Method, device and medium for improving text recall effect based on keyword extraction

Abstract

The application discloses a method, device and medium for improving text recall effect based on keyword extraction, comprising: inputting a query text into a pre-trained keyword extraction model, performing word vector embedding based on an ERNIE underlying model, and performing classification prediction through a full connection layer to extract entities and entity categories in the text as a target keyword set; performing a plus word mapping and a must word mapping on the extracted keywords, and combining and constructing the keywords into a text recall request of a search engine after fine-grained word segmentation processing and meeting a preset entity coverage ratio condition; executing the text recall request, and outputting a recall result according to a text recall score. The application extracts entities and entity categories in the query text accurately, and constructs a retrieval condition in combination with a mapping mechanism meeting the coverage condition, effectively avoiding the loss of core entities in the text recall process, and significantly improving the accuracy of the retrieval result.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and information retrieval technology, specifically to a method, device, and medium for improving text recall based on keyword extraction. Background Technology

[0002] The development of current industrial text search and question-answering systems typically employs a cascaded "recall-ranking" architecture. In the recall phase, the system needs to quickly filter candidate documents relevant to the user's query from a massive text database. As business scenarios become more complex, taking medical consultation scenarios as an example, user queries often contain multiple core entities simultaneously, such as multiple terms related to "drugs," "symptoms," or "diseases." However, existing text search systems, when handling such queries with multiple core entities, lack the ability to fine-grainedly analyze multiple entities and complex intents, as well as a dynamic matching mechanism. They typically adopt a coarse-grained retrieval strategy: the system often only extracts the word with the highest segmentation weight as the sole, mandatory matching condition, while downgrading the remaining core entities in the query to ordinary, weakly matched terms. This single-word-result strategy has serious flaws in practical applications. The system over-relies on a single local feature, fragmenting the user's overall semantic understanding, causing the underlying search engine to recall a large number of documents containing only the highest-weighted entity, while losing other important entities in the query. For example, when a user searches for "the effects of Job's tears and bamboo shavings," the system, limited by the aforementioned extensive strategy, only sets "bamboo shavings" as the only required keyword. This results in a large number of documents in the search results that only describe "bamboo shavings" and completely omit "Job's tears." This loss of core entities significantly reduces the efficiency of the subsequent fine-tuning stage and the accuracy of the final search results.

[0003] Therefore, this application proposes a method for improving text recall based on keyword extraction to solve the above problems. Summary of the Invention

[0004] The main objective of this invention is to provide a method for improving text retrieval performance based on keyword extraction. This method aims to solve the technical problems of core entity retrieval omissions and low retrieval accuracy caused by existing text retrieval systems when processing multi-entity queries, due to their reliance on a single forced matching condition and lack of a refined weight gain mechanism.

[0005] The present invention solves the above-mentioned technical problems by adopting the following technical solutions: A method for improving text recall based on keyword extraction includes: The query text is input into a pre-trained keyword extraction model, and the entity and its corresponding entity category are extracted from the initial query text using the ERNIE-based low-level embedding network and the fully connected layer classification prediction network in the keyword extraction model. Based on the entity category, using the preset list of required word categories and the scoring word mapping dictionary, the required word set and the scoring word set are obtained respectively; at the same time, global fine-grained word segmentation is performed on the initial query text, the fine-grained word segment with the highest segmentation weight is extracted, and the fine-grained word segment with the highest segmentation weight is added to the required word set; Fine-grained word segmentation is performed on the entities in the set of required words and the set of bonus words to obtain a set of sub-segments carrying word segmentation weights; and sub-segments are greedily selected in descending order of word segmentation weights until the sum of the characters of the selected sub-segments meets the preset entity length coverage condition as a proportion of the total number of characters of the corresponding entity. The sub-segments that meet the entity length coverage condition are combined to construct a text recall request body, wherein the sub-segments corresponding to the must-appear words are configured as forced matching conditions with a minimum hit ratio threshold, and the sub-segments corresponding to the bonus words are configured as gain matching conditions with corresponding retrieval gain weights. The text recall request body is executed, or the text recall request body is reconstructed after a fallback strategy is triggered when the preset number of entities is met. Candidate texts are recalled from the preset text library, and the recall results are output according to the text recall score.

[0006] Preferably, the list of required word categories and the scoring word mapping dictionary are allowed to contain overlapping items.

[0007] Preferably, the step of extracting entities and their corresponding entity categories from the initial query text specifically includes: using the classification prediction network combined with the sequence labeling mechanism, employing the BIO tagging system to accurately delineate the boundaries of each entity in the initial query text, and predicting the entity category to which each entity belongs.

[0008] Preferably, the preset entity length coverage condition is greater than or equal to one-third; the process of greedily selecting sub-segments in descending order of segmentation weight until the preset entity length coverage condition is met is as follows: calculate the ratio of the sum of the character counts of the currently selected sub-segments to the total number of characters of the corresponding original entity; determine whether the ratio is greater than or equal to one-third; if the ratio is less than one-third, continue to select the adjacent sub-segments with the second highest segmentation weight from the remaining sub-segments for accumulation until the ratio is greater than or equal to one-third.

[0009] Preferably, the step of constructing the text recall request body specifically includes: filling the sub-segments corresponding to the must-find words that meet the entity length coverage condition into the forced matching array in the distributed search engine request body, and configuring the logical operator inside the forced matching array as OR; configuring the minimum hit ratio threshold to 60%, that is, requiring that the search results must hit at least 60% of the sub-segments in the forced matching array.

[0010] Preferably, the step of constructing the text recall request body further includes: filling the sub-segments corresponding to the scoring words that meet the entity length coverage condition, as well as the must-find words themselves, into the gain matching array in the distributed search engine request body; and dynamically allocating the retrieval gain weight to different sub-segments in the gain matching array according to the scoring word mapping dictionary and the basic weight of the corresponding entity, so as to realize multi-level enhancement of retrieval weight when calculating the comprehensive recall score.

[0011] Preferably, the triggering fallback strategy is executed after reconstructing the text recall request body, specifically including: determining whether the number of entities in the set of guaranteed words is greater than eight; if it is greater than eight, then triggering the fallback strategy: canceling the configuration rules in the original text recall request body, using the fine-grained word segment with the highest word segmentation weight in the initial query text as a separate forced matching condition; and adding the remaining fine-grained words in the initial query text, except for the fine-grained word segment with the highest weight, equally as no-gain word segmentation to the gain matching condition, and executing the reconstructed text recall request.

[0012] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0013] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0014] As can be seen from the above technical solution, the present invention provides a method for improving text retrieval performance based on keyword extraction, comprising: inputting query text into a pre-trained keyword extraction model, performing word vector embedding using an ERNIE-based underlying model, and performing classification prediction through a fully connected layer to extract entities and entity categories from the text as a set of target keywords; performing additive word mapping and guaranteed word mapping on the extracted keywords respectively, and combining them into a text retrieval request for a search engine after fine-grained word segmentation and meeting a preset entity coverage ratio condition; executing the text retrieval request and outputting the retrieval results according to the text retrieval score. The present invention, by accurately extracting entities and entity categories from the query text and constructing retrieval conditions by combining a mapping mechanism that meets the coverage conditions, effectively avoids the loss of core entities during text retrieval and significantly improves the accuracy of retrieval results.

[0015] Compared with the prior art, the present invention has the following advantages: 1. This invention extracts multiple key entities from a query and transforms them into mandatory matching conditions with a minimum hit ratio threshold. This can cover core entities as much as possible while ensuring a basic recall rate, thereby effectively alleviating the problem of core information loss in multi-entity query scenarios and improving the accuracy of search results.

[0016] 2. This invention employs a selection strategy based on segmentation weights after fine-grained word segmentation and sets expected entity length coverage conditions. Therefore, it can reduce the loss of entity information caused by excessive segmentation by word segmentation tools, thereby better preserving the complete semantic expression of entities in the underlying search engine and ensuring the integrity of entity semantics.

[0017] 3. By combining forced matching with gain matching and dynamically allocating retrieval weights using a word mapping dictionary, this invention can accurately interpret users' complex search intentions and reduce interference from redundant words, thereby constructing a multi-level retrieval weight mechanism and improving the system's adaptability to complex queries.

[0018] 4. This invention, by supplementing the extraction loop with the highest global weight words and designing a corresponding strategy fallback mechanism for complex query situations where the number of entities exceeds a preset threshold, can smoothly transition to a mode that only retains the highest weight words as the mandatory matching condition. This ensures the stable operation and basic recall baseline of the system when facing high loads or extremely long text queries, effectively improving the retrieval stability and robustness of the system.

[0019] It should be understood that the descriptions in this section are not intended to identify key or essential features of embodiments of the invention, nor are they intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Of course, implementing any product of the invention does not necessarily require achieving all of the advantages described above simultaneously. Attached Figure Description

[0020] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the keyword-based text retrieval strategy process of the present invention; Figure 2 This is a schematic diagram of the keyword greedy coverage process based on sub-segmentation weights according to the present invention; Figure 3 This is a schematic diagram of the distributed search engine request body assembly process of the present invention; Figure 4 The document quality score retrieved using traditional methods; Figure 5 Document quality scoring for applying this invention. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] The method for improving text recall based on keyword extraction proposed in this invention first obtains the user's initial query text and extracts entities. Then, based on a preset dictionary, the entities are divided into guaranteed words and bonus words, while introducing the globally highest weighted word for mandatory protection. Next, fine-grained word segmentation is performed on each term, and greedily truncation is performed according to weight until the entity coverage condition is met. Finally, the number of entities is determined to decide whether to trigger a fallback strategy or construct a text recall request body with minimum hit ratio and bonus weight, and execute the recall operation. The overall process is as follows: Figure 1 As shown. Specifically includes: The query text is input into a pre-trained keyword extraction model, and the entity and its corresponding entity category are extracted from the query text using the ERNIE-based low-level embedding network and the fully connected layer classification prediction network in the keyword extraction model. Based on the entity category, using a preset list of required word categories and a bonus word mapping dictionary, the extracted entities are divided into a set of required words and a set of bonus words respectively. Fine-grained word segmentation is performed on the entities in the set of required words and the set of bonus words to obtain a set of sub-segments carrying word segmentation weights; and sub-segments are greedily selected in descending order of word segmentation weights until the sum of the characters of the selected sub-segments meets the preset entity length coverage condition as a proportion of the total number of characters of the corresponding entity. The sub-segments that meet the entity length coverage condition are combined to construct a text recall request body, wherein the sub-segments corresponding to the must-appear words are configured as forced matching conditions with a minimum hit ratio threshold, and the sub-segments corresponding to the bonus words are configured as gain matching conditions with corresponding retrieval gain weights. The text recall request body is executed, or the text recall request body is reconstructed after a fallback strategy is triggered when the preset number of entities is met. Candidate texts are recalled from the preset text library, and the recall results are output according to the text recall score.

[0023] Specifically, the list of required word categories and the scoring word mapping dictionary are allowed to contain overlapping items; Furthermore, when obtaining the set of guaranteed words and the set of bonus words respectively, the process also includes: Perform global fine-grained word segmentation on the initial query text, extract the fine-grained word segment with the highest weight, and add the fine-grained word segment with the highest weight to the set of words that must be output.

[0024] The extraction of entities and their corresponding entity categories from the query text specifically includes: using the classification prediction network combined with a sequence labeling mechanism, employing a BIO labeling system to accurately delineate the boundaries of each entity in the query text, and predicting the entity category to which each entity belongs.

[0025] The preset entity length coverage condition is greater than or equal to one-third; The process of greedily selecting sub-segments in descending order of segmentation weight until the preset entity length coverage condition is met is as follows: Calculate the ratio of the sum of the character counts of the currently selected sub-segments to the total character count of the corresponding original entity; Determine whether the ratio is greater than or equal to one-third; If the ratio is less than one-third, the next adjacent sub-segment with the second-highest weight is selected from the remaining sub-segments and accumulated until the ratio is greater than or equal to one-third.

[0026] The step of constructing the text recall request body specifically includes: The sub-segments corresponding to the mandatory words that meet the entity length coverage condition are filled into the forced matching array in the distributed search engine request body, and the logical operator inside the forced matching array is configured as OR. The minimum hit rate threshold is set to 60%, which means that the search results must hit at least 60% of the sub-words in the forced matching array.

[0027] The step of constructing the text recall request body also includes: The sub-segments corresponding to the bonus words that satisfy the entity length coverage condition are filled into the gain matching array in the distributed search engine request body; Based on the aforementioned word mapping dictionary and the basic weights of the corresponding entities, the retrieval gain weights are dynamically assigned to different sub-words in the gain matching array to achieve multi-level enhancement of retrieval weights when calculating the comprehensive recall score.

[0028] The trigger fallback strategy is executed after reconstructing the text recall request body, specifically including: Determine whether the number of entities in the set of guaranteed words is greater than eight; If there are more than eight, the fallback strategy is triggered: cancel the configuration rules in the original text recall request body, and use the fine-grained word segment with the highest word segmentation weight in the initial query text as a separate forced matching condition; Then, all the fine-grained words in the initial query text except for the fine-grained word with the highest weight are added equally as gain-free words to the gain matching condition, and the reconstructed text recall request is executed.

[0029] The following uses a medical consultation inquiry as an example. The overall process of this method is as follows: S1. Obtain user queries and perform keyword extraction: S101. Obtain the initial query text: The system receives the initial query request input by the user and extracts the initial query text as: "The role of coix seed and bamboo shavings in traditional Chinese medicine prescriptions;".

[0030] S102. Extract entities using a pre-trained model: The initial query text is input into a pre-trained keyword extraction model. The model's underlying layer extracts high-dimensional semantic features based on the ERNIE pre-trained language model, performs classification prediction through a fully connected layer, and employs the BIO sequence labeling mechanism to accurately delineate entity boundaries. The output named entities and their basic weights are as follows: "Coix Seed": Starting position 0, entity category is "medicine", basic weight 3; "Bamboo Shavings": start position 4, entity category is "drug", base weight 3; "Traditional Chinese Medicine Prescription": start position 7, entity category is "other", base weight 2; "effect": start position 12, entity category is "other", base weight 3.

[0031] S2. Perform strategy offloading and global high-weight word segmentation extraction: S201. Global high-weight word extraction: The system performs global fine-grained word segmentation on the initial query text, extracts one fine-grained segmented word with the highest global word segmentation weight (for example, the one with the highest weight in this embodiment is "Bamboo Shavings"), and directly adds it to the mandatory word set unconditionally, so as to prevent the loss of core intent caused by model extraction omission.

[0032] S202. Division of mandatory words and bonus words: The system is pre-configured with a mandatory word category list (for example, including the "drug" category) and a bonus word mapping dictionary, and the list and the dictionary are allowed to contain overlapping items. According to the extraction result, the system divides "Coix Seed" and "Bamboo Shavings" whose entity category is "drug" into mandatory words, and adds them to the mandatory word set; divides "Traditional Chinese Medicine Prescription" and "effect" whose entity category is "other" into bonus words, and adds them to the bonus word set.

[0033] S3. Perform fine-grained word segmentation and entity length coverage truncation, the execution process is as Figure 2 shown: S301. Fine-grained word segmentation processing: Perform fine-grained word segmentation processing on each term in the mandatory word set and the bonus word set respectively, to obtain a sub-segmentation sequence carrying part-of-speech (tag) and word segmentation weight (weight). For example: the fine segmentation weight of "Coix Seed" is 0.3729; the fine segmentation weight of "Bamboo Shavings" is 0.4021; the fine segmentation weight of "Traditional Chinese Medicine Prescription" is 0.1668; the fine segmentation weight of "middle" is 0.0167; the fine segmentation weight of "effect" is 0.0412.

[0034] S302. Coverage truncation calculation: Greedily select sub-segmentations sequentially in descending order of segmentation weight, until the sum of the character counts of the selected sub-segmentations meets the entity length coverage condition that the proportion of the total character count of the corresponding original term is greater than or equal to one third (1 / 3). Specifically: (1) For "Coix Seed": select "Coix Seed", the character length ratio is 3 / 3 = 100% > 1 / 3, which meets the condition, and stop the selection. (2) For "Bamboo Shavings": select "Bamboo Shavings", the character length ratio is 2 / 2 = 100% > 1 / 3, which meets the condition, and stop the selection. (3) For "Chinese Herbal Formula": the high-weight "Chinese Herbal Formula" is preferentially selected, the character length ratio is 3 / 3 = 100% > 1 / 3, which meets the condition, and stop the selection.

[0035] S4. Construct the search engine request body and perform fallback processing, the execution process is as Figure 3 shown as follows: S401. Extreme value fallback strategy judgment: Determine whether the number of entities in the mandatory term set is greater than eight. If it is greater than eight, the fallback strategy is triggered: the original assembly rule is canceled, only the fine-grained segmentation with the highest segmentation weight in the initial query text is used alone as a mandatory matching condition (must), and all other fine-grained segmentations are equally added as bonus terms with no gain (boost=1) to the bonus matching condition (should), so as to ensure the stability of the system under high load.

[0036] S402. Construction of mandatory matching and bonus matching under normal strategy: In this embodiment, the number of mandatory entities does not exceed eight, and the normal assembly process is executed: Fill the mandatory term sub-segmentations ("Coix Seed", "Bamboo Shavings") that meet the coverage condition into the mandatory matching array (must) in the request body of the distributed search engine (such as ES), and configure the internal operator as or. Since the total number of mandatory sub-segmentations is 2, the minimum hit proportion threshold is set to sixty percent (60%), that is, 2 × 60% = 1.2, and the minimum_should_match parameter is 1 after rounding down.

[0037] The bonus term sub-segmentations ("Chinese Herbal Formula", "Action") that meet the coverage condition, as well as the mandatory terms themselves, are jointly filled into the bonus matching array (should). According to the bonus term mapping dictionary, the search bonus weight is dynamically allocated: configure boost: 1.5 for "Coix Seed" and "Bamboo Shavings", configure boost: 1.2 for "Action", and configure boost: 1 for "Chinese Herbal Formula".

[0038] S5. Execute text recall operation: The underlying layer executes the text recall request body constructed above, recalls candidate texts from the preset text library, calculates the score after combining the forced matching and the weights of each level of gain, and outputs the sorted results according to the text recall score.

[0039] To verify the superiority of the method of the present invention, 200 user queries were randomly selected from the system log. The retrieval process remained unchanged, and only the text recall strategy was replaced to perform an end-to-end retrieval experiment.

[0040] Based on the scoring of the search results, the following evaluation metrics exist:

[0041] Representative: The proportion of results rated A in the final search results, i.e., the proportion of excellent search results.

[0042]

[0043] Representative: The proportion of results rated A or B in the final search results, i.e., the proportion of qualified search results.

[0044] The following table shows the experimental results:

[0045] As can be seen from the above experiments, the method of the present invention significantly improves the AB_grade rate and also improves the A_grade rate, thus alleviating the problem of search results not accurately matching user queries to a certain extent. Examples of the benefits brought by the method of the present invention include... Figure 4 , Figure 5 As shown.

[0046] It is understood that the system provided in the embodiments of the present invention corresponds to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

[0047] This application also provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.

[0048] Memory, used to store computer programs; When the processor executes the program stored in memory, it implements the above-mentioned method for improving text recall based on keyword extraction. Specifically, it includes the following steps: performing ERNIE-based keyword extraction and entity category prediction, dictionary mapping and diffing of sure-fire words and bonus words, greedy coverage truncation based on fine-grained word segmentation weights, constructing a text recall request body with a minimum hit ratio, and triggering an extreme value fallback strategy.

[0049] The communication bus mentioned in the above-mentioned electronic devices can be a standard bus for interconnecting peripheral components or an extended industrial standard structure bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc.

[0050] The communication interface is used for communication between the aforementioned electronic device and other external computing devices (such as the server that executes the distributed search engine).

[0051] The memory may include random access memory or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0052] The processors mentioned above can be general-purpose processors, including central processing units, network processors, etc.; they can also be digital signal processors, application-specific integrated circuits, field-programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0053] In conjunction with the methods provided in the embodiments of the present invention, the above embodiments can be implemented entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the text retrieval control flow or function based on keyword extraction as described in the embodiments of this application can be generated entirely or partially. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that integrates one or more available media.

[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention to the underlying network of the keyword extraction model, the truncation threshold for entity length coverage, the allocation logic of retrieval gain weights, and the extreme value of the number of entities triggering the fallback strategy, etc., should be included within the protection scope of the present invention.

[0055] Furthermore, the technical solutions of the various embodiments can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

Claims

1. A method for improving text recall based on keyword extraction, characterized in that, include: The query text is input into a pre-trained keyword extraction model, and the entity and its corresponding entity category are extracted from the query text using the ERNIE-based low-level embedding network and the fully connected layer classification prediction network in the keyword extraction model. Based on the entity category, using a preset list of required word categories and a bonus word mapping dictionary, the extracted entities are divided into a set of required words and a set of bonus words respectively. Fine-grained word segmentation is performed on the entities in the set of required words and the set of bonus words to obtain a set of sub-segments carrying word segmentation weights; and sub-segments are greedily selected in descending order of word segmentation weights until the sum of the characters of the selected sub-segments meets the preset entity length coverage condition as a proportion of the total number of characters of the corresponding entity. The sub-segments that meet the entity length coverage condition are combined to construct a text recall request body, wherein the sub-segments corresponding to the must-appear words are configured as forced matching conditions with a minimum hit ratio threshold, and the sub-segments corresponding to the bonus words are configured as gain matching conditions with corresponding retrieval gain weights. The text recall request body is executed, or the text recall request body is reconstructed after a fallback strategy is triggered when the preset number of entities is met. Candidate texts are recalled from the preset text library, and the recall results are output according to the text recall score.

2. The method for improving text recall based on keyword extraction as described in claim 1, characterized in that, The list of required word categories and the scoring word mapping dictionary are allowed to contain overlapping items; Furthermore, when obtaining the set of guaranteed words and the set of bonus words respectively, the process also includes: Perform global fine-grained word segmentation on the initial query text, extract the fine-grained word segment with the highest weight, and add the fine-grained word segment with the highest weight to the set of words that must be output.

3. The method for improving text recall based on keyword extraction as described in claim 1, characterized in that, The extraction of entities and their corresponding entity categories from the query text specifically includes: using the classification prediction network combined with a sequence labeling mechanism, employing a BIO labeling system to accurately delineate the boundaries of each entity in the query text, and predicting the entity category to which each entity belongs.

4. The method for improving text recall based on keyword extraction as described in claim 1, characterized in that, The preset entity length coverage condition is greater than or equal to one-third; The process of greedily selecting sub-segments in descending order of segmentation weight until the preset entity length coverage condition is met is as follows: Calculate the ratio of the sum of the character counts of the currently selected sub-segments to the total character count of the corresponding original entity; Determine whether the ratio is greater than or equal to one-third; If the ratio is less than one-third, the next adjacent sub-segment with the second-highest weight is selected from the remaining sub-segments and accumulated until the ratio is greater than or equal to one-third.

5. The method for improving text recall based on keyword extraction as described in claim 1, characterized in that, The step of constructing the text recall request body specifically includes: The sub-segments corresponding to the mandatory words that meet the entity length coverage condition are filled into the forced matching array in the distributed search engine request body, and the logical operator inside the forced matching array is configured as OR. The minimum hit rate threshold is set to 60%, which means that the search results must hit at least 60% of the sub-words in the forced matching array.

6. The method for improving text recall based on keyword extraction as described in claim 5, characterized in that, The step of constructing the text recall request body also includes: The sub-segments corresponding to the bonus words that satisfy the entity length coverage condition are filled into the gain matching array in the distributed search engine request body; Based on the word mapping dictionary and the basic weights of the corresponding entities, the retrieval gain weights are dynamically assigned to different sub-words in the gain matching array to achieve multi-level enhancement of retrieval weights when calculating the comprehensive recall score.

7. The method for improving text recall based on keyword extraction as described in claim 1, characterized in that, The trigger fallback strategy is executed after reconstructing the text recall request body, specifically including: Determine whether the number of entities in the set of guaranteed words is greater than eight; If there are more than eight, the fallback strategy is triggered: cancel the configuration rules in the original text recall request body, and use the fine-grained word segment with the highest word segmentation weight in the initial query text as a separate forced matching condition; Then, all the fine-grained words in the initial query text except for the fine-grained word with the highest weight are added equally as gain-free words to the gain matching condition, and the reconstructed text recall request is executed.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the processor, it causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

Images