A Vectorized Retrieval Optimization Method and System for Email Scenarios

Abstract

This invention relates to the field of data retrieval technology and provides a vectorized retrieval optimization method and system for email scenarios. The method includes: determining an appropriate range for topic tag discrimination based on the correspondence between the topic tag discrimination of each sample and the retrieval load value; determining the corresponding topic tag compression setting from samples whose topic tag discrimination is within the appropriate range, and using it as the standard topic tag compression setting. This avoids the loss of distinguishing clues between target emails and distracting emails due to excessive topic tag compression, and also avoids interference from duplicate tags and low-value tags in the email topic vector expression due to insufficient compression. This improves the distinguishing ability between emails with similar core topics, reduces the retrieval burden on the searcher in obtaining the target email from the retrieval result set, and improves the accuracy and usability of the vectorized email retrieval results.

Description

Technical Field

[0001] This invention belongs to the field of data retrieval technology, and in particular relates to a vectorized retrieval optimization method and system for email scenarios. Background Technology

[0002] With the continuous increase in the number of corporate emails, government emails, and project collaboration emails, users are no longer satisfied with simple keyword matching when searching for historical emails in email systems. Current technologies typically convert email subjects, bodies, or other email content into vector representations and use vector similarity to recall emails semantically similar to the search query. This approach can overcome problems such as inconsistent keywords and different expression methods to a certain extent, improving the recall capability of email retrieval.

[0003] In email scenarios, email subjects often accumulate multiple subject tags through replies, forwarding, security gateway processing, manual annotation, or system processing, such as "Re," "Fwd," "External," "Urgent," "Final Version," and "System Reminder." For emails circulating through multiple rounds, the same core subject may correspond to a large number of emails with similar themes, and their subject tags may be duplicated, similar, or multi-layered. To avoid excessive subject tags leading to verbose subject text and to reduce the interference of duplicate tags on vector representation, existing technologies typically compress subject tags before vectorization retrieval, for example, merging multiple subject tags with similarity thresholds into a single compressed tag.

[0004] However, existing subject tag compression methods often focus on the compression rules themselves, typically aiming to reduce duplicate tags, lower text noise, or improve processing efficiency. They rarely consider whether the compressed subject tags can still effectively distinguish the target email from distracting emails with similar core topics. In actual search results, the target email often appears alongside several emails with similar core topics. Over-compression of subject tags may lose crucial clues for distinguishing the target email; under-compression may retain too many low-value tags, interfering with the expression of the email's subject vector. Therefore, determining a more suitable subject tag compression method for the current core topic by combining historical search results is a problem that needs to be addressed in optimizing email vectorized retrieval. Summary of the Invention

[0005] The purpose of this invention is to provide a vectorized retrieval optimization method and system for email scenarios, aiming to solve the problems mentioned in the background art.

[0006] This invention is implemented as follows: a vectorized retrieval optimization method for email scenarios, the method comprising:

[0007] S1. Determine the current core topic corresponding to the current email retrieval request, and obtain historical email retrieval samples with the same core topic as the current core topic. The samples include historical target emails and historical interference emails with similar core topics to the historical target emails.

[0008] S2. Based on the distinction between historical target emails and historical interference emails after subject tag compression, determine the subject tag distinguishability of each sample and calculate the retrieval load value of each sample.

[0009] S3. Based on the correspondence between the topic tag discrimination of each sample and the retrieval load value, determine the appropriate range of topic tag discrimination;

[0010] S4. From the samples whose topic label discrimination is within a suitable range, determine the corresponding topic label compression setting and use it as the standard topic label compression setting;

[0011] S5. Based on the standard subject tag compression settings, perform subject tag compression on the subject of the email to be retrieved corresponding to the current email retrieval request, and perform vectorized retrieval on the compressed email subject.

[0012] As a further limitation of the technical solution of the present invention, the sample is a sample carrying topic clues; the historical target email is a target email confirmed by the searcher in the corresponding historical email retrieval process; and the historical interference email is an email that appears in the same search result set as the historical target email in the corresponding historical email retrieval process but has not been confirmed as a target email by the searcher.

[0013] As a further limitation of the technical solution of this embodiment of the invention, the core theme is the semantic information of the subject used to characterize the main content of the searched item, extracted from the email retrieval request; the same core theme means that the similarity between the semantic information of the sample and the semantic information of the current email retrieval request meets a first similarity threshold; the similar core theme means that the similarity between the semantic information of the subject of historical interference emails and the semantic information of the subject of historical target emails meets a second similarity threshold.

[0014] As a further limitation of the technical solution of this embodiment of the invention, the calculation process of the topic tag distinguishability includes:

[0015] Obtain the subject of the target email and each historical interference email from the same historical email retrieval sample after compression by subject tagging, and extract the retained subject tags of the target email and each historical interference email respectively.

[0016] The retained subject tags of the historical target emails are determined as a candidate tag set;

[0017] For any historical interference email, determine whether the candidate tag in the candidate tag set appears in the retained subject tag of the historical interference email, and calculate the appearance ratio of the candidate tag in the retained subject tag of the historical interference email;

[0018] When the candidate tag does not appear in the retained subject tags of any historical interference email, or when the appearance ratio of the candidate tag in the retained subject tags of any historical interference email is lower than a preset ratio threshold, the retained subject tag of any historical interference email is determined as a relative difference subject tag.

[0019] The topic tag discrimination of the historical email retrieval sample is calculated based on the number of relatively different topic tags, the tag type weight corresponding to the relatively different topic tags, and the degree of overlap between the relatively different topic tags and the candidate tags.

[0020] As a further limitation of the technical solution of the present invention, the retrieval load value is used to characterize the degree of retrieval burden that a retrieval user needs to bear in obtaining historical target emails from the retrieval result set corresponding to the historical email retrieval sample;

[0021] The retrieval load value is determined based on at least one of the following retrieval behavior parameters: the sorting position of the historical target email in the retrieval result set, the number of historical interference emails opened by the retrieval user before confirming the historical target email, the browsing time of the retrieval user in the retrieval result set, the number of times the retrieval user modifies the retrieval request, and the number of times the retrieval user returns to the retrieval result set.

[0022] As a further limitation of the technical solution of this embodiment of the invention, step S3 specifically includes:

[0023] S31. Sort several historical email retrieval samples in ascending order of subject tag distinguishability to obtain a sample sequence;

[0024] S32. Sequentially obtain the retrieval load value corresponding to each historical email retrieval sample in the sample sequence, and determine the changing trend between topic tag discrimination and retrieval load value;

[0025] S33. When the trend shows that the retrieval load value first decreases and then increases as the topic tag distinguishability increases, the topic tag distinguishability interval corresponding to the low load state of the retrieval load value is determined as the appropriate interval of the topic tag distinguishability.

[0026] As a further limitation of the technical solution of the present invention, when the trend of change does not show the trend of the retrieval load value first decreasing and then increasing as the subject tag distinguishability increases, the process of determining the appropriate range of the subject tag distinguishability is terminated, and the subject tag compression processing of the email subject corresponding to the current email retrieval request is performed on the current subject tag compression setting, and the compressed email subject is vectorized for retrieval.

[0027] As a further limitation of the technical solution of this embodiment of the invention, step S4 includes:

[0028] S41. Determine the historical email retrieval sample corresponding to the appropriate interval and use it as a reference sample;

[0029] S42. Obtain the topic tag compression settings used by the reference sample in the historical retrieval process, and use the topic tag compression settings as the standard topic tag compression settings;

[0030] S43. When there are multiple standard topic tag compression settings, calculate the similarity of topic semantic information between the core topic corresponding to each reference sample and the current core topic, and select the topic tag compression setting corresponding to the reference sample with the highest similarity of topic semantic information as the standard topic tag compression setting.

[0031] As a further limitation of the technical solution of this embodiment of the invention, step S5 specifically includes:

[0032] S51. According to the standard subject tag compression settings, merge, retain or delete the subject tags in the subject of the email to be retrieved corresponding to the current email retrieval request to obtain the compressed email subject.

[0033] S52. Input the compressed email subject into the vectorization model to generate an email subject vector, and perform similarity retrieval based on the email subject vector.

[0034] A vectorized retrieval optimization system for email scenarios, the system comprising:

[0035] The core topic determination module is used to determine the current core topic corresponding to the current email retrieval request and obtain historical email retrieval samples with the same core topic as the current core topic. The samples include historical target emails and historical interference emails with similar core topics to the historical target emails.

[0036] The sample parameter calculation module is used to determine the subject tag discrimination degree of each sample based on the distinction status of historical target emails and historical interference emails after subject tag compression, and to calculate the retrieval load value of each sample.

[0037] The appropriate interval determination module is used to determine the appropriate interval of topic tag discrimination based on the correspondence between the topic tag discrimination of each sample and the retrieval load value.

[0038] The standard setting determination module is used to determine the corresponding topic tag compression setting from samples whose topic tag discrimination is within a suitable range, and use it as the standard topic tag compression setting;

[0039] The vectorized retrieval module is used to compress the subject of the email to be retrieved corresponding to the current email retrieval request based on the standard subject tag compression settings, and then perform vectorized retrieval on the compressed email subject.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] To address the issue of duplicate subject tags in email vectorized retrieval caused by multiple rounds of replies, forwarding, or system processing, this invention moves beyond simply compressing subject tags using fixed rules. Instead, it combines historical target emails, historical interference emails, and retrieval load values ​​to determine an appropriate range for subject tag distinguishability, and accordingly obtains standard subject tag compression settings. This avoids over-compression of subject tags, which can lead to the loss of distinguishing clues between target and interference emails, and also prevents under-compression, which can result in duplicate or low-value tags interfering with the subject vector expression. This improves the ability to distinguish emails with similar core topics, reduces the retrieval burden on users obtaining target emails from the search results set, and enhances the accuracy and usability of email vectorized retrieval results. Attached Figure Description

[0042] Figure 1 A flowchart of the method provided in the embodiments of the present invention;

[0043] Figure 2 This is a flowchart illustrating the method for obtaining a suitable range of topic tag distinguishability in the embodiments of the present invention;

[0044] Figure 3 This is a flowchart illustrating the process of determining standard topic tag compression settings in the method provided in this embodiment of the invention;

[0045] Figure 4 This is a flowchart illustrating the application of standard topic tag compression settings in the method provided in this embodiment of the invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0047] Figure 1 A flowchart of the method provided by an embodiment of the present invention is shown.

[0048] Specifically, a vectorized retrieval optimization method for email scenarios includes the following steps:

[0049] S1. Determine the current core topic corresponding to the current email retrieval request, and obtain historical email retrieval samples with the same core topic as the current core topic. The samples include historical target emails and historical interference emails with similar core topics to the historical target emails.

[0050] The sample is a sample carrying topic clues; the historical target email is the target email confirmed by the searcher during the corresponding historical email retrieval process; the historical interference email is an email that appears in the same search result set as the historical target email during the corresponding historical email retrieval process, but has not been confirmed as the target email by the searcher.

[0051] The core topic is the semantic information of the subject extracted from the email retrieval request to represent the main content of the retrieval item; the same core topic means that the similarity between the semantic information of the subject corresponding to the sample and the semantic information of the subject corresponding to the current email retrieval request meets the first similarity threshold; the similar core topic means that the similarity between the semantic information of the subject of historical interference emails and the semantic information of the subject of historical target emails meets the second similarity threshold.

[0052] In this embodiment of the invention, vectorized email retrieval refers to converting the searchable content in an email into a vector representation and determining emails related to the email retrieval request based on the similarity between the vectors. The searchable content may include information such as email subject, email body, email attachment summary, email sender, email recipient, and email time. In existing technologies, email text is typically cleaned, segmented, and semantically extracted first, then corresponding email vectors are generated using a vectorization model. Subsequently, the email retrieval request is also converted into a query vector. By calculating the similarity between the query vector and the email vector, emails semantically close to the email retrieval request are recalled. For example, if a searcher inputs "customer urgently confirms final contract version," the system does not only search for emails containing that exact phrase, but also recalls emails semantically related to "customer confirmation," "final contract version," and "urgent processing."

[0053] This invention focuses on a vectorized email retrieval scenario that incorporates subject-specific clues. In this scenario, email retrieval requests typically do not solely target the email subject field. Instead, they utilize the semantic information of the subject within the retrieval request to recall emails related to the search query from the email database. In other words, the email subject is not the sole retrieval object, but the semantic information it carries participates in the retrieval process and plays a crucial role in filtering and differentiating similar emails.

[0054] Emails that have undergone multiple rounds of replies, forwarding, security gateway processing, manual annotation, or system processing typically have subject tags superimposed in their subject lines. For example, the subject line might contain tags such as "Re," "Fwd," "External," "Urgent," "Final Version," or "System Reminder." For the same email matter, as different people reply, forward, or the system gateway processes it automatically, the subject line might gradually change to a format like "Re: Fwd: [External] [Urgent] Contract Final Version Confirmation." These subject tags can reflect the email's reply status, forwarding status, external source, urgency, or version status, but they can also increase the length of the email subject line text due to repeated superposition and affect the semantic expression of the subject line during vectorization.

[0055] In existing email vectorization retrieval processes, to avoid multi-level, repetitive subject tags increasing text length and interfering with vector representation, subject tags in email subjects are typically compressed before retrieval. For example, when the similarity between "Re," "reply," and "answer" meets a preset similarity threshold, they can be merged into the same "reply" compressed tag; similarly, when the similarity between "Fwd," "forward," and "FW" meets a preset similarity threshold, they can be merged into the same "forward" compressed tag. This method reduces the impact of duplicate tags on email subject vectorization retrieval.

[0056] The topic tag compression settings may not be exactly the same in different historical retrieval processes. This difference can arise naturally from objective factors and is not necessarily artificially set for the purposes of this invention. For example, different email system versions at different times lead to different topic tag recognition rules; different email gateways or security policies add tags such as "external" and "system reminder" in different ways, resulting in different compression targets; different email clients or language environments express reply and forward tags differently, resulting in different tag similarity judgment results; different users or departments may enable different email archiving rules, retrieval configurations, or topic cleaning rules, resulting in different topic tag compression settings actually used in historical retrieval processes. Therefore, multiple historical email retrieval samples with different topic tag compression settings can naturally be formed in historical retrieval data.

[0057] However, in actual email search results, the historical target emails that the searcher ultimately confirms often appear alongside several historical distracting emails with similar core themes. Since these emails have already undergone initial screening by the search system, their core themes are often quite similar, making further differentiation difficult based solely on the core themes. For example, if a searcher wants to find emails titled "Client urgently confirms final version of contract," the search results may simultaneously include emails such as "Contract final version confirmed," "Re: Contract final version confirmed," "Fwd: Contract final version confirmed," and "[External] [Urgent] Contract final version confirmed." In this case, the subject tags that are retained after compression may become an important clue to distinguish historical target emails from historical distracting emails.

[0058] Those skilled in the art, through analysis of numerous completed historical email retrieval samples, have discovered that under the same core topic, different topic tag compression settings lead to varying degrees of difference in the compressed topic tags between emails. The relationship between this difference and the retrieval burden on the searcher to obtain the target historical email is not a simple monotonic one. When compression is excessive, too many topic tags that distinguish the target historical email from historical distracting emails are merged or deleted, requiring the searcher to manually identify them among multiple emails with similar core topics, resulting in a high retrieval burden. When compression is appropriate, repetitive low-value tags are compressed, while topic tags reflecting differences in external source, urgency, version status, etc., are retained, making it easier for the searcher to identify the target historical email from historical distracting emails, resulting in a lower retrieval burden. When compression is insufficient, too many repetitive or low-value tags still participate in vectorized retrieval, potentially weakening the effective semantic expression of the email topic, and requiring the searcher to still bear a high filtering burden. Therefore, based on historical email retrieval samples, suitable topic tag compression settings for retrieving similar core topics can be analyzed and applied to current email retrieval requests.

[0059] In step S1, the current core topic corresponding to the current email retrieval request is first determined. The current core topic is the semantic information extracted from the current email retrieval request that characterizes the main content of the retrieval item. For example, for "finding emails requesting urgent confirmation of the final version of a contract from a customer," the current core topic could include "customer confirmation" and "final version of the contract"; for "finding emails forwarded by a supplier confirming a quote," the current core topic could include "supplier quote confirmation." The current core topic can be obtained from the current email retrieval request through keyword extraction, entity recognition, semantic summarization, vectorization models, or a combination thereof.

[0060] After determining the current core topic, historical email retrieval samples with the same core topic as the current core topic are retrieved from the historical retrieval database. These historical email retrieval samples are samples carrying topic clues, including historical target emails and historical interference emails with similar core topics. Historical target emails refer to emails confirmed as target emails by the searcher during the corresponding historical email retrieval process. Examples include emails that were clicked and remained on for more than a preset time, opened and replied to or forwarded, added to favorites, had attachments downloaded, or were marked as successful retrieval results by the system. Historical interference emails refer to emails that appear in the same search result set as historical target emails during the corresponding historical email retrieval process but were not confirmed as target emails by the searcher.

[0061] By selecting historical email retrieval samples that share the same core theme as the current core topic, subsequent analysis can be based on historical data with similar search intent and email distribution to the current email retrieval request, avoiding the introduction of historical search behaviors with irrelevant topics into the current processing. For example, if the current email retrieval request involves "confirmation of the final version of the contract," then historical email retrieval samples that also involve "confirmation of the final version of the contract" or have the same semantics should be prioritized, rather than historical email retrieval samples with significantly different themes such as "confirmation of meeting minutes" or "sample delivery arrangements."

[0062] In this embodiment of the invention, the shared core topic can be determined by the similarity between the topic semantic information corresponding to the sample and the topic semantic information corresponding to the current email retrieval request. Specifically, the topic semantic information corresponding to the sample and the topic semantic information corresponding to the current email retrieval request can be input into a semantic vector model to obtain corresponding topic semantic vectors, and the cosine similarity between the two topic semantic vectors can be calculated. When the similarity meets a first similarity threshold, it is determined that the sample and the current email retrieval request have the same core topic. In addition to cosine similarity, Euclidean distance, dot product similarity, edit distance, and semantic similarity can also be used for determination.

[0063] The first similarity threshold is used to control the degree of topic consistency between historical email retrieval samples and the current email retrieval request. The first similarity threshold can be set based on topic clustering results in the historical retrieval database, manually labeled samples, system default configurations, or business scenario requirements. For example, in scenarios with strict core topic requirements, a higher first similarity threshold can be set to ensure a high degree of consistency between the obtained historical email retrieval samples and the current email retrieval request; in scenarios where email topics are more dispersed, the first similarity threshold can be appropriately lowered to improve the recall of historical email retrieval samples.

[0064] The similarity of core themes can be determined by the similarity between the semantic information of the themes of historical interference emails and the semantic information of the themes of historical target emails. Specifically, the semantic information of the themes of historical target emails and historical interference emails can be extracted, and corresponding semantic vectors can be generated for each. When the similarity between the semantic vector of the themes of historical interference emails and the semantic vector of the themes of historical target emails meets a second similarity threshold, it is determined that the historical interference emails and the historical target emails have similar core themes. The second similarity threshold is used to limit the degree of proximity between historical interference emails and historical target emails, so that the historical interference emails selected in the sample are indeed emails that are easily confused with historical target emails, rather than ordinary search results that are obviously unrelated to the search query.

[0065] The first and second similarity thresholds can be the same or different. Typically, the first similarity threshold is used to determine whether a historical email retrieval sample is suitable for migration to the current email retrieval request, while the second similarity threshold is used to determine whether historical interference emails within the same historical email retrieval sample are sufficiently close to the historical target email. Both can be determined based on historical retrieval accuracy, the size of the historical retrieval result set, topic clustering effects, or manual verification results. By combining the first and second similarity thresholds, it can be ensured that the historical email retrieval samples obtained in step S1 are both comparable to the current email retrieval request and contain sufficient historical data to reflect the distinction between the target email and interference emails.

[0066] Furthermore, the vectorized retrieval optimization method for email scenarios also includes the following steps:

[0067] S2. Based on the distinction between historical target emails and historical interference emails after subject tag compression, determine the subject tag discrimination degree for each sample and calculate the retrieval load value for each sample.

[0068] The calculation process for the topic tag distinguishability includes:

[0069] Obtain the subject of the target email and each historical interference email from the same historical email retrieval sample after compression by subject tagging, and extract the retained subject tags of the target email and each historical interference email respectively; determine the retained subject tags of the target email as a candidate tag set; for any historical interference email, determine whether the candidate tag in the candidate tag set appears in the retained subject tags of any historical interference email, and calculate the appearance ratio of the candidate tag in the retained subject tags of any historical interference email;

[0070] When the candidate tag does not appear in the retained subject tags of any historical interference email, or when the appearance ratio of the candidate tag in the retained subject tags of any historical interference email is lower than a preset ratio threshold, the retained subject tag of any historical interference email is determined as a relative difference subject tag.

[0071] The topic tag discrimination of the historical email retrieval sample is calculated based on the number of relatively different topic tags, the tag type weight corresponding to the relatively different topic tags, and the degree of overlap between the relatively different topic tags and the candidate tags.

[0072] The retrieval load value is used to characterize the retrieval burden that a retriever needs to bear to obtain historical target emails from the retrieval result set corresponding to the historical email retrieval sample; the retrieval load value is determined based on at least one of the following retrieval behavior parameters: the sorting position of the historical target email in the retrieval result set, the number of historical interference emails opened by the retriever before confirming the historical target email, the browsing time of the retriever in the retrieval result set, the number of times the retriever modifies the retrieval request, and the number of times the retriever returns to the retrieval result set.

[0073] In this embodiment of the invention, step S2 is used to further quantify the distinction between historical target emails and historical interference emails after subject tag compression, based on the already determined historical email retrieval samples. It should be noted that this invention does not only focus on the subject tag compression settings themselves, such as compression threshold levels, tag merging rules, or tag retention rules, but rather on the actual distinction results formed between historical target emails and historical interference emails after subject tag compression. In other words, subject tag distinguishability is used to describe the differences at the compression result level, rather than simply describing the compression rules themselves.

[0074] Specifically, within the same historical email retrieval sample, both historical target emails and historical interference emails have undergone corresponding subject tag compression settings. Therefore, retained subject tags can be extracted from their compressed email subjects. Retained subject tags can be obtained through a preset tag dictionary, regular expressions, subject prefix recognition rules, or tag classification models. For example, tags such as "Re," "Fwd," "External," "Urgent," "Final Version," and "System Reminder" can be identified and retained after compression.

[0075] When calculating the subject tag discriminability, the retained subject tags of historical target emails can be first determined as a candidate tag set. Then, the relationship between the retained subject tags of each historical interference email and the candidate tag set is compared. If no candidate tags appear in the retained subject tags of a historical interference email, or the appearance ratio of candidate tags is lower than a preset ratio threshold, it indicates that there is a discriminatory difference between the historical interference email and the historical target email at the compressed subject tag level. In this case, the retained subject tags of the historical interference email are determined as relative difference subject tags. The subject tag discriminability of the historical email retrieval sample is calculated by combining the number of relative difference subject tags, the tag type weight corresponding to the relative difference subject tags, and the degree of overlap between the relative difference subject tags and the candidate tags.

[0076] The above calculation process can be implemented in practice. For example, the system can represent the retained subject tags of each email as a tag set; record the tag set of historical target emails as a candidate tag set; perform set matching between the tag set of each historical interference email and the candidate tag set, count the proportion of candidate tags appearing in historical interference emails, and further count the number and type of relatively different subject tags. The tag type weight can be pre-configured according to the distinguishing value of the tag; for example, the distinguishing value of tags such as "urgent," "final version," and "external" can be higher than that of the ordinary "Re" tag; the degree of overlap can be determined by the ratio of the number of intersections to the number of unions between two tag sets. Thus, the subject tag distinguishability can be obtained through rule-based calculation, weighted calculation, or normalized calculation.

[0077] The significance of setting subject tag discriminability lies in transforming the question of "whether compressed subject tags still help distinguish between historical target emails and historical interference emails" into a calculable quantity. Existing technologies typically focus on how to compress subject tags to reduce repetition or text noise, while this invention further focuses on whether the subject tags retained after compression still retain the ability to distinguish target emails. This value can provide a basis for subsequently determining the appropriate range of subject tag discriminability.

[0078] Simultaneously, step S2 also calculates the retrieval load value corresponding to each historical email retrieval sample. The retrieval load value characterizes the retrieval burden a searcher bears when retrieving the historical target email from the retrieval result set corresponding to the historical email retrieval sample. This value can be obtained based on common user behavior logs in existing retrieval systems, such as the ranking position of the historical target email in the retrieval result set, the number of historical distracting emails opened by the searcher before confirming the historical target email, the duration of browsing the retrieval result set, the number of times the search request is modified, and the number of times the retrieval result set is returned. For example, if the historical target email is ranked high and the searcher confirms the historical target email without opening too many historical distracting emails, the corresponding retrieval load value is low; if the searcher repeatedly modifies the search request and opens multiple historical distracting emails before confirming the historical target email, the corresponding retrieval load value is high.

[0079] Furthermore, the vectorized retrieval optimization method for email scenarios also includes the following steps:

[0080] S3. Based on the correspondence between the topic tag discrimination of each sample and the retrieval load value, determine the appropriate range of topic tag discrimination.

[0081] Specifically, Figure 2 A flowchart is shown to obtain the appropriate range for topic tag discrimination.

[0082] The determination of the appropriate range for topic tag discrimination based on the correspondence between the topic tag discrimination of each sample and the retrieval load value specifically includes the following steps:

[0083] Step S31: Sort several historical email retrieval samples in ascending order of subject tag distinguishability to obtain a sample sequence;

[0084] Step S32: Sequentially obtain the retrieval load value corresponding to each historical email retrieval sample in the sample sequence, and determine the changing trend between topic tag discrimination and retrieval load value;

[0085] Step S33: When the trend shows that the retrieval load value first decreases and then increases as the topic tag distinguishability increases, the topic tag distinguishability interval corresponding to the low load state of the retrieval load value is determined as the appropriate interval of the topic tag distinguishability.

[0086] When the trend does not show that the retrieval load value first decreases and then increases with the increase of topic tag discrimination, the process of determining the appropriate range of topic tag discrimination is terminated, and the topic tag compression processing of the email subject corresponding to the current email retrieval request is performed based on the current topic tag compression setting, and the compressed email subject is vectorized for retrieval.

[0087] In this embodiment of the invention, step S3 is used to determine, based on multiple historical email retrieval samples, whether there is a pattern of change between the topic tag distinguishability and the retrieval load value that can be used for the current retrieval in the historical retrieval scenario corresponding to the current core topic. In other words, step S3 does not directly participate in email recall, but is used to identify a suitable range of topic tag distinguishability from historical email retrieval samples for the current email retrieval request.

[0088] Specifically, in step S31, several historical email retrieval samples are sorted in ascending order of topic tag discriminability to obtain a sample sequence. This sorting method transforms the compressed discriminability of topic tags in different historical email retrieval samples into a continuous and observable sample arrangement, facilitating subsequent analysis of the trend of retrieval load value changing with topic tag discriminability.

[0089] In step S32, the retrieval load value corresponding to each historical email retrieval sample in the sample sequence is obtained sequentially, and the changing trend between the subject tag discrimination and the retrieval load value is determined. This process can be achieved through trend analysis, interval statistics, sliding window averaging, curve fitting, or piecewise fitting. For example, the sample sequence can be divided into multiple intervals according to the subject tag discrimination, the average retrieval load value in each interval can be calculated, and then the overall trend can be determined based on the change in the average value of each interval.

[0090] In step S33, when the trend shows that the retrieval load value first decreases and then increases as the subject tag discrimination increases, it indicates that when the subject tag discrimination is too low, there are insufficient subject tag distinguishing clues between the target email and historical interference emails; when the subject tag discrimination is moderate, the searcher can more easily obtain the historical target email from the historical interference emails; when the subject tag discrimination is too high, too many or too complicated differences in subject tags are retained, which increases the burden on the searcher to identify the historical target email. Therefore, the subject tag discrimination range corresponding to the low retrieval load value is determined as the appropriate range for subject tag discrimination.

[0091] The significance of an appropriate range for subject tag discrimination lies in providing a basis for subsequently determining standard subject tag compression settings. In other words, this appropriate range represents the subject tag discrimination state corresponding to a low retrieval burden for the searcher when obtaining historical target emails from historical email retrieval samples with the same or similar core topics. Subject tag compression settings can then be selected from the historical email retrieval samples corresponding to this appropriate range for use in the current email retrieval request.

[0092] When the trend does not show a pattern where the retrieval load value first decreases and then increases with the increase of topic tag discrimination, it indicates that the historical email retrieval samples corresponding to the current core topic do not exhibit a stable and usable interval pattern. In this case, to avoid affecting the retrieval results by adjusting the topic tag compression settings based on unstable patterns, this embodiment of the invention terminates the process of determining the appropriate interval for topic tag discrimination, and performs topic tag compression processing on the email topics of the emails to be retrieved corresponding to the current email retrieval request based on the current topic tag compression settings, and performs vectorized retrieval on the compressed email topics.

[0093] Furthermore, the vectorized retrieval optimization method for email scenarios also includes the following steps:

[0094] Step S4: Determine the corresponding topic tag compression setting from the samples whose topic tag discrimination is within the appropriate range, and use it as the standard topic tag compression setting.

[0095] Specifically, Figure 3 A flowchart for determining standard topic tag compression settings is shown.

[0096] The process of determining the corresponding topic tag compression setting from samples whose topic tag discrimination is within a suitable range and using it as the standard topic tag compression setting specifically includes the following steps:

[0097] Step S41: Determine the historical email retrieval sample corresponding to the appropriate interval and use it as a reference sample;

[0098] Step S42: Obtain the topic tag compression settings used by the reference sample in the historical retrieval process, and use the topic tag compression settings as the standard topic tag compression settings;

[0099] Step S43: When there are multiple standard topic tag compression settings, calculate the similarity of topic semantic information between the core topic corresponding to each reference sample and the current core topic, and select the topic tag compression setting corresponding to the reference sample with the highest similarity of topic semantic information as the standard topic tag compression setting.

[0100] Step S5: Based on the standard subject tag compression settings, perform subject tag compression on the subject of the email to be retrieved corresponding to the current email retrieval request, and perform vectorized retrieval on the compressed email subject.

[0101] Specifically, Figure 4 A flowchart illustrating the application of standard topic tag compression settings is shown.

[0102] The process of compressing the email subject based on standard subject tagging settings, and then performing vectorized retrieval on the compressed email subject, specifically includes the following steps:

[0103] Step S51: According to the standard subject tag compression settings, merge, retain or delete the subject tags in the subject of the email to be retrieved corresponding to the current email retrieval request to obtain the compressed email subject.

[0104] Step S52: Input the compressed email subject into the vectorization model to generate an email subject vector, and perform similarity retrieval based on the email subject vector.

[0105] In this embodiment of the invention, step S4 is used to convert the suitable range of subject tag discrimination obtained in step S3 into a standard subject tag compression setting that can be directly applied to the current email retrieval request. Since the suitable range itself represents a compressed discrimination state, while the current email retrieval process actually needs to execute the rules for merging, retaining, or deleting subject tags, it is necessary to further determine the subject tag compression setting used in the historical retrieval process from historical email retrieval samples whose subject tag discrimination is within the suitable range.

[0106] Specifically, in step S41, historical email retrieval samples with subject tag discrimination within a suitable range are determined as reference samples. The historical retrieval process corresponding to this reference sample has already shown a low retrieval load value, so the subject tag compression settings used therein can be used as a reference for the current email retrieval request.

[0107] In step S42, the topic tag compression settings used by the reference sample during the historical retrieval process are obtained and used as the standard topic tag compression settings. The topic tag compression settings may include an approximation threshold between topic tags, mergeable topic tag categories, topic tag categories to be retained, and low-value topic tag categories to be deleted. For example, for reply-type tags such as "Re," "Reply," and "Answer," a merging rule can be used; for tags with strong distinguishing power such as "External," "Urgent," and "Final Version," a retention rule can be used; and for recurring system reminder-type tags, deletion or weakening rules can be used.

[0108] In step S43, if there are multiple reference samples whose topic tag distinguishability is within a suitable range, the similarity of topic semantic information between the core topic corresponding to each reference sample and the current core topic is further calculated, and the topic tag compression setting corresponding to the reference sample with the highest similarity of topic semantic information is selected as the standard topic tag compression setting. This method avoids rule distortion caused by simply averaging multiple compression settings, making the standard topic tag compression setting closer to the specific core topic corresponding to the current email retrieval request.

[0109] In step S5, based on the standard subject tag compression settings, the subject tags of the email to be retrieved corresponding to the current email retrieval request are compressed. Specifically, in step S51, subject tags in the subject of the email to be retrieved can be merged, retained, or deleted according to the standard subject tag compression settings to obtain a compressed email subject. For example, multiple duplicate reply tags can be merged into one reply tag, retaining subject tags with distinguishing significance such as "external," "urgent," and "final version," while deleting system tags with high repetition or low distinguishing value.

[0110] In step S52, the compressed email subject is input into the vectorization model to generate an email subject vector, and a similarity search is performed based on the email subject vector. By first processing the email subject using standard subject tag compression settings and then performing vectorization search, the current email search request can use the subject tag compression method that performed better in historical samples when processing emails with similar core subjects, thereby reducing search interference caused by over-compression or under-compression.

[0111] The overall beneficial effect of this invention lies in the fact that it does not solely determine how subject tags should be processed based on the compression rules themselves. Instead, it combines historical target emails, historical interference emails, and searcher behavior from historical email retrieval samples to analyze the correspondence between the actual distinguishability of compressed subject tags and the retrieval burden, thereby determining a standard subject tag compression setting suitable for the current core topic. This avoids the problem in existing technologies where fixed compression settings cannot adapt to different core topic email retrieval scenarios.

[0112] This invention addresses the aforementioned core research points. For emails that have undergone multiple rounds of replies, forwarding, security gateway processing, manual annotation, or system processing, resulting in multi-layered topic tags, this invention neither simply deletes nor retains all topic tags. Instead, it determines a more suitable topic tag compression method under similar core topics based on historical samples. This reduces the interference of repetitive, low-value topic tags on vector representation while retaining topic tag clues that help distinguish historical target emails from historical interference emails, thereby reducing the retrieval burden for searchers seeking target emails.

[0113] In terms of application prospects, this invention is applicable to email vectorization retrieval scenarios in enterprise email, government email, customer service email systems, project collaboration email systems, contract approval email systems, and knowledge management platforms. Especially in scenarios with a large number of emails, numerous replies and forwards on the same issue, and multiple topic tags superimposed in email subjects, this invention can improve the ability to distinguish between similar emails, making the retrieval results more likely to present the emails that users truly need. This method can also be combined with existing keyword retrieval, full-text retrieval, semantic vector retrieval, and email metadata filtering mechanisms as an optimization step in topic tagging processing before email vectorization retrieval.

[0114] In another preferred embodiment of the present invention, a vectorized retrieval optimization system for email scenarios is also included, the system comprising:

[0115] The core topic determination module is used to determine the current core topic corresponding to the current email retrieval request, and to obtain historical email retrieval samples with the same core topic as the current core topic. The samples include historical target emails and historical interference emails with similar core topics to the historical target emails.

[0116] The sample parameter calculation module is used to determine the subject tag discrimination degree of each sample based on the distinction status of historical target emails and historical interference emails after subject tag compression, and to calculate the retrieval load value of each sample.

[0117] The appropriate interval determination module is used to determine the appropriate interval of topic label discrimination based on the correspondence between the topic label discrimination of each sample and the retrieval load value.

[0118] The standard setting determination module is used to determine the corresponding topic tag compression setting from samples whose topic tag discrimination is within a suitable range, and use it as the standard topic tag compression setting.

[0119] The vectorized retrieval module is used to compress the subject of the email to be retrieved corresponding to the current email retrieval request based on the standard subject tag compression settings, and then perform vectorized retrieval on the compressed email subject.

[0120] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0121] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0122] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0123] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

[0124] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A vectorized retrieval optimization method for email scenarios, characterized in that, The method includes: S1. Determine the current core topic corresponding to the current email retrieval request, and obtain historical email retrieval samples with the same core topic as the current core topic; the samples include historical target emails and historical interference emails with similar core topics to the historical target emails; S2. Based on the distinction between historical target emails and historical interference emails after subject tag compression, determine the subject tag distinguishability of each sample and calculate the retrieval load value of each sample. S3. Based on the correspondence between the topic tag discrimination of each sample and the retrieval load value, determine the appropriate range of topic tag discrimination; S4. From the samples whose topic label discrimination is within a suitable range, determine the corresponding topic label compression setting and use it as the standard topic label compression setting; S5. Based on the standard subject tag compression settings, perform subject tag compression on the subject of the email to be retrieved corresponding to the current email retrieval request, and perform vectorized retrieval on the compressed email subject.

2. The vectorized retrieval optimization method for email scenarios according to claim 1, characterized in that, The sample is a sample carrying topic clues; the historical target email is the target email confirmed by the searcher during the corresponding historical email retrieval process; the historical interference email is an email that appears in the same search result set as the historical target email during the corresponding historical email retrieval process, but has not been confirmed by the searcher as the target email.

3. The vectorized retrieval optimization method for email scenarios according to claim 1, characterized in that, The core topic is the semantic information of the subject extracted from the email retrieval request to represent the main content of the retrieval item; the same core topic means that the similarity between the semantic information of the subject corresponding to the sample and the semantic information of the subject corresponding to the current email retrieval request meets the first similarity threshold; the similar core topic means that the similarity between the semantic information of the subject of historical interference emails and the semantic information of the subject of historical target emails meets the second similarity threshold.

4. The vectorized retrieval optimization method for email scenarios according to claim 1, characterized in that, The calculation process for the topic tag distinguishability includes: Obtain the subject of the target email and each historical interference email from the same historical email retrieval sample after compression by subject tagging, and extract the retained subject tags of the target email and each historical interference email respectively. The retained subject tags of the historical target emails are determined as a candidate tag set; For any historical interference email, determine whether the candidate tag in the candidate tag set appears in the retained subject tag of the historical interference email, and calculate the appearance ratio of the candidate tag in the retained subject tag of the historical interference email; When the candidate tag does not appear in the retained subject tags of any historical interference email, or when the appearance ratio of the candidate tag in the retained subject tags of any historical interference email is lower than a preset ratio threshold, the retained subject tag of any historical interference email is determined as a relative difference subject tag. The topic tag discrimination of the historical email retrieval sample is calculated based on the number of relatively different topic tags, the tag type weight corresponding to the relatively different topic tags, and the degree of overlap between the relatively different topic tags and the candidate tags.

5. The vectorized retrieval optimization method for email scenarios according to claim 1, characterized in that, The retrieval load value is used to characterize the retrieval burden that a retriever needs to bear in order to obtain historical target emails from the retrieval result set corresponding to the historical email retrieval sample; The retrieval load value is determined based on at least one of the following retrieval behavior parameters: the sorting position of the historical target email in the retrieval result set, the number of historical interference emails opened by the retrieval user before confirming the historical target email, the browsing time of the retrieval user in the retrieval result set, the number of times the retrieval user modifies the retrieval request, and the number of times the retrieval user returns to the retrieval result set.

6. The vectorized retrieval optimization method for email scenarios according to claim 1, characterized in that, Step S3 specifically includes: S31. Sort several historical email retrieval samples in ascending order of subject tag distinguishability to obtain a sample sequence; S32. Sequentially obtain the retrieval load value corresponding to each historical email retrieval sample in the sample sequence, and determine the changing trend between topic tag discrimination and retrieval load value; S33. When the trend shows that the retrieval load value first decreases and then increases as the topic tag distinguishability increases, the topic tag distinguishability interval corresponding to the low load state of the retrieval load value is determined as the appropriate interval of the topic tag distinguishability.

7. The vectorized retrieval optimization method for email scenarios according to claim 6, characterized in that, When the trend does not show that the retrieval load value first decreases and then increases with the increase of topic tag discrimination, the process of determining the appropriate range of topic tag discrimination is terminated, and the topic tag compression processing of the email subject corresponding to the current email retrieval request is performed based on the current topic tag compression setting, and the compressed email subject is vectorized for retrieval.

8. The vectorized retrieval optimization method for email scenarios according to claim 6, characterized in that, Step S4 includes: S41. Determine the historical email retrieval sample corresponding to the appropriate interval and use it as a reference sample; S42. Obtain the topic tag compression settings used by the reference sample in the historical retrieval process, and use the topic tag compression settings as the standard topic tag compression settings; S43. When there are multiple standard topic tag compression settings, calculate the similarity of topic semantic information between the core topic corresponding to each reference sample and the current core topic, and select the topic tag compression setting corresponding to the reference sample with the highest similarity of topic semantic information as the standard topic tag compression setting.

9. The vectorized retrieval optimization method for email scenarios according to claim 1, characterized in that, Step S5 specifically includes: S51. According to the standard subject tag compression settings, merge, retain or delete the subject tags in the subject of the email to be retrieved corresponding to the current email retrieval request to obtain the compressed email subject. S52. Input the compressed email subject into the vectorization model to generate an email subject vector, and perform similarity retrieval based on the email subject vector.

10. A vectorized retrieval optimization system for email scenarios, characterized in that, The system includes: The core topic determination module is used to determine the current core topic corresponding to the current email retrieval request and obtain historical email retrieval samples with the same core topic as the current core topic. The samples include historical target emails and historical interference emails with similar core topics to the historical target emails. The sample parameter calculation module is used to determine the subject tag discrimination degree of each sample based on the distinction status of historical target emails and historical interference emails after subject tag compression, and to calculate the retrieval load value of each sample. The appropriate interval determination module is used to determine the appropriate interval of topic tag discrimination based on the correspondence between the topic tag discrimination of each sample and the retrieval load value. The standard setting determination module is used to determine the corresponding topic tag compression setting from samples whose topic tag discrimination is within a suitable range, and use it as the standard topic tag compression setting; The vectorized retrieval module is used to compress the subject of the email to be retrieved corresponding to the current email retrieval request based on the standard subject tag compression settings, and then perform vectorized retrieval on the compressed email subject.

Images