximum matching method to segment sentences in the candidate set, remove stop words and punctuation marks, mark the remaining words with parts of speech, remove modal particles, auxiliary words, and interrogative words, and classify the sentiment polarity of words into positive, negative, and neutral.

[0023] Preferably, the process of calculating the event trigger degree based on sentence position gain, part of speech, and sentiment polarity, and extracting event trigger words, uses the following formula for calculating the event trigger degree:

[0024] Trigger(m)=Score(m)attr(m)|emo(m)|

[0025] Where attr(m) is the part-of-speech weight of word m, ranging from 0 to 1, and emo(m) is the sentiment polarity weight of word m, ranging from -1 to 1.

[0026] Based on the calculated triggering results, the word with the highest triggering degree is determined as the triggering word of the text event.

[0027] An event trigger word extraction device includes at least one processor and at least one processor-communicated memory, the memory storing instructions executable by at least one processor, the instructions being executed by at least one processor to enable at least one processor to execute the instructions described in any one of the methods.

[0028] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0029] This invention proposes formulas for calculating paragraph attenuation coefficients and sentence position value coefficients based on the influence of different paragraphs, sentences at different positions within the same paragraph, and word part-of-speech and sentiment polarity on event triggering intensity. Based on the attenuation coefficients and value coefficients, a candidate set of event triggering sentences is selected as the extraction pool for triggering words. Event triggering words are extracted based on the positional gain, part-of-speech, and sentiment polarity of words in the global text. This invention can be applied to event triggering word extraction at the level of long texts, achieving the goals of accuracy, speed, and wide scope. Attached image description:

[0030] Figure 1 The flowchart of this invention

[0031] Figure 2 The text image for Example 1 Detailed Implementation

[0032] The present invention will be further described in detail below with reference to experimental examples and specific embodiments. However, this should not be construed as limiting the scope of the above-mentioned subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0033] Example

[0034] This example uses the Chengdu Daily's report "Chengdu's Wild Birds' Circle of Friends Expands Again" from February 14, 2022, to illustrate the extraction of event trigger words.

[0035] like Figure 1 As shown, a method for extracting topic sentences from text includes the following steps:

[0036] Step 1: Enter text. In this example, enter the content of the Chengdu Daily report from February 14, 2022, titled "Chengdu's Wild Bird Network Expands Again."

[0037] Step 2: Sentence segmentation and sentence feature vector calculation. The report content is segmented into sentences, with a period, exclamation mark, or question mark marking the end of each sentence. Events are usually described in declarative sentences, so sentences ending with exclamation marks and question marks are deleted after segmentation. Then, the variable-length sentences are generated using the Doc2Vec unsupervised algorithm.

[0038] Step 3: Calculate the trigger decay coefficient d(i) for different paragraphs. Different paragraphs in the same text contribute differently to event triggering. For texts such as news and newspaper articles, the contribution to text event triggering decreases sharply as the paragraph number increases. Figure 2 The diagram shown is a schematic of the paragraph attenuation coefficient in this embodiment, and its calculation formula is as follows:

[0039] d i =d0e -αi

[0040] Wherein, d0 is the initial paragraph decay coefficient of the text, which is set to d0 = 1 in this embodiment; α is the exponential decay constant, which is set to α = 2 in this embodiment.

[0041] Step 4: Calculate the value coefficient V(n) of sentences n at different positions within the same paragraph. Generally, the first and last sentences of a paragraph have a higher probability of containing summary statements, i.e., a higher probability of containing event-triggered sentences (key sentences containing event-triggered words), while the probability of event-triggered sentences appearing in other positions is lower. As shown, for sentences n at different positions within the same paragraph (n is the sentence number of the current sentence in the paragraph), this invention calculates its value coefficient V(n) based on the vector distance between n and the first sentence s and the last sentence e of the paragraph. The calculation method is expressed as follows:

[0042]

[0043] Where Vs and Ve are the value coefficients of the first sentence s and the last sentence e of the paragraph, respectively, and d(n,s) and d(n,e) are the vector distances between the current sentence n and the first sentence s and the last sentence e of the paragraph, respectively. In this embodiment, Vs = 1 and Ve = 0.9, that is, when the sentence is the first sentence of the paragraph, its value coefficient is 1, and when the sentence is the last sentence of the paragraph, its value coefficient is 0.9.

[0044] Step 5: Obtain the event trigger sentence candidate set {C1,C2,…,Cp}. Calculate the sentence score Score(n|i) based on the paragraph decay coefficient d(i) from Step 3 and the sentence value coefficient V(n) from Step 4 (where i is the paragraph number of the sentence, and m is the sentence number within paragraph i). For example, the first sentence of the third paragraph, "The Painted Duck mainly breeds in Korea, Japan, Siberia, and other places. Its numbers were once extremely abundant, but due to overhunting, its numbers decreased significantly. According to the 1990 Asian Midwinter Waterbird Survey organized by the International Waterfowl Research Bureau, the overwintering population in China was only 4384," is represented by Score(1|2). Sort the sentence scores and select the top p scores to form the event trigger sentence candidate set {C1,C2,…,Cp}. In this embodiment, the top 10 sentences are selected to form the trigger word candidate set, i.e., p = 10. The sentence score Score(m|i) is calculated as follows:

[0045] Score(n|i)=d i V(n)

[0046] Step 6: Segment the sentences in the candidate set and label the word nature attr and the word sentiment polarity emo. Segment the candidate sentences based on the Bidirectional Maximum Matching method (Bi-MM), remove the stop words (such as "de", "di", "ma", etc.) and punctuation marks, and label the word nature (noun, verb, modal particle, interrogative word, etc.). Event trigger words are mostly nouns or verbs, so words with the nature of modal particles, auxiliary words, interrogative words, etc. are removed. Taking the first sentence in the third paragraph of the article "The Baikal Teal mainly breeds in places such as South Korea, Japan, and Siberia. Its population was once extremely abundant. However, due to overhunting, its population significantly decreased. According to the Asian Midwinter Waterfowl Census organized by the International Waterfowl Research Bureau in 1990, the number of wintering populations in China was only 4,384" as an example, after segmenting the sentence and removing the stop words, the word groups obtained are {"Baikal Teal" (n), "breed" (v), "South Korea" (n), "Japan" (n), "Siberia" (n), "di" (n), "population" (n), "abundant" (v), "hunt" (v), "decrease" (v), "year" (n), "International" (n), "Waterfowl" (n), "Research Bureau" (n), "organize" (v), "Asia" (n), "Midwinter" (n), "Waterfowl" (n), "Census" (v), "China" (n), "wintering" (v), "population" (n)}. Then, divide the word sentiment polarity into positive, negative, and neutral. "Baikal Teal" (n), "breed" (v), "South Korea" (n), "Japan" (n), "Siberia" (n), "di" (n), "population" (n), "year" (n), "International" (n), "Waterfowl" (n), "Research Bureau" (n), "organize" (v), "Asia" (n), "Midwinter" (n), "Waterfowl" (n), "Census" (v), "China" (n), "wintering" (v), "population" (n) are neutral, "hunt" (v), "decrease" (v) are negative, and "abundant" (v) is positive. Take the score Score(n|i) of the sentence where the word in Step 5 is located as the global position gain of the word. Finally, the features of a word include the sentence position gain Score, the word nature attr, and the sentiment polarity emo.

[0047] Step 7: Extract event trigger words. Calculate the event trigger degree Trigger of the word according to the word position gain Score, the word nature attr, and the sentiment polarity emo. If there are the same words, retain the word with a higher trigger degree. The word with the highest trigger degree becomes the event trigger word of the text. The calculation method of the event trigger degree Trigger(m) of the word m is as follows:

[0048] Trigger(m) = Score(m)attr(m)|emo(m)|

[0049] Where attr(m) is the part-of-speech weight (0~1) of word m. In this embodiment, if the part of speech of m is verb v, then attr(m) = 0.6; if the part of speech of m is noun n, then attr(m) = 0.4. emo(m) is the emotional polarity weight (-1~1) of word m. In this embodiment, if the emotional polarity is positive, then emo(m) = 0.5; if the emotional polarity is negative, then emo(m) = -0.6; if the emotional polarity is neutral, then emo(m) = -0.4.

[0050] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting event trigger words, characterized in that, Includes the following steps: S1. Enter text; S2. Segment the text into sentences and calculate the feature vectors of the segmented sentences; S3. Calculate the trigger decay coefficient for different paragraphs; S4. Calculate the value coefficient of sentence n in different positions within the same paragraph; S5. Calculate the sentence score based on the paragraph attenuation coefficient and the sentence value coefficient, and use the sentence score as the word position gain to obtain the candidate set of event-triggered sentences; S6. Segment the sentences in the candidate set of event-triggered sentences and mark their parts of speech and sentiment polarity; S7. Calculate the event triggering degree based on sentence position gain, part of speech, and sentiment polarity, and extract event triggering words; In step S4, the value coefficient of sentences n at different positions in the same paragraph is calculated based on the directional distance between sentence n and the first sentence s and the last sentence e of the paragraph. The calculation formula is as follows: Wherein, Vs and Ve are the value coefficients of the first sentence s and the last sentence e of the paragraph, respectively. , These are the vector distances between the current sentence n, the first sentence s, and the last sentence e, respectively.

2. The method for extracting event trigger words according to claim 1, characterized in that, In step S2, the text is segmented into sentences using periods, exclamation marks, and question marks as segmentation criteria. The feature vector of the segmented sentences is calculated by using the Doc2Vec unsupervised algorithm to generate a fixed-length feature vector for sentences ending with periods.

3. The method for extracting event trigger words according to claim 1, characterized in that, In step S3, the formula for calculating the trigger attenuation coefficient of different paragraphs is as follows: in, d 0 represents the initial paragraph decay coefficient of the text. >0 represents the exponential decay constant.

4. The method for extracting event trigger words according to claim 1, characterized in that, In step S5, the sentence score is calculated based on the paragraph attenuation coefficient and the sentence value coefficient, and the calculation formula is as follows: .

5. The method for extracting event trigger words according to claim 1, characterized in that, In step S6, the process of segmenting sentences in the candidate set of event-triggered sentences and labeling them with parts of speech and sentiment polarity involves using a bidirectional maximum matching method to segment sentences in the candidate set, removing stop words and punctuation marks, labeling the remaining words with parts of speech, removing modal particles, auxiliary words, and interrogative words, and classifying the sentiment polarity of words into positive, negative, and neutral.

6. The method for extracting event trigger words according to claim 1, characterized in that, In step S7, the event trigger degree is calculated based on sentence position gain, part of speech, and sentiment polarity, and event trigger words are extracted. The formula for calculating the event trigger degree is as follows: Where attr(m) is the part-of-speech weight of word m, ranging from 0 to 1, and emo(m) is the sentiment polarity weight of word m, ranging from -1 to 1. Based on the calculated triggering degree results, the word with the highest triggering degree is determined as the triggering word of the text event.

7. An event trigger word extraction device, characterized in that, A memory comprising at least one processor and at least one processor communicatively connected, the memory storing instructions executable by at least one processor, the instructions being executed by at least one processor to enable at least one processor to execute the instructions as described in any one of claims 1 to 6.

Images