阅读说明：本技术 面向金融领域的因果关系提取方法和系统 (Financial field-oriented causal relationship extraction method and system ) 是由 高楠 董嘉豪 冯伟强 周越 俞凯乐 葛婧 于 2021-06-28 设计创作，主要内容包括：面向金融领域的混合因果关系提取方法,包括：步骤1：扩充训练语句、训练模型；步骤2：提取中心词；步骤3：查中心词表；步骤4：以中心词为界划分句子为两边,再进行相应的句子预处理,删去多余的句子；步骤5：矫正中心词位置；步骤6：分类型进行因果识别；步骤7：模型调优。本发明还包括面向金融领域的混合因果关系提取系统。本发明以金融领域的因果语料作为训练集训练模型,同时建立中心词库,实现面向金融领域的文本的混合因果关系的提取,对指导事件溯因、问答、把握行业动态等方面提供一定支持。(The mixed causal relationship extraction method for the financial field comprises the following steps: step 1: expanding training sentences and training models; step 2: extracting the central word; and step 3: looking up a central word list; and 4, step 4: dividing sentences into two sides by taking the central words as boundaries, then carrying out corresponding sentence preprocessing, and deleting redundant sentences; and 5: correcting the position of the central word; step 6: identifying cause and effect according to types; and 7: and (6) optimizing the model. The invention also comprises a mixed causal relationship extraction system oriented to the financial field. The method takes causal corpora in the financial field as a training set training model, and meanwhile, a central word library is established, so that extraction of mixed causal relationships of texts facing the financial field is realized, and certain support is provided for guidance event tracing, question answering, industry dynamics grasping and the like.)

1. The mixed causal relationship extraction method for the financial field comprises the following steps:

step 1: expanding training sentences and training models; acquiring related statement expansion corpora from research and report and articles in the financial field, respectively labeling sentences with cause relationship and sentences without cause-effect relationship, and training the sentences as training sets to judge the existence of cause-effect relationship;

step 2: extracting the central word; finding out the central word in the sentence, expanding a central word list, and labeling the common position and the common mode of the central word list;

and step 3: looking up a central word list; obtaining an input sentence, searching a central word list for the sentence with the central word, and obtaining the common position and the common mode of the central word; if no central word exists, the model is changed into three categories (non-causal, pre-causal and post-causal pre-causal), and then partial sentences are divided according to a permutation and combination mode;

and 4, step 4: dividing sentences into two sides by taking the central words as boundaries; the sentences are divided into left and right sides by taking the central words as boundaries, and are divided into sentences according to punctuation marks, then corresponding sentence preprocessing is carried out, and redundant sentences are deleted;

and 5: correcting the position of the central word; correcting the position of the central word with possible errors;

step 6: identifying cause and effect according to types; the method comprises the steps of carrying out separate processing on different positions of three central words with a central word in the middle, a central word in the front and a central word in the back, obtaining two lists by finding out an entity and judging whether two-two causal relations exist or not, realizing the identification of one-cause multi-effect, multi-cause one-effect, one-cause one-effect and multi-cause multi-effect, and finally judging to obtain one of the cause lists and one of the result lists;

and 7: adjusting and optimizing the model; and manually distinguishing and parameter tuning are carried out on the obtained preliminary model result, and the causal judgment effect of the model is improved through data set targeted capacity expansion and over-parameter tuning.

2. The financial domain-oriented mixed causality extraction method as recited in claim 1, wherein: finding out the central words in the sentence described in step 2 includes "cause", "because"; the expanded core word list is used for searching synonyms to find other words capable of indicating cause and effect relationship and adding the words into the core word list; the common positions are three in the middle, the front and the back, and the common mode is two modes of representing the pre-cause consequence and representing the pre-cause consequence.

3. The financial domain-oriented mixed causality extraction method as recited in claim 1, wherein: the correcting the position of the central word in the step 5 specifically comprises the following steps:

a) the central word should be in the middle, but the sentence without noun on the left will reset its position to the front; otherwise, the position is set as the back;

b) correcting when the central word is in front and the right side has no two sentences with nouns, setting the right side as a middle sentence with only one entity and setting the right side as a back sentence without the entity;

c) the central word is corrected when there are no two sentences with nouns on the back side and one sentence with an entity on the left side is set as the middle, and the sentence without the entity on the left side is set as the front.

4. The financial domain-oriented mixed causality extraction method as recited in claim 1, wherein: the step 6 of separately processing the following three situations where the headword is located at different positions specifically includes:

c) the central word is in the middle: finding out the nearest entities on the left side and the right side, marking as left and right, and respectively placing the left and right entities into a left list and a right list; the rest sentences on the left side and right are sequentially combined and are placed in a trained BERT model to judge the causal relationship, if the causal relationship exists, the rest sentences on the left side and left are sequentially combined and are placed in the model to judge the causal relationship, and if the causal relationship exists, the rest sentences on the right side and left are added in a right list; finally, judging which is a reason list and which is a result list according to the common mode of the central words;

d) the central word is preceded: recording an entity nearest behind a central word as a head, adding the head into a head list, combining the rest sentences with the head in sequence, putting the head and the rest sentences into a trained BERT model for judging the cause-effect relationship, putting the head list into a tail list if the cause-effect relationship exists, otherwise putting the head list into a temporary list, wherein the temporary list is used for temporarily storing sentences possibly including the sentence where the cause/the result exists, and putting the head list into the head list if the cause-effect relationship can be formed by half of sentences in the temporary list with the tail list, and finally judging which one of the head list and the tail list is a cause list and which one is a result list;

e) the core word is followed: and recording an entity nearest to the head of the central word as tail, adding the tail into a tail list, combining the rest sentences with the tail in sequence, putting the tail and the rest sentences into a trained BERT model for judging the cause-effect relationship, putting the tail list into a head list if the cause-effect relationship exists, otherwise putting the tail list into a temporary list, temporarily storing sentences possibly including the sentence where the cause/the result exists, putting the tail list into the temporary list if half of sentences capable of following the head list in the temporary list form the cause-effect relationship, and finally judging which is a cause list and which is a result list.

5. The financial domain-oriented mixed causality extraction method as recited in claim 1, wherein: the targeted capacity expansion of the data set in step 7 specifically includes: after the model is trained, a plurality of pieces of data are taken randomly for detection, each result is analyzed and evaluated, and if the type of errors is common, a plurality of pieces of data are found for the type to supplement a training set; marking sentences with the prefixes of 'title', 'report' and 'discovery' as causality-free sentences and adding the causality-free sentences into a training set; if the error is one case, adding the error into the training set after correcting the error; the above steps are repeated for three times;

the super-parameter tuning specifically comprises the following steps: according to the BerT related paper, the recommended hyper-parameters are preferably adjusted to be, least-rate (Adam) 5e-5,3e-5,2e-5Number of epochs 2,3, 4; and obtaining a plurality of models according to the recommended hyper-parameters, and selecting the best model to improve the discrimination effect by comparing precision, call and f1 values of the models on the test set.

6. A system for implementing a hybrid causal relationship extraction method for the financial domain as claimed in claim 1, wherein: comprises a training sentence expanding and model training module, a central word extracting module, a central word table searching module, a sentence dividing module, a central word position correcting module, a cause and effect identifying module and a model optimizing module which are connected in sequence, wherein,

the training sentence expanding and model training module comprises: acquiring related statement expansion corpora from research and report and articles in the financial field, respectively labeling sentences with cause relationship and sentences without cause-effect relationship, and training the sentences as training sets to judge the existence of cause-effect relationship;

the headword draws the module and includes: finding out the central word in the sentence, expanding a central word list, and labeling the common position and the common mode of the central word list;

the central word table searching module comprises: obtaining an input sentence, searching a central word list for the sentence with the central word, and obtaining the common position and the common mode of the central word; if no central word exists, the model is changed into three categories (non-causal, pre-causal and post-causal pre-causal effects), and then the divided sentence division module which can realize partial sentences according to the permutation and combination mode comprises: the sentences are divided into left and right sides by taking the central words as boundaries, and are divided into sentences according to punctuation marks, then corresponding sentence preprocessing is carried out, and redundant sentences are deleted;

the central word position correction module comprises: correcting the position of the central word with possible errors;

the cause and effect identification module includes: the method comprises the steps of carrying out separate processing on different positions of three central words with a central word in the middle, a central word in the front and a central word in the back, obtaining two lists by finding out an entity and judging whether two-two causal relations exist or not, realizing the identification of one-cause multi-effect, multi-cause one-effect, one-cause one-effect and multi-cause multi-effect, and finally judging to obtain one of the cause lists and one of the result lists;

the model tuning module comprises: and manually distinguishing and parameter tuning are carried out on the obtained preliminary model result, and the causal judgment effect of the model is improved through data set targeted capacity expansion and over-parameter tuning.

Technical Field

The invention relates to a method and a system for relation extraction and cause and effect discrimination, in particular to a method and a system for extracting mixed cause and effect relations in the financial field, which realize the extraction of explicit and implicit cause and effect relations in financial language segments.

Background

Under the background of the era of economic globalization, the research on the causal relationship of financial events has important referential significance for establishing the national macro regulation and control policy. For example, there are a number of manual summaries, including a number of causal relationships, in the company's financial reports on the market. The identification of the causal relationship can help people to know the coming and going arteries and veins among events, obtain the evolution relationship of the events and be beneficial to prediction and decision. Meanwhile, with the rapid development of natural language processing technology, a great deal of research foundation exists in the field of text event extraction and event cause and effect extraction, and a rapid screening and discovering method for cause and effect in financial events is still lacked in the existing research. The project is designed to be based on a pretraining language model of BERT, causal corpora in the financial field is used as a training set training model, a central word library is established, extraction of mixed causal relationships of texts facing the financial field is achieved, and certain support is provided for guidance event traceability, question answering, industry dynamics control and the like.

The extraction of the causal relationship in the financial field has the following problems:

(1) the financial industry is rapidly developed, the transaction activity is active, the content of event information is huge, and the issued information text has the characteristics of long length, complex syntax structure and the like, so that the processing of the text has great difficulty and difficulty.

(2) Causal relationships confirm the presence of outliers. For financial texts with complex syntactic structures, a sentence often has a plurality of cause and effect relationships, the cause and the effect in a group of cause and effect may be one or more, and all the causes and the effects corresponding to the central word need to be extracted.

(3) There is a difficulty in extracting the headword. The central word is a component of the link cause body and the effect body in the cause and effect relationship of the sentence and is a word capable of clearly expressing the occurrence of things. If the central word cannot be found correctly in the text, the following extraction of the causal pairs has deviation in different degrees or is directly wrong.

Cause and effect Extraction (Causality Extraction) is a relationship Extraction task in natural language processing, and is used for mining event pairs with cause and effect in texts. In recent years, a method combining neural network and machine learning avoids high-cost characteristic engineering in a traditional event causal relationship extraction method, and can capture implicit and fuzzy causal relationships in texts. In the financial field, with the continuous development of economic life, financial events occur continuously, and a large amount of event information is generated. Compared with other fields, events occurring in the financial field often have complexity, relevance and professionality, and if the information relationship is directly judged and processed, great difficulty, accuracy and credibility are affected to a certain extent. How to find the implied potential laws from the massive financial event data and scientifically analyze the causal relationship information of the financial events also become a problem to be solved urgently. Therefore, the extraction of the financial event cause and effect relationship is endowed with importance and necessity. Financial logic is extracted from research and report in the financial field to construct a case map, which plays an important role in guiding event tracing, question answering, grasping industry dynamics and the like. Meanwhile, with the rapid development of natural language processing technology, a great deal of research foundation exists in the field of text event extraction and event cause and effect extraction, and a rapid screening and discovering method for cause and effect in financial events is still lacked in the existing research. Therefore, the project is to construct a financial causal relationship recognition model based on a BERT pre-training language model, implicit and explicit modes are achieved, causal relationship extraction of one-cause-one-effect, multiple-cause-multiple-effect, one-cause-multiple-effect and one-cause-multiple-effect provides future dynamic information of the field for financial decision and other practical application requirements, and therefore the problems of high enterprise risk prediction analysis cost, low efficiency, high threshold and low timeliness are solved.

Disclosure of Invention

The invention aims to overcome the defects of the prior art and provides a causal relationship extraction method and a causal relationship extraction system for the financial field.

The invention discloses a BERT pre-training language model-based financial causal relationship recognition model, which realizes implicit and explicit extraction of causal relationships of one cause, multiple causes and multiple effects, one cause and multiple effects and one cause and multiple effects, and provides future dynamic information in the field for financial decision and other practical application requirements, thereby solving the problems of high cost, low efficiency, high threshold and low timeliness of enterprise risk prediction and analysis.

The invention relates to a causal relationship extraction method for the financial field, which comprises the following steps:

step 1: obtaining related sentence expansion linguistic data from research and report and articles in the financial field, respectively marking sentences with cause relation and sentences without cause-effect relation, and taking the sentences as training sets to train BERT models to judge the existence of cause-effect relation. The model architecture is shown in FIG. 2.

Step 2: and extracting the central words in the sentences, expanding a central word list, and labeling the common positions and common modes of the central words.

And step 3: the method comprises the steps of obtaining an input sentence, searching a central word list for the sentence with the central word, and obtaining the common position and the common mode of the central word. If no central word exists, the model is changed into three categories (non-causal, ante-causal and post-causal pre-causal), and then partial sentences are divided according to a permutation and combination mode.

And 4, step 4: the sentences are divided into left and right sides by taking the central words as boundaries, and are divided into sentences according to punctuation marks, then corresponding sentence preprocessing is carried out, and redundant sentences are deleted.

And 5: and correcting the position of the central word which is possibly wrong.

Step 6: the method comprises the steps of carrying out separate processing on different positions of three central words with the central word in the middle, the central word in the front and the central word in the back, obtaining two lists by finding out entities and judging whether the two-two causal relationship exists or not, realizing the identification of one-cause multi-effect, multi-cause one-effect, one-cause one-effect and multi-cause multi-effect, and finally judging to obtain one of the cause lists and one of the result lists.

And 7: and manually distinguishing and parameter tuning are carried out on the obtained preliminary model result, and the causal judgment effect of the model is improved through data set targeted capacity expansion and over-parameter tuning.

Preferably, the finding of the central word in the sentence in step 2 includes "cause", "because"; the expanded core word list is used for searching synonyms to find other words capable of indicating cause and effect relationship and adding the words into the core word list; the common positions are three in the middle, the front and the back, and the common mode is two modes of representing the pre-cause consequence and representing the pre-cause consequence.

The correcting the position of the central word in the step 5 specifically comprises the following steps:

a) the central word should be in the middle, but the sentence without noun on the left will reset its position to the front; otherwise, the position is set as the back;

b) correcting when the central word is in front and the right side has no two sentences with nouns, setting the right side as a middle sentence with only one entity and setting the right side as a back sentence without the entity;

c) the central word is corrected when there are no two sentences with nouns on the back side and one sentence with an entity on the left side is set as the middle, and the sentence without the entity on the left side is set as the front.

The step 6 of separately processing the following three situations where the headword is located at different positions specifically includes:

a) the central word is in the middle: finding out the nearest entities on the left side and the right side, marking as left and right, and respectively placing the left and right entities into a left list and a right list; the rest sentences on the left side and right are sequentially combined and are placed in a trained BERT model to judge the causal relationship, if the causal relationship exists, the rest sentences on the left side and left are sequentially combined and are placed in the model to judge the causal relationship, and if the causal relationship exists, the rest sentences on the right side and left are added in a right list; finally, judging which is a reason list and which is a result list according to the common mode of the central words;

b) the central word is preceded: recording an entity nearest behind a central word as a head, adding the head into a head list, combining the rest sentences with the head in sequence, putting the head and the rest sentences into a trained BERT model for judging the cause-effect relationship, putting the head list into a tail list if the cause-effect relationship exists, otherwise putting the head list into a temporary list, wherein the temporary list is used for temporarily storing sentences possibly including the sentence where the cause/the result exists, and putting the head list into the head list if the cause-effect relationship can be formed by half of sentences in the temporary list with the tail list, and finally judging which one of the head list and the tail list is a cause list and which one is a result list;

c) the core word is followed: and recording an entity nearest to the head of the central word as tail, adding the tail into a tail list, combining the rest sentences with the tail in sequence, putting the tail and the rest sentences into a trained BERT model for judging the cause-effect relationship, putting the tail list into a head list if the cause-effect relationship exists, otherwise putting the tail list into a temporary list, temporarily storing sentences possibly including the sentence where the cause/the result exists, putting the tail list into the temporary list if half of sentences capable of following the head list in the temporary list form the cause-effect relationship, and finally judging which is a cause list and which is a result list.

The targeted capacity expansion of the data set in step 7 specifically includes: after the model is trained, a plurality of pieces of data are taken randomly for detection, each result is analyzed and evaluated, and if the type of errors is common, a plurality of pieces of data are found for the type to supplement a training set; marking sentences with the prefixes of 'title', 'report' and 'discovery' as causality-free sentences and adding the causality-free sentences into a training set; if the error is one case, adding the error into the training set after correcting the error; the above steps are repeated for three times;

the hyper-parameter tuning specifically comprises the following steps: according to the BerT related paper, the recommended hyper-parameters are preferably adjusted to be, least-rate (Adam) 5e-5,3e-5,2e-5Number of epochs 2,3, 4; and obtaining a plurality of models according to the recommended hyper-parameters, and selecting the best model to improve the discrimination effect by comparing precision, call and f1 values of the models on the test set.

The invention also comprises a mixed causality extraction system oriented to the financial field, which comprises a training sentence expansion and model training module, a central word extraction module, a central word table searching module, a sentence division module, a central word position correction module, a causality identification module and a model tuning and optimization module which are sequentially connected.

The method has the advantages that the cause and effect extraction of the complex sentences in the financial field is realized, implicit cause and effect relations and explicit cause and effect relations are included, meanwhile, four cause and effect relation types including one cause, multiple causes, multiple effects, one cause, multiple effects and one cause and multiple effects can be extracted, future dynamic information in the field is provided for financial decision and other practical application requirements, and a new method is provided for solving the problems of high cost, low efficiency, high threshold, low timeliness and the like of enterprise risk prediction analysis.

Drawings

FIG. 1 is a general flow diagram of the process of the present invention.

FIG. 2 is a block diagram of the BERT model invoked by the present invention.

FIG. 3 is a diagram illustrating a specific flow of causal relationship extraction.

Detailed Description

The technical scheme of the invention is further explained by combining the attached drawings.

The invention relates to a causal relationship extraction method for the financial field, which comprises the following steps:

step 1: marking the sentences containing the causal relationship as 1 and the sentences not containing the causal relationship as 0 to obtain approximately equal positive and negative example sentences which are used as a training set for training the BERT model to obtain a trained BERT causal judgment model, and identifying the existence of the causal relationship of the input sentences.

Step 2: and (3) extracting the central words such as 'cause', 'cause' and the like from the existing cause and effect sentences to establish a central word list. And then synonym search is carried out, other words which can indicate the cause and effect relationship can be found and added into the central word list, and the central word library is expanded. And then manually distinguishing and labeling the common positions (three in the middle, in the front and in the back) and the common modes (two types of pre-cause and post-cause characterization) of all the central words.

And step 3: obtaining an input sentence, searching whether the sentence contains a central word or not according to a central word table, if so, inquiring the central word in the central word table to obtain the common position of the central word and the common mode (the following common reason or result). Such as "cause", which is located in the middle, followed by "cause" which is generally the result, and followed by "cause" which is generally the cause. If the central word contained in the word list is not found, and the central word is identified in the sentence according to the word segmentation, the position of the central word is in the middle of the reason and the result by default. If the input sentence does not contain the central word, the model is changed into three classifications (non-causal, pre-causal and post-causal), and then the partial sentences can be divided according to the arrangement and combination mode.

And 4, step 4: the sentence is divided into left and right sides by taking the central word as a boundary, and the sentence is divided according to punctuation marks. Then, sentence preprocessing is carried out, and since too short sentences like 'for example', 'at the same time' and the like generally cannot represent reasons or results, sentences with the length smaller than 4 are removed; in addition, as "investment advice: the deductive type qualified secondary capital tools are firstly introduced, the capital sufficiency rate is improved, so that the capital supplement pressure brought by gradual deduction of the bank due to the fact that no deductive or the secondary debt of the sublevel is listed as the unqualified capital tools is effectively relieved, and the bank obtains the capital required for coping with the risk weighted asset expansion caused by the new business development. In this example, the "investment advice" obviously cannot be used as a cause or result, i.e., the short sentence of all nouns cannot characterize the cause or result, and should be removed.

And 5: correcting the position of the central word:

a) the central word should be in the middle, but the sentence with no noun on the left will have its position reset to the front. Otherwise, it is set as the rear.

b) The correction is carried out when the central word is in the front and the right side has no two sentences with nouns, the right side has only one sentence with an existence entity and is set as the middle, and the right side has no sentence with an existence entity and is set as the back.

c) The central word is corrected when there are no two sentences with nouns on the back side and one sentence with an entity on the left side is set as the middle, and the sentence without the entity on the left side is set as the front.

Step 6: the following three situations with different positions of the central words are separately processed:

a) the central word is in the middle: find the nearest entities on the left and right sides, note as left and right, put into the left list and the right list respectively. And the rest sentences on the left side are sequentially combined with right and are placed in a trained BERT model to judge the causal relationship, if the causal relationship exists, the rest sentences on the left side are added into the left list, similarly, the rest sentences on the right side are sequentially combined with left and are placed in the model to judge the causal relationship, and if the causal relationship exists, the rest sentences on the right side are added into the right list. And finally, judging which is a reason list and which is a result list according to the common mode of the central words.

b) The central word is preceded: and recording an entity which is the nearest behind the central word as a head, adding the head into a head list, combining the rest sentences with the head in sequence, putting the head and the rest sentences into a trained BERT model for judging the cause-effect relationship, putting the head list into a tail list if the cause-effect relationship exists, otherwise putting the head list into a temporary list, wherein the temporary list is used for temporarily storing sentences which may be sentences in which the causes/results exist, and putting the head list into the head list if the cause-effect relationship exists and half sentences in the temporary list form the cause-effect relationship with the tail list, and finally judging which one of the head list and the tail list is a cause list and which one is a result list.

c) The core word is followed: and recording an entity nearest to the head of the central word as tail, adding the tail into a tail list, combining the rest sentences with the tail in sequence, putting the tail and the rest sentences into a trained BERT model for judging the cause-effect relationship, putting the tail list into a head list if the cause-effect relationship exists, otherwise putting the tail list into a temporary list, temporarily storing sentences possibly including the sentence where the cause/the result exists, putting the tail list into the temporary list if half of sentences capable of following the head list in the temporary list form the cause-effect relationship, and finally judging which is a cause list and which is a result list.

And 7: the causal discrimination result of the model is optimized by the following two ways:

a) and carrying out targeted capacity expansion on the data set. After the model is trained, 30 pieces of data are randomly detected, each result is analyzed and evaluated, and if the type of errors is common, 50 pieces of data are found for the type to supplement a training set. The experiments show that sentences prefixed by ' title ', ' report ', ' found ' and the like are easily mistaken for being causal, for example, example sentence ' reporter observation shows that the net profit increase of the first three quarters of multi-family insurance enterprises benefits from the profit increase of the equity market. The recognition result of the first result model shows that the sentence with the reason is 'income increase of the equity market', the sentence with the result is 'net profit increase of multi-family insurance enterprises in the first three quarters' and 'reporter observation discovery', and the 'reporter observation discovery' is wrongly recognized as causal relationship. Therefore, sentences with prefixes of 'title', 'report' and the like are found and marked to be added into the training set after being free of cause and effect. If the error is an example, the error is corrected and added to the training set. The above steps are repeated for three times.

b) And (6) optimizing the super-parameters. According to the BerT related paper, the recommended hyper-parameters are preferably adjusted to a lean-rate (Adam):5e-5,3e-5,2e-5Number of epochs:2,3, 4. And obtaining a plurality of models according to the recommended hyper-parameters, and selecting the best model to improve the discrimination effect by comparing precision, call and f1 values of the models on the test set.

The invention also comprises a mixed causality extraction system oriented to the financial field, which comprises a training sentence expansion and model training module, a central word extraction module, a central word table searching module, a sentence division module, a central word position correction module, a causality identification module and a model tuning and optimization module which are connected in sequence, wherein,

the training sentence expanding and model training module comprises: acquiring related statement expansion corpora from research and report and articles in the financial field, respectively labeling sentences with causal relationship and sentences without causal relationship, and training a BERT model by taking the sentences as a training set to judge the existence of the causal relationship;

the headword draws the module and includes: finding out the central word in the sentence, expanding a central word list, and labeling the common position and the common mode of the central word list;

the central word table searching module comprises: obtaining an input sentence, searching a central word list for the sentence with the central word, and obtaining the common position and the common mode of the central word; if no central word exists, the model is changed into three classifications (non-causal, ante-causal and post-causal), and then partial sentences are divided according to a permutation and combination mode

The sentence division module comprises: the sentences are divided into left and right sides by taking the central words as boundaries, and are divided into sentences according to punctuation marks, then corresponding sentence preprocessing is carried out, and redundant sentences are deleted;

the central word position correction module comprises: correcting the position of the central word with possible errors;

the cause and effect identification module includes: the method comprises the steps of carrying out separate processing on different positions of three central words with a central word in the middle, a central word in the front and a central word in the back, obtaining two lists by finding out an entity and judging whether two-two causal relations exist or not, realizing the identification of one-cause multi-effect, multi-cause one-effect, one-cause one-effect and multi-cause multi-effect, and finally judging to obtain one of the cause lists and one of the result lists;

the model tuning module comprises: and manually distinguishing and parameter tuning are carried out on the obtained preliminary model result, and the causal judgment effect of the model is improved through data set targeted capacity expansion and over-parameter tuning.

Causal extraction results example:

example 1:

inputting a sentence: reporters observed that the net profit increase of the first three quarters of multi-risk enterprises benefited from the revenue increase of the equity market.

The resulting reason sentence is: { "gain increase of equity market" }

The results obtained are: { "increase of net profit of former three quarter multi-family insurance enterprises" }

Aiming at the problems that the text processing difficulty in the financial field is high, the cause and effect relationship is determined to have disorder points, the extraction of central words is difficult and the like, the pre-training language model based on the BERT takes cause and effect linguistic data in the financial field as a training set training model, and meanwhile, a central word library is established, so that the extraction of mixed cause and effect relationship of the text facing the financial field is realized, and certain support is provided for the aspects of guiding event tracing, question answering, industry dynamic control and the like.

The invention has been illustrated by the above examples, but it should be noted that the examples are for illustrative purposes only and do not limit the invention to the scope of the examples. Although the invention has been described in detail with reference to the foregoing examples, it will be appreciated by those skilled in the art that: the technical solutions described in the foregoing examples can be modified or some technical features can be equally replaced; second, these modifications or substitutions do not depart from the scope of the present invention. The scope of the invention is defined by the appended claims and their equivalents.