Knowledge acquisition method and system based on causal inference

Abstract

The invention discloses a knowledge acquisition method and system based on causal inference, and belongs to the field of natural language processing. The method comprises the following steps: firstly, constructing a causal graph by identifying a causal dependency relationship between events in a text; secondly, inputting a process text containing missing steps and candidate steps, extracting events, judging whether a causal relationship exists or not through a causal graph, and taking related causal information as additional features; and respectively calculating probabilities that the candidate steps contain and do not contain the dominant entity, coding through the two models, and calculating probabilities that the candidate steps are used as missing steps in combination with a multi-layer perceptron. And finally, comparing the probabilities of all the candidate steps with a negative class threshold, and determining the candidate step which most accords with the condition or judging that the correct step does not exist. According to the method, the causal relationship between events is additionally introduced, so that the more accurate knowledge acquisition capability is realized.

Description

Technical Field

[0001] The present invention belongs to the field of natural language processing, and specifically relates to a knowledge acquisition method and system based on causal inference. Background Art

[0002] Knowledge acquisition methods aim to mine useful information from large amounts of text, such as events that meet specific constraints. Among them, how to extract missing events in process text is one of the difficulties in the current field of knowledge acquisition.

[0003] Process text is a description of a process, that is, the movement or change of one or more entities over a period of time. Process text is widely present in daily life. Popular science articles, product manuals, recipes, etc. that describe a scientific phenomenon are relatively standard process texts. These texts are usually composed of sentences arranged in chronological order, and each sentence represents a specific event. Therefore, process text can be regarded as a sequence of events related to one or more entities. Process text often has missing problems. Using knowledge acquisition methods to extract missing events and supplement process text can help downstream systems better use process text to complete planning, question-answering and other tasks.

[0004] In process texts, different events are arranged in chronological order. There is usually a causal relationship between events related to the same entity, that is, the necessary condition for the subsequent event to occur is that the prerequisite event has been completed. However, existing knowledge acquisition methods usually ignore this causal relationship and only use the semantic information of the text to extract, which will reduce the accuracy of knowledge acquisition. In order to obtain a more accurate knowledge acquisition method, it is necessary and effective to introduce the causal relationship between events. Summary of the invention

[0005] In view of the technical problems existing in the prior art, the purpose of the present invention is to provide a more effective knowledge acquisition method and system. The method combines causal inference technology. First, a causal graph between different events is constructed through a large number of existing process texts. When extracting a missing event in a process text, the causal relationship between the event and other events in the process text is fully considered, thereby achieving more accurate knowledge acquisition capabilities.

[0006] The technical solution adopted by the present invention to achieve the above-mentioned purpose is:

[0007] A knowledge acquisition method based on causal inference, the steps of which include:

[0008] (1) Collect process texts, identify the causal dependencies between events in the process texts, and construct a causal graph of events;

[0009] (2) Input a process text with missing steps and the corresponding candidate steps, extract events in the process text and candidate steps, and determine whether there is a causal relationship between the candidate event and the existing events in the process text based on the extracted events and the causal graph. If so, use the event as an additional feature;

[0010] (3) Calculate the probability of each candidate step containing a dominant entity and not containing a dominant entity; for the case where the candidate step contains a dominant entity, encode the process text containing the missing step and the concatenation result of the candidate step through a Lattice-BERT model, and then use a multi-layer perceptron to calculate the probability that the candidate step is the missing step in the process text; for the case where the candidate step does not contain a dominant entity, encode the concatenation result of the process text containing the missing step, the candidate step and the embedding representation of the additional feature through another Lattice-BERT model, and then use a multi-layer perceptron to calculate the probability that the candidate step is the missing step in the process text; combine all the probabilities of these two cases to obtain the final probability that each candidate step is the missing step in the process text;

[0011] (4) Determine the difference between the final probability corresponding to each candidate step and the preset negative threshold. If the final probabilities corresponding to all candidate steps are smaller than the negative threshold, then all candidate steps are determined not to be missing steps in the process text. Otherwise, the candidate step corresponding to the maximum final probability greater than the negative threshold is determined to be a missing step in the process text.

[0012] Furthermore, the step of identifying the causal dependency between events in step (1) includes:

[0013] Taking event a as the outcome, find all other events a′ that appear before event a in all process texts, and use the propensity score matching method to estimate the average causal effect ψ of events a and a′ a′,a , if ψ a′,a Greater than a preset threshold τ c , then event a′ is determined to be the cause event of event a;

[0014] Taking event a as the cause, find all other events a′ that appear after event a in all process texts, and use the propensity score matching method to estimate the average causal effect ψ of events a and a′ a′,a , if ψ a′,a Greater than a preset threshold τ c , then event a′ is determined to be the result event of event a;

[0015] The causal dependency between events is determined based on the determined cause events and result events.

[0016] Furthermore, in step (2), verbs in the process text and candidate steps are extracted, and entities and tools accompanying the action processing are extracted, and the extracted verbs, entities and tools constitute an event.

[0017] Furthermore, in step (2), each sentence in the process text is subjected to syntactic analysis, and each verb on the syntactic tree is regarded as an event; if the verb has a nominal child node on the syntactic tree, then the child node is the entity and tool of the action.

[0018] Furthermore, the step of determining whether there is a causal relationship between the candidate event and the existing event in the process text in step (2) includes:

[0019] Extracting events from the process text includes extracting all event sets A contained in the steps before the missing step before , and the set A of all events contained in the step after the missing step after ;

[0020] Extract event a in the candidate step choice , find a from the causal graph choice Corresponding cause event set and the resulting event collection

[0021] A before and Perform event matching. If the match is successful, it is considered that there is a cause event of the candidate step in the process text;

[0022] Will and A after Perform event matching. If the match is successful, it is considered that there is a result event of the candidate step in the process text;

[0023] The above matching events are used as additional features.

[0024] Furthermore, the step of calculating the probability of each candidate step containing the dominant entity and not containing the dominant entity in step (3) includes:

[0025] Identify the entities contained in all steps in the process text and use them as candidate leading entities;

[0026] Using the pre-trained dominant entity recognition model, predict the probability P of each candidate dominant entity becoming the dominant entity e ;

[0027] Each entity e in the candidate step c Match with all candidate dominant entities, and select from the candidate dominant entities the entities that match each entity e in the candidate step c The entity with the highest matching degreep The entity pair (e c ,e p ), each entity e in the candidate step c Corresponding to an entity pair (e c ,e p );

[0028] For each entity pair (e c ,e p )), where entity e p The corresponding probability P e As an entity c As the probability of the dominant entity; all entities have probability P e The maximum probability among them is taken as the probability P(ing∈C i );

[0029] The probability that each candidate step does not contain the dominant entity

[0030] Furthermore, the additional features in step (3) are encoded through a causal prefix encoder to generate corresponding embedded representations, and the embedded representations are introduced into another Lattice-BERT model using soft prompts during concatenation.

[0031] Furthermore, in step (3), all the probabilities of the two cases are combined to obtain the final probability that each candidate step is a missing step in the process text:

[0032]

[0033] Among them, ing represents the dominant entity, P(ing∈C i ) represents candidate step C i The probability of containing a dominant entity; Represents candidate step C i The probability of not containing the dominant entity; P(C i |ing∈C i ) indicates that in candidate step C i Under the condition of containing the dominant entity, candidate step C i as the probability of missing steps; Indicates that in candidate step C i Without the dominant entity, candidate step C i as the probability of missing a step.

[0034] Furthermore, the negative class threshold in step (4) is set to 0.4.

[0035] A knowledge acquisition system based on causal inference includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0036] The beneficial effects achieved by the present invention are as follows:

[0037] 1.) The present invention constructs a causal graph of events through a causal estimation method based on propensity score matching, accurately captures the causal dependencies between events, and avoids the erroneous event associations that may be caused by traditional methods that rely solely on time sequence or context information.

[0038] 2.) In the process of knowledge acquisition, the causal guidance of the present invention effectively verifies the rationality of the candidate steps, ensuring that when completing the missing steps in the process text, the inserted steps have causes and consequences that conform to the causal logic, thereby improving the correctness of the reasoning results.

[0039] 3.) The present invention introduces a dominant entity recognition model to evaluate whether a candidate step contains a dominant entity, thereby effectively improving the accuracy of candidate step matching, especially when the missing step is related to the dominant entity, further improving the accuracy of system judgment.

[0040] 4.) Based on the traditional knowledge acquisition method, the present invention additionally introduces the causal relationship between events. By constructing a causal graph with a large amount of process text, causal guidance is added to the knowledge acquisition process, which enhances the model's understanding of causal information and improves the accuracy of missing step prediction and knowledge acquisition.

[0041] 5.) The present invention sets a negative threshold to accurately handle the situation of "no correct answer", avoiding the system's erroneous prediction steps, thereby further improving the reliability and intelligence level of the overall prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Flow chart of the knowledge acquisition method based on causal inference in an embodiment. DETAILED DESCRIPTION

[0043] The present invention is further described in detail below in conjunction with the accompanying drawings. The embodiments given are only used to explain the present invention but not to limit the scope of the present invention.

[0044] This embodiment specifically discloses a knowledge acquisition method based on causal inference, and its main process is as follows: Figure 1 The specific processing steps are described as follows.

[0045] 1. Construction of causal graph:

[0046] In a reasonable procedure text, the order of events is not arbitrary, because events affect the state of entities, such as shape and temperature. This method not only generates the necessary events, but also puts them in the right place. Specifically, the correctness of the position is judged according to the order constraint, that is, checking which events should be placed before the inserted event and which events should be placed after the inserted event. That is, for each event a to be inserted, it should be placed in the cause event set After and Result event collections Before. Order constraints can be viewed as causal dependencies: The event in is the cause of a, The action in is the result of a.

[0047] This method uses sufficient process text collected in advance and a causal estimation technique based on the Propensity Score Matching method to identify the causal dependency between events and construct a causal graph of events. Specifically, for two events a and a′ that have appeared in the recipe data set, the average causal effect ψ between them is a′ ,a)) can be calculated by the following formula:

[0048]

[0049] in represents the instance that is most similar to i in i's opposite group (here the instance is a specific recipe). i represents the value of the outcome of instance i, a′ i represents the value of the intervention measure (Treatment) of instance i, z i represents the value of the confusion factor (Confounder) of instance i, L(z i ) represents the propensity score (Propensity Score) of instance i.

[0050] Propensity score L(z) = P(a′) i =1|Z=z) is the intervention measure a′ given a set of observable confounding factors z. i The probability of taking the value of 1. In this method, the events that appear before events a and a' in any process text in the set are regarded as potential confusion factors, and the confusion factor set Z is constructed. For a specific process text instance i, by detecting whether the event in Z appears in i, the confusion factor z of instance i can be obtained. i This method uses the logistic regression method to fit the L(z) function.

[0051] Calculate the average causal effect ψ of events a and a' a′,a After that, if ψ a′,a >τ c , then put a' as the cause event into According to the preliminary experimental results, this embodiment will c Set to 0.1. The construction of is similar, that is, event a is taken as the cause, and every other event a′ that appears after a in all process texts is exhaustively taken as a candidate result, and then the above propensity score matching process is repeated to determine the result event and put it into middle.

[0052] 2. Cause and effect guidance:

[0053] After constructing the causal diagram of the event in the previous step, when processing the specific process text later, this causal diagram can be used to determine whether the candidate options are reasonable.

[0054] After receiving the process text and candidate steps (Candidate Step), they are first normalized into event sequences. Each event can be formally written as (v, igs, tools), that is, a verb v is the core, usually accompanied by the entity igs and tools tools for action processing. Specifically, this method will perform syntactic analysis on each sentence in the process text and treat each verb on the syntax tree as an event. If the verb has noun child nodes on the syntax tree, then these child nodes are the entities and tools of the action. For example, the phrase "cut onions" will be converted into actions (cut, onions, ).

[0055] Enter a process text with a missing step and the corresponding candidate step, and record the event in the candidate step as a choice , in the input process text, the set of all events in the step before a missing step is A before , the set of all events contained in the step after the missing step is A after In the causal graph constructed in the previous step, find a choice Corresponding cause event set and the resulting event collection In order to explore whether there is a potential cause event before the candidate step, before and Internal events are matched. As long as there is a match, it is considered that the original process text contains the cause event of the current candidate step. It can be formally written as:

[0056]

[0057] The same method can be used to determine whether there is a result event for the current candidate step in the original process text:

[0058]

[0059] I cause and I effect It will be used as an additional feature in the regression prediction step below.

[0060] 3. Entity Guidance:

[0061] The missing step in the process text is likely to contain entities that change state. In other words, if the candidate step does not contain a dominant entity, then the possibility of it being a missing step in the input process text will be significantly reduced. In order to model this dependency, the candidate step C i The probability of being the correct answer (i.e. missing step) is P(C i ) is split into two parts including the dominant entity and excluding the dominant entity, as shown in the following formula:

[0062]

[0063] The above formula can be further expressed as the following expansion:

[0064]

[0065] Among them, ing represents the dominant entity, P(C i ,ing∈C i ) represents the candidate step C containing the dominant entity i As the probability of the correct answer; Denotes a candidate step C that does not contain a dominant entity i As the probability of the correct answer; P(ing∈C i ) represents candidate step C i The probability of containing a dominant entity; Represents candidate step C i The probability of not containing the dominant entity; P(C i |ing∈C i ) indicates that in candidate step C i Under the condition of containing the dominant entity, candidate step C i as the probability of missing steps; Indicates that in candidate step C i Without the dominant entity, candidate step C i as the probability of missing a step.

[0066] P(ing∈C i) is calculated as follows: First, identify the entities involved in all steps in the process text and list them as candidate dominant entities. Then, apply the pre-trained dominant entity recognition model, that is, the model trained based on Lattice-BERT, to quantitatively evaluate the possibility of each candidate dominant entity becoming a dominant entity, and obtain the probability P of becoming a dominant entity. e Then, for candidate step C i Entity e in c An analysis is performed to determine whether these entities match the entities in all steps of the process text. The matching uses a comparative analysis method to compare the entities in the candidate step with the entities in all steps of the process text to find the entity pairs with the highest matching degree (e c ,e p )). For each selected entity pair (e c ,e p )), the probability P of the candidate dominant entity becoming the dominant entity e As the probability that the corresponding entity in the candidate step is the dominant entity. Finally, since the candidate step may contain multiple entities, it is necessary to consider the possibility that all entities in the candidate step are the dominant entity. To this end, a method of maximizing the probability is adopted, that is, selecting the maximum value from the probabilities of all matched entities as the dominant entity. This maximum value represents the comprehensive probability that the candidate step contains the dominant entity, that is, P(ing∈C i ).

[0067] For candidate step C i Two different Lattice-BERT models are used to model the two cases of containing and excluding the dominant entity. and P(C i |ing∈C i ).

[0068] For candidate step C i In the case where the leading entity is not included in the sentence, the present invention concatenates the process text containing the missing steps and the candidate steps as the input of the Lattice-BERT model. Then the sentence representation encoded by the model is input into a multi-layer perceptron (MLP) for regression prediction to obtain the probability

[0069] And for candidate step C i In addition to the method of concatenating the process text and the candidate steps, an additional feature I is introduced into the input. cause and I effect, that is, whether there is a cause event or result event of the event in the candidate step in the current process text. This feature is calculated in the previous causal guidance step. The feature is introduced by using the Soft)prompt method. Soft)prompt is a technique in natural language processing (NLP) for guiding large pre-trained language models (such as the GPT series). This method involves creating and using trainable embeddings as part of the model input to achieve specific guidance of the model behavior. Specifically, the I obtained in the causal guidance step is used cause and I effect As input, put it into a causal prefix encoder Causal)Prefix)Emb to get the corresponding embedding representation, considering I cause and I effect The values ​​of are all 0 or 1, so there are four different embedding representations, namely (I cause ,I effect )=(0,0),(0,1),(1,0) or (1,1). The corresponding embedding representation is concatenated with the process text and the word representation of the candidate step, and then put into another Lattice-BERT model for encoding. The sentence representation obtained after encoding is passed through an MLP for regression prediction to obtain the probability P(C i |ing∈C i ).

[0070] According to the four probabilities obtained by entity guidance and regression prediction, substituting them into the above expansion formula, we can calculate the final probability P(C i ).

[0071] 4. NA result prediction:

[0072] In the previous step, the probability of each candidate step being the correct answer is calculated, but this does not apply to the candidate step of the NA (Not Applicable) type "None of the above answers are correct", that is, no candidate step is the correct answer (i.e., a missing step). Therefore, this method needs to handle this type separately during training and prediction. Specifically, this method sets a negative class threshold. When the probability of all non-NA candidate steps being the correct answer is less than this threshold, the model directly predicts "None of the above answers are correct." When and only when there is a candidate step whose probability of being the correct answer is greater than this threshold, the candidate step is used as the final prediction result. When the probability of multiple candidate steps being the correct answer is greater than this threshold, the one with the largest predicted probability among all candidate steps is taken as the final prediction result. The negative class threshold is used as a hyperparameter, and this embodiment empirically sets it to 0.4.

[0073] Experimental test - knowledge extraction ability test:

[0074] The test data set used in this test is an event extraction data set based on process text. This data set is in the form of multiple-choice questions, and the missing event in the missing process text is selected from four candidate event options. The test indicator is the extraction accuracy, that is, the frequency of the option with the highest prediction score being the standard answer. The test results are shown in Table 1.

[0075] Table 1) Effect of knowledge extraction methods on event extraction dataset based on process text

[0076] method Extraction accuracy Randomly select answer 25％ Baseline method (Lattice-Bert) 62.2％ The present invention 81.8％

[0077] From the test results in Table 1, it can be seen that the extraction accuracy of the knowledge extraction method based on causal inference of the present invention is as high as 81.8%, which is significantly improved compared with the accuracy of 25% of the baseline method of randomly selecting answers and the accuracy of 62.2% of the baseline method that also uses the Lattice-BERT neural network but does not introduce causal inference.

[0078] Although the specific embodiments of the present invention are disclosed for the purpose of illustration, the purpose is to help understand the content of the present invention and implement it accordingly, those skilled in the art will understand that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the best embodiment, and the scope of the present invention is subject to the scope defined in the claims.

Claims

1. A knowledge acquisition method based on causal inference, characterized in that the steps include: (1) Collect process texts, identify the causal dependencies between events in the process texts, and construct a causal graph of events; (2) Input a process text with missing steps and the corresponding candidate steps, extract events in the process text and candidate steps, and determine whether there is a causal relationship between the candidate event and the existing events in the process text based on the extracted events and the causal graph. If so, use the event as an additional feature; (3) Calculate the probability of each candidate step containing a dominant entity and not containing a dominant entity; for the case where the candidate step contains a dominant entity, encode the process text containing the missing step and the concatenation result of the candidate step through a Lattice-BERT model, and then calculate the probability that the candidate step is the missing step in the process text through a multi-layer perceptron; for the case where the candidate step does not contain a dominant entity, encode the concatenation result of the process text containing the missing step, the candidate step and the embedding representation of the additional feature through another Lattice-BERT model, and then calculate the probability that the candidate step is the missing step in the process text through a multi-layer perceptron; Combine all the probabilities for these two cases to get the final probability that each candidate step is the missing step in the process text; (4) Determine the difference between the final probability corresponding to each candidate step and the preset negative threshold. If the final probabilities corresponding to all candidate steps are smaller than the negative threshold, then all candidate steps are determined not to be missing steps in the process text. Otherwise, the candidate step corresponding to the maximum final probability greater than the negative threshold is determined to be a missing step in the process text.

2. The method according to claim 1, characterized in that The steps of identifying causal dependencies between events in step (1) include: Taking event a as the outcome, find all other events a′ that appear before event a in all process texts, and use the propensity score matching method to estimate the average causal effect ψ of events a and a′ a′,a , if ψ a′,a Greater than a preset threshold τ c , then event a′ is determined to be the cause event of event a; Taking event a as the cause, find all other events a′ that appear after event a in all process texts, and use the propensity score matching method to estimate the average causal effect ψ of events a and a′ a′,a , if ψ a′,a Greater than a preset threshold τ c , then event a′ is determined to be the result event of event a; The causal dependency between events is determined based on the determined cause events and result events.

3. The method according to claim 1, characterized in that In step (2), verbs in the process text and candidate steps are extracted, and entities and tools accompanying the action processing are extracted. The extracted verbs, entities and tools constitute an event.

4. The method according to claim 1, characterized in that In step (2), each sentence in the process text is subjected to syntactic analysis, and each verb on the syntactic tree is regarded as an event; if the verb has a nominal child node on the syntactic tree, then the child node is the entity and tool of the action.

5. The method according to claim 1, characterized in that The step of determining whether there is a causal relationship between the candidate event and the existing event in the process text in step (2) includes: Extracting events from the process text includes extracting all event sets A contained in the steps before the missing step before , and the set A of all events contained in the step after the missing step after ; Extract event a in the candidate step choice , find a from the causal graph choice Corresponding cause event set and the resulting event collection A before and Perform event matching. If the match is successful, it is considered that there is a cause event of the candidate step in the process text; Will and A after Perform event matching. If the match is successful, it is considered that there is a result event of the candidate step in the process text; The above matching events are used as additional features.

6. The method according to claim 1, characterized in that The step of calculating the probability of each candidate step containing the dominant entity and not containing the dominant entity in step (3) includes: Identify the entities contained in all steps in the process text and use them as candidate leading entities; Using the pre-trained dominant entity recognition model, predict the probability P of each candidate dominant entity becoming the dominant entity e ; Each entity e in the candidate step c Match with all candidate dominant entities, and select from the candidate dominant entities the entities that match each entity e in the candidate step c The entity with the highest matching degree p The entity pair (e c ,e p ), each entity e in the candidate step c Corresponding to an entity pair (e c ,e p ); For each entity pair (e c ,e p ), where entity e p The corresponding probability P e As an entity c As the probability of the dominant entity; all entities have probability P e The maximum probability among them is taken as the probability P(ing∈C i ); The probability that each candidate step does not contain the dominant entity 7. The method according to claim 1, characterized in that In step (3), the additional features are encoded through the causal prefix encoder to generate the corresponding embedded representation, and the embedded representation is introduced into another Lattice-BERT model using soft prompts during splicing.

8. The method according to claim 1, characterized in that In step (3), all the probabilities of the two cases are combined to obtain the final probability that each candidate step is a missing step in the process text: Among them, ing represents the dominant entity, P(C i ,ing∈C i ) represents the candidate step C containing the dominant entity i As the probability of the correct answer; Denotes a candidate step C that does not contain a dominant entity i As the probability of the correct answer; P(ing∈C i ) represents candidate step C i The probability of containing a dominant entity; Represents candidate step C i The probability of not including a dominant entity; P(C i |ing∈C i ) indicates that in candidate step C i Under the condition of containing the dominant entity, candidate step C i as the probability of missing steps; Indicates that in candidate step C i Without the dominant entity, candidate step C i as the probability of missing a step.

9. The method according to claim 1, characterized in that In step (4), the negative class threshold is set to 0.

4.

10. A knowledge acquisition system based on causal inference, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method according to any one of claims 1 to 9 when executing the computer program.

Images