Large model relation extraction method and system based on reason verification and feedback enhancement

Abstract

The invention provides a reason verification and feedback enhancement-based large model relation extraction method and system. The method comprises the following steps of collecting unbiased and biased reasons for a large language model by adopting a causal intervention and observation method; training a reason supervisor according to the unbiased reason and the biased reason obtained in the step 1; and iteratively verifying the output reason by adopting the trained reason supervisor, and providing feedback according to the verified deviation reason to correct the large language model so as to guide the large language model to make more accurate prediction. A reason supervisor is trained, prediction of the large language model can be verified through the reason supervisor, and deviation prediction of the large language model in a relation extraction task can be corrected through verification feedback.

Description

Technical Field

[0001] The present invention belongs to the technical field of large model relationship extraction, and in particular relates to a large model relationship extraction method and system based on reason verification and feedback enhancement. Background Art

[0002] With the rapid development of information technology and the Internet, from search engines to social media to smart assistants, big model relationship extraction technology provides us with smarter and more personalized services and experiences. In search engines, it can accurately understand the user's search intent and quickly present the most relevant search results, making information retrieval more efficient and convenient. On social media platforms, big model relationship extraction can not only strengthen the connection between users, but also discover users' interests and topic trends, promote the depth and breadth of social interaction, and provide users with a richer social experience. Through this technology, smart assistants and chatbots can not only understand users' needs more accurately, but also provide smarter and more intimate services and suggestions based on personalized preferences and habits. In the field of medical health, the application of big model relationship extraction technology can accelerate the acquisition of medical information and the making of medical decisions, which has positive significance for improving the quality of medical services.

[0003] The relation extraction task aims to extract semantic relations between entities in text, which is an important task in information extraction. Unlike previous fine-tuning strategies based on small language models, recent studies have leveraged the powerful instruction understanding ability and rich intrinsic knowledge of large language models to improve the performance of relation extraction. Although this approach has made significant progress, methods based on large language models may suffer from relation bias when forming relation extraction. In addition, existing relation extraction methods based on large language models focus on pre-selection of contextual demonstrations or instruction design to improve performance, so the current research on relation extraction based on large language models still lacks verification and feedback mechanisms to correct biased predictions. Summary of the invention

[0004] One object of the present invention is to address the deficiencies of the prior art and provide a large model relationship extraction method based on reason verification and feedback enhancement, which trains a reason supervisor and uses it not only to verify the predictions of the large language model, but also to provide new demonstration improvement feedback for correction during the reasoning process.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0006] A large model relationship extraction method based on reason verification and feedback enhancement includes the following steps:

[0007] Step 1: Use causal intervention and observation methods to collect unbiased and biased reasons for the large language model;

[0008] Step 2: Train the justification supervisor with the unbiased reasons and biased reasons obtained in step 1;

[0009] Step 3: Use the trained justification supervisor to iteratively verify the output justification and provide feedback to correct the large language model based on the verified deviation justification to guide the large language model to make more accurate predictions.

[0010] Furthermore, the implementation in step 1 includes two steps: inducing unbiased reasons and observing biased reasons.

[0011] Furthermore, the implementation methods of inducing unbiased reasons include:

[0012] Cut off the causal direction that makes the bias B affect the prediction Y, that is, only provide the text of the golden label in the input prompt, so that the golden label guides the large language model to generate the reason R through context learning. Observe that the reason generated by the large language model is

[0013] Intervene on the reason R, that is, only provide the observation reason in the input prompt The text is then used to guide the large language model to generate the label Y of the given text through context learning; if the observed value of Y is equal to the golden label y i , then as grounds for non-bias.

[0014] Furthermore, the implementation methods of observing the reasons for deviation include:

[0015] For a labeled sample {x i ,y i}, where x represents the input sample sentence and head and tail entities, y represents the relationship label between the head and tail entities, i represents the sample sequence number, and randomly selects a context example set D with different labels dii , where D dii From the labeled sample set, and D dii The label of each demonstration is not equal to the label y of the sample to be observed i ;

[0016] Set context example I to be from D dii Sample Among them, r u Indicates the explanation of the sample relationship label, j indicates the sample number, that is, only the sample is provided in the input prompt. Guide the large language model to generate predicted labels through contextual learning;

[0017] If the predicted labels of the observed large language model are inconsistent with the sample annotation labels, the reasons generated by the observed large language model are regarded as biased reasons.

[0018] Furthermore, in step 2, the contrastive learning method is used to train the reason supervisor so that the reason supervisor acquires the following capabilities:

[0019] Identify reasons for bias and non-bias;

[0020] Learn the differences in the various reasons for the deviations.

[0021] Furthermore, the specific method of training the reason supervisor includes:

[0022] Design positive and negative pairs for comparative training, where the unbiased reasons with the same golden relationship label and the biased reasons under the same bias are used as positive pairs, and the biased and unbiased reasons from the same sample are used as negative pairs; use the positive and negative pairs to train the reason supervisor so that the reason supervisor can distinguish between biased and unbiased reasons;

[0023] And according to different deviation situations, corresponding biased reasons are used to train the reason supervisor, so that the reason supervisor can distinguish different deviation situations and provide feedback according to the biased reasons during reasoning.

[0024] Furthermore, the loss of contrastive learning is calculated as follows:

[0025]

[0026] in, and represents two pairs of sets, and represents two sets of negative pairs, r1 and r2 represent the deviation reasons in the two sets of positive pairs, τ represents the temperature hyperparameter, and the dot product is used as the similarity function sim();

[0027] During the contrast training process, update R γ The parameter γ is used to make L cl minimize.

[0028] Further, the method of verifying the output reason using the trained reason supervisor includes:

[0029] The set of biased reasons R collected from step 1 b and the unbiased reason set R u Select the subset S related to the large language model prediction label y b and S u , which is then used as an anchor point to determine whether the current output justification is close to the biased or unbiased group, S b and Su The definition is as follows:

[0030] S b ={{r b ,y b}|{r b ,y b}∈R b ,y b =y}

[0031] S u ={{r u ,y u}|{r u ,y u}∈R u ,y u =y}

[0032] In the formula, r b Denotes the subset S of deviation reasons b The reason for the deviation, r u represents the unbiased reason subset S u The unbiased reason in y b Indicates the reason for deviation b The corresponding biased relationship type, y u represents an unbiased relationship type;

[0033] Then, the indicator score for judging whether the prediction reason r of the large language model is biased is calculated as follows:

[0034]

[0035] Among them, the dot product is used as the similarity function sim(), when p b When it is greater than 0, it means that the characteristic of r is better than S u The characteristic field is closer to S b , then r and the corresponding relationship prediction y are considered to be biased and need to be corrected.

[0036] Further, the method of providing feedback based on the verified deviation reasons to correct the large language model includes:

[0037] When the prediction reason r is verified to be biased, first in S b In the method of semantic similarity retrieval of reason text, the most similar deviation reason to r is selected, which is expressed as Defined as:

[0038]

[0039] k represents the search ranking range set in the similarity search process;

[0040] Then, from Dl Select and The labeled samples corresponding to the biased reasons in are used as feedback demonstrations D fb , defined as:

[0041]

[0042] in, Corresponding to the sample x i Reasons and predictions for deviations;

[0043] In the selection feedback demo D fb Afterwards, the large language model f θ By adding D fb As an example in context learning, we can regenerate the reason r and the predicted label y, that is, {r, y} = f θ (D fb ,x);

[0044] After generation, the aforementioned reason verification step will be re-executed, that is, the whole process will be iterated until the reason r generated by the large language model is verified to be unbiased in the reason verification step, and finally the corresponding unbiased relationship prediction y will be output.

[0045] Another object of the present invention is to provide a system for implementing the above-mentioned large model relationship extraction method based on reason verification and feedback enhancement, comprising:

[0046] A data collection module for collecting unbiased and biased justifications using causal intervention and observation methods for large language models;

[0047] A supervisor training module, used to train a reason supervisor based on the collected unbiased reasons and biased reasons;

[0048] The large language model correction module is used to iteratively verify the output reasons using the trained reason supervisor, and to correct the large language model by providing feedback based on the verified deviation reasons, so as to guide the large language model to make better predictions.

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows: in view of the problem that there is still a lack of verification and feedback mechanism for correcting biased predictions in the research on relation extraction of large language models, the present invention provides a relation extraction framework that checks whether there is a bias in the relation prediction of the large language model by verifying reasons, and provides automatic feedback for the relevance accordingly. The framework is used to solve two problems existing in the existing methods in relation extraction: one is that it is designed for other tasks, such as reasoning tasks, and the feedback goals are also tailored for these tasks, so it cannot be directly applied to the feedback framework design of the relation extraction task, and the other is that the existing methods do not include any demonstration in their feedback; the present invention proposes a novel automatic feedback framework, which trains a reason supervisor, and uses it not only to verify predictions, but also to correct the biased predictions of the large language model in the relation extraction task through verification feedback; in addition, the present invention proposes a causal intervention and observation method for collecting unbiased and biased reasons, thereby training the reason supervisor, and using the trained reason supervisor through the verification-feedback-correction process to iteratively improve the ability of the large language model in correcting biased predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A flowchart of a method for collecting unbiased and biased reasons for an embodiment of the present invention;

[0051] Figure 2 A diagram showing the steps of training a reason supervisor according to an embodiment of the present invention;

[0052] Figure 3 A diagram of the verification feedback correction process of the present invention using a justification supervisor during reasoning. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present invention will be described clearly and completely below in combination with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0055] The present invention will be further described below in conjunction with specific embodiments, but the present invention is not limited thereto.

[0056] Since training a reason supervisor requires collecting unbiased and biased reasons, and the purpose and method of collecting reasons in the existing verification method are not suitable for relationship extraction tasks, in view of this, an embodiment of the present invention provides a large model relationship extraction method based on reason verification and feedback enhancement, including the following steps:

[0057] Step 1: Use causal intervention and observation methods to collect unbiased and biased reasons for the large language model;

[0058] The implementation of this step includes: "inducing unbiased reasons" and "observing biased reasons". Figure 1 (a), the implementation method thereof includes:

[0059] Sub-step 1: Cut off the causal direction that may cause the bias B to affect the prediction Y, that is, do not give any text other than the golden label in the input prompt, and let the golden label guide the large language model to generate the reason R through context learning, formally expressed as the intervention operation do(Y) = y i and do(Y)→R, where y i It is said that the reason for observing the generation of large language models is

[0060] Sub-step 2: Perform a similar intervention operation on the reason R, that is, do not give any reasons other than observation in the input prompt The text outside is formally expressed as do(R)→Y, then guide the large language model to generate the label observation Y of the sample through context learning; if the observed value of Y is equal to the golden label y i , then as grounds for non-bias.

[0061] In the "Observe the reason for the deviation" step, as shown in the attached Figure 1 (b), the implementation thereof includes:

[0062] For a labeled sample {x i ,y i}(x represents the input sample sentence and head and tail entities, y represents the relationship label between the head and tail entities, and i represents the sample number). First, we randomly select a context example set D with different labels. dii , where D dii From the labeled sample set, and D dii The label of each demonstration is not equal to the label y of the sample to be observed i ;

[0063] Next, set context example I to be from D dii Sample (x represents the input sample sentence and head and tail entities, r u Indicates the explanation of the sample relationship label, y indicates the relationship label between the head and tail entities, and j indicates the sample number), that is, no information other than the sample is given in the input prompt. The text outside is formally represented as an intervention operation The large model is guided to generate predicted labels through context learning. At this time, if the predicted labels of the large language model are observed to be inconsistent with the sample labels, the reasons generated by the observed large language model are regarded as biased reasons. The unbiased reasons and biased reasons obtained in the above steps will be added to the unbiased set R u and the deviation set R b middle.

[0064] Step 2: Reason supervisor training. Train the reason supervisor with the unbiased reasons and biased reasons obtained in step 1;

[0065] like Figure 2 As shown, this step uses the unbiased reasons and the biased reasons obtained in step 1 as training data. At this stage, it is expected that the reason supervisor can verify whether the output reason is biased, and provide different feedback for different bias situations to correct the initial prediction. To achieve this, this embodiment uses contrastive learning to train the reason supervisor to obtain two capabilities: distinguishing biased and unbiased reasons and learning the differences between various biased reasons.

[0066] In this embodiment, two positive and negative pairs are designed for comparative training. Specifically, the unbiased reasons (e.g. ) and the reasons for the deviation in the same deviation situation (e.g. ) as two positive pairs. The biased and unbiased reasons (e.g. ) as negative pairs, and train the reason supervisor through positive and negative pairs to enable it to distinguish biased and unbiased reasons. Then, the corresponding biased reason pairs are used in different bias situations (e.g. and ) train a justification supervisor such that the trained justification supervisor can distinguish different biased cases and provide feedback according to the biased justifications during inference.

[0067] In this embodiment, the contrast loss is calculated as follows:

[0068]

[0069] in and represents two pairs of sets, and represents two sets of negative pairs, r1 and r2 represent the deviation reasons in two sets of positive pairs. This embodiment uses the dot product as the similarity function sim() and adds a temperature hyperparameter τ to pay more attention to difficult pairs. In the rational comparison training process, update R γ The parameter γ makes L cl minimize.

[0070] Step 3: Use the trained justification supervisor to iteratively verify the output justification and provide feedback to correct the large language model based on the verified deviation justification to guide the large language model to make more accurate predictions.

[0071] In the reason verification and feedback stage, this embodiment uses the trained reason supervisor R γ to iteratively verify the output reasons and provide feedback based on the verified deviation reasons to guide the large language model to make better predictions. Figure 3 As shown, in this embodiment, it is divided into three steps, and the implementation process of each step is described below.

[0072] 1) Reason verification

[0073] In order to verify whether the reason prediction is biased, we need to get the deviation set R from step 1 b and the unbiased set R u Select the subset S that is relevant to the prediction y of the large prediction model b and S u , which is then used as an anchor to determine whether the current output justification is close to the biased or unbiased group. b and S u The definition is as follows:

[0074] S b ={{r b ,y b}|{r b ,y b}∈R b ,y b =y}

[0075] S u ={{r u ,y u}|{r u ,y u}∈R u ,y u =y}

[0076] In the formula, r b Denotes the subset S of deviation reasons b The reason for the deviation, r u represents the unbiased reason subset S u The unbiased reason in yb Indicates the reason for deviation b The corresponding biased relationship type, y u Represents the unbiased relationship type (i.e., the true label of the sample);

[0077] Then, the index score for judging whether the large language model relation extraction prediction reason r is biased is calculated as follows:

[0078]

[0079] Among them, the similarity function sim() is defined in step 2. When p b When it is greater than 0, it means that the characteristic of r is better than S u The characteristic field is closer to S b feature field, which means that r and the corresponding relationship prediction y should be considered biased and feedback is needed to correct them.

[0080] 2) Feedback demonstration retrieval

[0081] Once the output reason r is verified to be biased, a new set of contextual demonstrations needs to be retrieved based on the features of r to guide the large language model to make better predictions. Specifically, first, in S b In the method of semantic similarity retrieval of reason text, the most similar deviation reason to r is selected, which is expressed as Defined as:

[0082]

[0083] k represents the search ranking range set in the similarity search process;

[0084] Then, from D l Select and The labeled samples corresponding to the biased reasons in are used as feedback demonstrations D fb , defined as:

[0085]

[0086] in, Corresponding to the sample x i Reasons and predictions for deviations.

[0087] 3) Correction through contextual learning

[0088] In the selection feedback demo D fb Afterwards, we use the large language model f θ By adding D fb As an example in context learning, we can regenerate the reason r and the predicted label y, that is, {r, y} = f θ (D fb,x). After generation, the aforementioned reason verification step will be re-executed, that is, the whole process will be iterated until the reason r generated by the large model is verified to be unbiased in the reason verification step, and the corresponding unbiased relationship prediction y will be output eventually. For example, the input sample "sentence: Allen has worked at Apple for 3 years. Head entity: 'Allen'; tail entity: 'Apple'", without iterative correction, the large language model may output an incorrect relationship prediction "head entity and tail entity relationship: 'person: creator'" (this may be due to the interference of the large model training corpus or the initial wrong context example). After iterative correction, the large language model gives the correct prediction "head entity and tail entity relationship: 'person: employee'".

[0089] The embodiment of the present invention further provides a system for implementing the above-mentioned large model relationship extraction method based on reason verification and feedback enhancement, comprising:

[0090] A data collection module for collecting unbiased and biased justifications using causal intervention and observation methods for large language models;

[0091] A supervisor training module, used to train a reason supervisor based on the collected unbiased reasons and biased reasons;

[0092] The large language model correction module is used to iteratively verify the output reasons using the trained reason supervisor, and to correct the large language model by providing feedback based on the verified deviation reasons, so as to guide the large language model to make better predictions.

[0093] The above are only preferred embodiments of the present invention, and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the contents of the specification of the present invention should be included in the protection scope of the present invention.

Claims

1. A large model relationship extraction method based on reason verification and feedback enhancement, characterized in that: The following steps are involved: Step 1: Use causal intervention and observation methods to collect unbiased and biased reasons for the large language model; Step 2: Train the justification supervisor with the unbiased reasons and biased reasons obtained in step 1; Step 3: Use the trained justification supervisor to iteratively verify the output justification and provide feedback to correct the large language model based on the verified deviation justification to guide the large language model to make more accurate predictions.

2. The large model relationship extraction method based on reason verification and feedback enhancement according to claim 1 is characterized in that: The implementation in step 1 consists of two steps: inducing unbiased reasons and observing biased reasons.

3. The large model relationship extraction method based on reason verification and feedback enhancement according to claim 2 is characterized in that: Ways to induce unbiased reasons include: Cut off the causal direction that makes the bias B affect the prediction Y, that is, only provide the text of the golden label in the input prompt, so that the golden label guides the large language model to generate the reason R through context learning. Observe that the reason generated by the large language model is Intervene on the reason R, that is, only provide the observation reason in the input prompt The text is then used to guide the large language model to generate the label Y of the given text through context learning; if the observed value of Y is equal to the golden label y i , then as grounds for non-bias.

4. The large model relationship extraction method based on reason verification and feedback enhancement according to claim 2 is characterized in that: Ways to implement observation biased justification include: For a labeled sample {x i ,y i }, where x represents the input sample sentence and head and tail entities, y represents the relationship label between the head and tail entities, i represents the sample sequence number, and randomly selects a context example set D with different labels dii , where D dii From the labeled sample set, and D dii The label of each demonstration is not equal to the label y of the sample to be observed i ; Set context example I to be from D dii Sample Among them, r u Indicates the explanation of the sample relationship label, j indicates the sample number, that is, only the sample is provided in the input prompt. Guide the large language model to generate predicted labels through contextual learning; If the predicted labels of the large language model are observed to be inconsistent with the sample annotation labels, the reasons generated by the observed large language model are regarded as biased reasons.

5. The large model relationship extraction method based on reason verification and feedback enhancement according to claim 1 is characterized in that: In step 2, the contrastive learning method is used to train the reason supervisor so that the reason supervisor can obtain the following capabilities: Identify reasons for bias and non-bias; Learn the differences in the various reasons for the deviations.

6. The large model relationship extraction method based on reason verification and feedback enhancement according to claim 5 is characterized in that: The specific methods for training the justification supervisor include: Design positive and negative pairs for comparative training, where the unbiased reasons with the same golden relationship label and the biased reasons under the same bias are used as positive pairs, and the biased and unbiased reasons from the same sample are used as negative pairs; use the positive and negative pairs to train the reason supervisor so that the reason supervisor can distinguish between biased and unbiased reasons; And according to different deviation situations, corresponding biased reasons are used to train the reason supervisor, so that the reason supervisor can distinguish different deviation situations and provide feedback according to the biased reasons during reasoning.

7. The large model relationship extraction method based on reason verification and feedback enhancement according to claim 6 is characterized in that: The loss for contrastive learning is calculated as follows: in, and represents two pairs of sets, and represents two sets of negative pairs, r1 and r2 represent the deviation reasons in the two sets of positive pairs, τ represents the temperature hyperparameter, and the dot product is used as the similarity function sim(); During the contrast training process, update R γ The parameter γ is used to make L ci minimize.

8. The large model relationship extraction method based on reason verification and feedback enhancement according to claim 1 is characterized in that: Methods for verifying output reasons using a trained justification supervisor include: The set of biased reasons R collected from step 1 b and the unbiased reason set R u Select the subset S related to the large language model prediction label y b and S u , which is then used as an anchor point to determine whether the current output justification is close to the biased or unbiased group, S b and S u The definition is as follows: S b ={{r b ,and b }|{r b ,and b }∈R b ,and b =and} S u ={{r u ,and u }|{r u ,and u }∈R u ,and u =and} In the formula, r b Denotes the subset S of deviation reasons b The reason for the deviation, r u represents the unbiased reason subset S u The unbiased reason in y b Indicates the reason for deviation b The corresponding biased relationship type, y u represents an unbiased relationship type; Then, the indicator score for judging whether the prediction reason r of the large language model is biased is calculated as follows: Among them, the dot product is used as the similarity function sim(), when p b When it is greater than 0, it means that the characteristic of r is better than S u The characteristic field is closer to S b , then r and the corresponding relationship prediction y are considered to be biased and need to be corrected.

9. The large model relationship extraction method based on reason verification and feedback enhancement according to claim 8 is characterized in that: Methods for correcting large language models by providing feedback based on validated bias justifications include: When the prediction reason r is verified to be biased, first in S b In the method of semantic similarity retrieval of reason text, the most similar deviation reason to r is selected, which is expressed as Defined as: k represents the search ranking range set in the similarity search process; Then, from D l Select and The labeled samples corresponding to the biased reasons in are used as feedback demonstrations D fb , defined as: in, Corresponding to the sample x i Reasons for and predictions of deviations; In the selection feedback demo D fb Afterwards, the large language model f θ By adding D fb As an example in context learning, we can regenerate the reason r and the predicted label y, that is, {r, y} = f θ (D fb ,x); After generation, the aforementioned reason verification step will be re-executed, that is, the whole process will be iterated until the reason r generated by the large language model is verified to be unbiased in the reason verification step, and finally the corresponding unbiased relationship prediction y will be output.

10. A system for implementing the large model relationship extraction method based on reason verification and feedback enhancement as described in any one of claims 1 to 9, characterized in that: include: A data collection module for collecting unbiased and biased justifications using causal intervention and observation methods for large language models; A supervisor training module, used to train a reason supervisor based on the collected unbiased reasons and biased reasons; The large language model correction module is used to iteratively verify the output reasons using the trained reason supervisor, and to correct the large language model by providing feedback based on the verified deviation reasons, so as to guide the large language model to make better predictions.

Images