An e-commerce information rapid retrieval method based on deep learning
By using deep learning-based semantic fragment modeling and set-level distance ranking, the problem of matching key attributes of queries and product information in e-commerce platforms was solved, enabling fast and accurate product retrieval and improving the accuracy and reliability of retrieval results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- YUNLIAN SHENGSHI (FUJIAN) TECH CO LTD
- Filing Date
- 2026-02-26
- Publication Date
- 2026-06-02
Smart Images
Figure CN122132552A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of e-commerce information retrieval technology, and in particular to a method for rapid e-commerce information retrieval based on deep learning. Background Technology
[0002] Traditional e-commerce platforms typically rely on inverted indexes, keyword matching, and relevance ranking for product retrieval. To improve query understanding, deep learning semantic representation has been widely adopted in recent years. This involves encoding query text and product text into vectors and calculating similarity to achieve semantic-level retrieval and ranking. In real-world systems, product information comes from complex sources, including basic information such as titles, categories, and brands, as well as textual content such as specifications, attribute keys, and detailed descriptions. User queries are often short, conversational, and ambiguous. To balance response speed and recall effectiveness, existing solutions often employ a multi-stage retrieval structure of coarse recall followed by fine ranking. Candidates are first recalled using rapid similarity or indexing, and then ranked using a more complex model, thereby achieving high throughput and low latency in large-scale product databases.
[0003] In existing technical solutions, when the query and product information are composed of multiple semantic fragments, existing semantic retrieval methods often compress the entire text into a single vector or use simple aggregation, which makes it difficult to accurately constrain the matching relationship of key attribute fragments. This can easily lead to results where the overall semantics are similar but the key specifications, models, capacities, and compatible objects are incorrect. If a fine-grained comparison of the fragment-level representations is performed directly across the entire database, it will increase the computational load and affect the retrieval speed.
[0004] Therefore, how to provide a fast e-commerce information retrieval method based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] One objective of this invention is to propose a fast e-commerce information retrieval method based on deep learning. This invention fully utilizes semantic fragment modeling, metric learning, and a two-stage retrieval computation mechanism. It constructs a set of semantic fragments for products by using field parsing and segmentation processing to extract basic product information, attribute and specification information, and product text description information. It also obtains a set of query semantic fragments by segmenting the user's query text. The invention details the vectorized encoding process of training an improved Triplet network using triplet samples based on historical e-commerce retrieval data, achieving vector representation learning of the query and product fragment sets. Furthermore, it uses vector aggregation to obtain representative query vectors and representative product vectors to quickly filter candidate products. Finally, it employs set-level Hausdorff distance to finely rank the query vector set and product vector set on the candidate set. This ensures retrieval speed while strengthening the constraint on mismatched key attribute fragments, reducing the probability of overall similar but incorrect key attribute retrieval results. It possesses the advantages of fast retrieval speed, high matching accuracy, and scalable engineering deployment.
[0006] A method for fast e-commerce information retrieval based on deep learning according to an embodiment of the present invention includes: Obtain the e-commerce product information database, perform field parsing and segmentation on the product information in the e-commerce product information database, and obtain a set of product semantic fragments; An improved Triplet network was constructed. Triplet samples were built based on historical e-commerce retrieval data, and the improved Triplet network was trained to obtain the trained improved Triplet network. Collect user-input e-commerce search query text, segment the e-commerce search query text, and generate a set of query semantic fragments; Based on the improved Triplet network after training, the set of product semantic fragments and the set of query semantic fragments are vectorized and encoded to obtain the set of product vectors and the set of query vectors, and the set of product vectors is stored to form a product vector set library. Vector aggregation is performed on each vector in the product vector set of each product in the query vector set and product vector set library to obtain the query representative vector and the product representative vector. Calculate the vector similarity between the query representative vector and the product representative vector, and select a preset number of candidate products from the product vector set library based on the vector similarity calculation results; Based on the query vector set and the corresponding product vector set of the candidate products, the Hausdorff distance algorithm is used to calculate the set-level Hausdorff distance for each candidate product, and the candidate products are sorted according to the set-level Hausdorff distance. The preset number of products in the sorted results are then output as the search results.
[0007] Optionally, the e-commerce product information database specifically includes basic product information, attribute and specification information, and product text description information; Optionally, the obtained set of product semantic fragments includes: Obtain product information for each product in the e-commerce product information database, parse the product information fields, and generate field-based product information data; The text of each field in the fieldized product information data is segmented, and each field text is divided into multiple semantic segments. Semantic segment identifiers are assigned to each segment to obtain the set of semantic segments corresponding to each field. The semantic fragment sets corresponding to each field of the same product are merged to generate a product semantic fragment set.
[0008] Optionally, the obtained improved Triplet network after training includes: An improved Triplet network is constructed, which consists of a triplet input unit, a shared coding unit, a fragment set preservation unit, and an aggregation computation unit. The historical search query text is segmented to obtain a set of anchor point semantic fragments. Based on the preset positive interaction conditions, a set of positive sample product semantic fragments and a set of negative sample product semantic fragments are generated and combined to form a triplet sample set. For each triplet sample in the triplet sample set, vectorize each semantic fragment in the anchor point semantic fragment set, the positive sample product semantic fragment set, and the negative sample product semantic fragment set to obtain the anchor point fragment vector set, the positive sample fragment vector set, and the negative sample fragment vector set. Vector aggregation is performed on the anchor point fragment vector set, the positive sample fragment vector set, and the negative sample fragment vector set respectively to obtain the anchor point set representation vector, the positive sample set representation vector, and the negative sample set representation vector. The positive sample distance value between the anchor point set representation vector and the positive sample set representation vector, and the negative sample distance value between the anchor point set representation vector and the negative sample set representation vector are calculated using a preset distance function. The training loss of triplet is calculated based on a preset interval value, and the training loss of each triplet sample in the triplet sample set is summarized to obtain the overall training loss. The parameters of the improved Triplet network are updated based on the overall training loss to obtain the trained improved Triplet network.
[0009] Optionally, the generation of the query semantic fragment set includes: Collect the e-commerce search query text input by the user, and perform character normalization, whitespace processing and symbol processing on the e-commerce search query text to obtain standardized query text; The standardized query text is segmented into multiple query semantic fragments, and each query semantic fragment is assigned a fragment identifier to generate a set of query semantic fragments.
[0010] Optionally, obtaining the product vector set and the query vector set includes: The set of product semantic fragments is used as product input, and the set of query semantic fragments is used as query input. The triple input unit receives the product input and the query input respectively, and performs input organization processing on the product input and query input to generate a sequence of semantic fragments to be encoded. The shared coding unit performs vectorized encoding on each semantic segment in the sequence of semantic segments to be encoded, introduces a multi-scale coding channel to extract features from the same semantic segment, and fuses the encoding results output by different coding channels to obtain the corresponding segment vectors, and forms a set of product segment vectors and a set of query segment vectors respectively. The fragment set retention unit performs filtering on the product fragment vector set and the query fragment vector set, introduces a dynamic fragment filtering mechanism, calculates the semantic importance of each fragment, and removes fragment vectors below the threshold based on the semantic importance, thus obtaining the filtered product fragment vector set and query fragment vector set. The aggregation calculation unit performs vector aggregation processing on the filtered product fragment vector set output and the filtered query fragment vector set output, respectively, to obtain the product set representation vector and the query set representation vector. The product vector sets are stored according to product identifiers to form a product vector set library.
[0011] Optionally, obtaining the query representative vector and the product representative vector includes: Perform vector aggregation on each query vector in the query vector set to obtain the query representative vector; For each product in the product vector set library, obtain the product vector set corresponding to the product, and perform vector aggregation on each product vector in the product vector set to obtain the product representative vector corresponding to the product.
[0012] Optionally, the step of selecting a preset number of candidate products from the product vector set library based on the vector similarity calculation results includes: Based on the query representative vector and the product representative vector, a similarity function is used to calculate the vector similarity score between the query representative vector and each product representative vector; The products in the product vector set library are sorted according to their vector similarity scores, and a preset number of products with the highest similarity scores are selected from the sorting results as a candidate product set.
[0013] Optionally, a preset number of items in the output sorting results are used as search results, including: Based on the query vector set, obtain the product vector set corresponding to each candidate product in the candidate product set; For each candidate product, the inter-vector distance between each query vector in the query vector set and each product vector in the product vector set corresponding to the candidate product is calculated one by one using a preset inter-vector distance function. The minimum inter-vector distance value corresponding to each query vector is then taken to obtain the set of minimum distance values corresponding to each query vector. The first direction set distance value is obtained by taking the maximum value of the minimum distance value set. The vector distance between each product vector in the product vector set corresponding to the candidate product and each query vector in the query vector set is calculated one by one using a preset vector distance function. The minimum distance value is taken for the vector distance corresponding to each product vector to obtain the minimum distance value set corresponding to each product vector. The second direction set distance value is obtained by taking the maximum value of the minimum distance value set corresponding to each product vector. The larger of the first direction set distance value and the second direction set distance value is determined as the set-level Hausdorff distance between the query vector set and the product vector set corresponding to the candidate product, and the set-level Hausdorff distance is determined as the fine ranking distance score of the candidate product. The candidate products are sorted according to their fine ranking distance scores, and a preset number of candidate products with the smallest fine ranking distance scores are selected from the sorted results and output as the search results.
[0014] The beneficial effects of this invention are: This invention proposes a fast e-commerce information retrieval method based on deep learning. It adopts a retrieval method that combines semantic fragment modeling, metric learning and set-level distance ranking to uniformly model the complex scenario in e-commerce retrieval where query and product information are composed of multiple semantic fragments.
[0015] This invention constructs a set of semantic fragments for products by parsing and segmenting the basic information, attributes, specifications, and textual descriptions of products in an e-commerce product information database. It also segments the user's search query text to obtain a set of semantic fragments for the query. Using historical e-commerce search data, it constructs triplet samples and trains a Triplet network, enabling the network to learn an effective distance metric between query and product semantic fragments, thus achieving a vectorized representation of products and queries. In the retrieval stage, this invention first achieves rapid filtering of candidate products through vector aggregation and vector similarity calculation. Then, within the candidate product range, it introduces set-level Hausdorff distance to perform fine-grained matching and ranking of the query vector set and the product vector set.
[0016] The method of this invention can effectively constrain the mismatch problem of key attribute fragments while ensuring the retrieval response speed, reduce the occurrence of retrieval results with similar overall semantics but incorrect key attributes, improve the accuracy and reliability of retrieval results, and at the same time have good engineering scalability and practical application value. Attached Figure Description
[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a deep learning-based method for fast e-commerce information retrieval proposed in this invention; Figure 2 This is a schematic diagram of the improved Triplet network structure of a fast e-commerce information retrieval method based on deep learning proposed in this invention. Figure 3 This is a flowchart of the Hausdorff distance algorithm used in a deep learning-based fast e-commerce information retrieval method proposed in this invention. Detailed Implementation
[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0019] refer to Figure 1 , Figure 2 and Figure 3 A method for fast e-commerce information retrieval based on deep learning, comprising: Obtain the e-commerce product information database, perform field parsing and segmentation on the product information in the e-commerce product information database, and obtain a set of product semantic fragments; An improved Triplet network was constructed. Triplet samples were built based on historical e-commerce retrieval data, and the improved Triplet network was trained to obtain the trained improved Triplet network. Collect user-input e-commerce search query text, segment the e-commerce search query text, and generate a set of query semantic fragments; Based on the improved Triplet network after training, the set of product semantic fragments and the set of query semantic fragments are vectorized and encoded to obtain the set of product vectors and the set of query vectors, and the set of product vectors is stored to form a product vector set library. Vector aggregation is performed on each vector in the product vector set of each product in the query vector set and product vector set library to obtain the query representative vector and the product representative vector. Calculate the vector similarity between the query representative vector and the product representative vector, and select a preset number of candidate products from the product vector set library based on the vector similarity calculation results; Based on the query vector set and the corresponding product vector set of the candidate products, the Hausdorff distance algorithm is used to calculate the set-level Hausdorff distance for each candidate product, and the candidate products are sorted according to the set-level Hausdorff distance. The preset number of products in the sorted results are then output as the search results.
[0020] In this embodiment, the e-commerce product information database specifically includes basic product information, attribute and specification information, and product text description information; In this embodiment, obtaining the set of product semantic fragments includes: Obtain product information for each product in the e-commerce product information database, parse the product information fields, and generate field-based product information data; The text of each field in the fieldized product information data is segmented, and each field text is divided into multiple semantic segments. Semantic segment identifiers are assigned to each segment to obtain the set of semantic segments corresponding to each field. The semantic fragment sets corresponding to each field of the same product are merged to generate a product semantic fragment set.
[0021] In this embodiment, obtaining the trained improved Triplet network includes: An improved Triplet network is constructed, which consists of a triplet input unit, a shared coding unit, a fragment set preservation unit, and an aggregation computation unit. The historical search query text is segmented to obtain a set of anchor point semantic fragments. Based on the preset positive interaction conditions, a set of positive sample product semantic fragments and a set of negative sample product semantic fragments are generated and combined to form a triplet sample set. For each triplet sample in the triplet sample set, vectorize each semantic fragment in the anchor point semantic fragment set, the positive sample product semantic fragment set, and the negative sample product semantic fragment set to obtain the anchor point fragment vector set, the positive sample fragment vector set, and the negative sample fragment vector set. Vector aggregation is performed on the anchor point fragment vector set, the positive sample fragment vector set, and the negative sample fragment vector set respectively to obtain the anchor point set representation vector, the positive sample set representation vector, and the negative sample set representation vector. The positive sample distance value between the anchor point set representation vector and the positive sample set representation vector, and the negative sample distance value between the anchor point set representation vector and the negative sample set representation vector are calculated using a preset distance function. The training loss of triplet is calculated based on a preset interval value, and the training loss of each triplet sample in the triplet sample set is summarized to obtain the overall training loss. The parameters of the improved Triplet network are updated based on the overall training loss to obtain the trained improved Triplet network.
[0022] In this embodiment, generating the query semantic fragment set includes: Collect the e-commerce search query text input by the user, and perform character normalization, whitespace processing and symbol processing on the e-commerce search query text to obtain standardized query text; The standardized query text is segmented into multiple query semantic fragments, and each query semantic fragment is assigned a fragment identifier to generate a set of query semantic fragments.
[0023] In this embodiment, obtaining the product vector set and the query vector set includes: The set of product semantic fragments is used as product input, and the set of query semantic fragments is used as query input. The triple input unit receives the product input and the query input respectively, and performs input organization processing on the product input and query input to generate a sequence of semantic fragments to be encoded. The shared coding unit performs vectorized encoding on each semantic segment in the sequence of semantic segments to be encoded, introduces a multi-scale coding channel to extract features from the same semantic segment, and fuses the encoding results output by different coding channels to obtain the corresponding segment vectors, and forms a set of product segment vectors and a set of query segment vectors respectively. The fragment set retention unit performs filtering on the product fragment vector set and the query fragment vector set, introduces a dynamic fragment filtering mechanism, calculates the semantic importance of each fragment, and removes fragment vectors below the threshold based on the semantic importance, thus obtaining the filtered product fragment vector set and query fragment vector set. The aggregation calculation unit performs vector aggregation processing on the filtered product fragment vector set output and the filtered query fragment vector set output, respectively, to obtain the product set representation vector and the query set representation vector. The product vector sets are stored according to product identifiers to form a product vector set library.
[0024] In this embodiment, obtaining the query representative vector and the product representative vector includes: Perform vector aggregation on each query vector in the query vector set to obtain the query representative vector; For each product in the product vector set library, obtain the product vector set corresponding to the product, and perform vector aggregation on each product vector in the product vector set to obtain the product representative vector corresponding to the product.
[0025] In this embodiment, selecting a preset number of candidate products from the product vector set library based on the vector similarity calculation results includes: Based on the query representative vector and the product representative vector, a similarity function is used to calculate the vector similarity score between the query representative vector and each product representative vector; The products in the product vector set library are sorted according to their vector similarity scores, and a preset number of products with the highest similarity scores are selected from the sorting results as a candidate product set.
[0026] In this embodiment, the preset number of products in the output sorting results are used as search results, including: Based on the query vector set, obtain the product vector set corresponding to each candidate product in the candidate product set; For each candidate product, the inter-vector distance between each query vector in the query vector set and each product vector in the product vector set corresponding to the candidate product is calculated one by one using a preset inter-vector distance function. The minimum inter-vector distance value corresponding to each query vector is then taken to obtain the set of minimum distance values corresponding to each query vector. The first direction set distance value is obtained by taking the maximum value of the minimum distance value set. The vector distance between each product vector in the product vector set corresponding to the candidate product and each query vector in the query vector set is calculated one by one using a preset vector distance function. The minimum distance value is taken for the vector distance corresponding to each product vector to obtain the minimum distance value set corresponding to each product vector. The second direction set distance value is obtained by taking the maximum value of the minimum distance value set corresponding to each product vector. The larger of the first direction set distance value and the second direction set distance value is determined as the set-level Hausdorff distance between the query vector set and the product vector set corresponding to the candidate product, and the set-level Hausdorff distance is determined as the fine ranking distance score of the candidate product. The candidate products are sorted according to their fine ranking distance scores, and a preset number of candidate products with the smallest fine ranking distance scores are selected from the sorted results and output as the search results.
[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to a search experiment on an e-commerce platform, taking home appliance category as an example. Users searched for "drum washing machine," "10kg," "energy efficiency level 1," and "quiet" on the e-commerce platform. The platform's product database contained a large number of products of the same brand, similar appearance, and price range, but with different capacities or energy efficiency levels. Traditional solutions based on whole-text vector matching or simple keyword matching tend to return results that look similar overall but have inconsistent key attributes. Users found that the attributes did not match after clicking on the details, leading to an increased return rate and a decrease in conversion.
[0028] In this scenario, the method of this invention first performs field parsing and segmentation on the basic information, attributes, specifications, and textual descriptions of products in the product information database. Brand, capacity, energy efficiency, noise level, model, etc., are split into a set of alignable semantic fragments, which are then encoded offline to form a product vector set library. After going online, the user-input query text is also segmented into a set of query semantic fragments and encoded to obtain a query vector set. The query vector set and each product vector set are then aggregated to obtain a representative query vector and a representative product vector. Based on vector similarity, a candidate set is quickly selected. Then, for each candidate product, the query fragment set and product fragment set are finely ranked using set-level Hausdorff distance. Finally, the top-ranked results are output, thus imposing constraints on mismatches in key attribute fragments during the fine-ranking stage.
[0029] Offline playback and low-traffic online A / B testing were conducted in a research and development data center. The offline playback data was taken from the home appliance search logs of the past 30 days, including approximately 1.2 million queries, approximately 8.6 million impressions, and approximately 740,000 clicks, with a fixed candidate set size of 300. Compared with the baseline that only uses representative vector similarity ranking, this method reduced the proportion of key attribute error results from 6.8% to 2.1% at the same recall scale, a relative decrease of approximately 69.1%. The NDCG of the Top 10 increased from 0.421 to 0.468, and the MRR increased from 0.312 to 0.347.
[0030] In terms of performance, under the same hardware, the average retrieval time increased from 78 milliseconds to 92 milliseconds, with the Hausdorff calculation for fine-grained ranking increasing by an average of 11 milliseconds, but the overall P95 was still kept below 165 milliseconds. Online 1% traffic validation showed that the return rate of search results pages decreased from 18.4% to 15.7%, and the add-to-cart conversion rate increased from 3.2% to 3.6%, validating the improvement in key attribute consistency while maintaining retrieval speed.
[0031] Table 1. Comparison of the effectiveness and performance of the proposed method and the baseline method in the home appliance retrieval scenario.
[0032] As shown in Table 1, during the validation period, the method of this invention exhibited stable and significant improvements in both relevance and critical error control. Regarding overall relevance, the method improved the Top 10 NDCG from 0.421 to 0.468, an improvement of 0.047, and the MRR from 0.312 to 0.347, an improvement of 0.035. This indicates that under the same search path, users are more likely to obtain product results matching their intent at higher positions, and the probability of hitting a valid result on the first attempt is higher, thus enhancing the effectiveness of search ranking.
[0033] This invention provides a direct suppression effect on the problem of overall similarity but incorrect key attributes in e-commerce retrieval. The proportion of results with incorrect key attributes in the Top 10 range decreased from 6.8% to 2.1%, a decrease of 4.7 percentage points, or about 69.1%, demonstrating the ability of set-level fine ranking to constrain key mismatch segments.
[0034] Further observation at the attribute level reveals that the capacity conflict error rate decreased from 3.1% to 0.9%, the energy efficiency conflict error rate decreased from 2.4% to 0.7%, and the model conflict error rate decreased from 1.3% to 0.5%. All three types of conflicts showed significant reductions, indicating that the method not only reduces attribute mismatches overall but also covers multiple high-frequency hard constraint fields in e-commerce, avoiding situations where it is effective for only one attribute but ineffective for others.
[0035] Online behavioral metrics also validated the actual benefits of the improvements. The detail page return rate decreased from 18.4% to 15.7%, reflecting a reduction in the number of users who clicked through but found the page did not meet their expectations and returned. The add-to-cart conversion rate increased from 3.2% to 3.6%, indicating that the improved availability of search results further boosted transaction intentions.
[0036] In terms of performance, the average retrieval time of this invention increased from 78ms to 92ms, the additional time for fine ranking was 11ms, and P95 increased from 148ms to 165ms. Although the introduction of set-level fine ranking brings certain computational overhead, the increase is within a controllable range. In addition, combined with the significant reduction in the error rate of key attributes and the synchronous improvement of user behavior indicators, this invention can demonstrate that it has achieved a better engineering balance between speed and accuracy.
[0037] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for rapid e-commerce information retrieval based on deep learning, characterized in that, include: Obtain the e-commerce product information database, perform field parsing and segmentation on the product information in the e-commerce product information database, and obtain a set of product semantic fragments; An improved Triplet network was constructed. Triplet samples were built based on historical e-commerce retrieval data, and the improved Triplet network was trained to obtain the trained improved Triplet network. Collect user-input e-commerce search query text, segment the e-commerce search query text, and generate a set of query semantic fragments; Based on the improved Triplet network after training, the set of product semantic fragments and the set of query semantic fragments are vectorized and encoded to obtain the set of product vectors and the set of query vectors, and the set of product vectors is stored to form a product vector set library. Vector aggregation is performed on each vector in the product vector set of each product in the query vector set and product vector set library to obtain the query representative vector and the product representative vector. Calculate the vector similarity between the query representative vector and the product representative vector, and select a preset number of candidate products from the product vector set library based on the vector similarity calculation results; Based on the query vector set and the corresponding product vector set of the candidate products, the Hausdorff distance algorithm is used to calculate the set-level Hausdorff distance for each candidate product, and the candidate products are sorted according to the set-level Hausdorff distance. The preset number of products in the sorted results are then output as the search results.
2. The method for rapid e-commerce information retrieval based on deep learning according to claim 1, characterized in that, The e-commerce product information database specifically includes basic product information, attribute and specification information, and product text description information; According to claim 1, the method for fast e-commerce information retrieval based on deep learning is characterized in that obtaining the set of product semantic fragments includes: Obtain product information for each product in the e-commerce product information database, parse the product information fields, and generate field-based product information data; The text of each field in the fieldized product information data is segmented, and each field text is divided into multiple semantic segments. Semantic segment identifiers are assigned to each segment to obtain the set of semantic segments corresponding to each field. The semantic fragment sets corresponding to each field of the same product are merged to generate a product semantic fragment set.
3. The method for rapid e-commerce information retrieval based on deep learning according to claim 1, characterized in that, The improved Triplet network obtained after training includes: An improved Triplet network is constructed, which consists of a triplet input unit, a shared coding unit, a fragment set preservation unit, and an aggregation computation unit. The historical search query text is segmented to obtain a set of anchor point semantic fragments. Based on the preset positive interaction conditions, a set of positive sample product semantic fragments and a set of negative sample product semantic fragments are generated and combined to form a triplet sample set. For each triplet sample in the triplet sample set, vectorize each semantic fragment in the anchor point semantic fragment set, the positive sample product semantic fragment set, and the negative sample product semantic fragment set to obtain the anchor point fragment vector set, the positive sample fragment vector set, and the negative sample fragment vector set. Vector aggregation is performed on the anchor point fragment vector set, the positive sample fragment vector set, and the negative sample fragment vector set respectively to obtain the anchor point set representation vector, the positive sample set representation vector, and the negative sample set representation vector. The positive sample distance value between the anchor point set representation vector and the positive sample set representation vector, and the negative sample distance value between the anchor point set representation vector and the negative sample set representation vector are calculated using a preset distance function. The training loss of triplet is calculated based on a preset interval value, and the training loss of each triplet sample in the triplet sample set is summarized to obtain the overall training loss. The parameters of the improved Triplet network are updated based on the overall training loss to obtain the trained improved Triplet network.
4. The method for rapid e-commerce information retrieval based on deep learning according to claim 1, characterized in that, The generated set of query semantic fragments includes: Collect the e-commerce search query text input by the user, and perform character normalization, whitespace processing and symbol processing on the e-commerce search query text to obtain standardized query text; The standardized query text is segmented into multiple query semantic fragments, and each query semantic fragment is assigned a fragment identifier to generate a set of query semantic fragments.
5. The method for rapid e-commerce information retrieval based on deep learning according to claim 1, characterized in that, The process of obtaining the product vector set and the query vector set includes: The set of product semantic fragments is used as product input, and the set of query semantic fragments is used as query input. The triple input unit receives the product input and the query input respectively, and performs input organization processing on the product input and query input to generate a sequence of semantic fragments to be encoded. The shared coding unit performs vectorized encoding on each semantic segment in the sequence of semantic segments to be encoded, introduces a multi-scale coding channel to extract features from the same semantic segment, and fuses the encoding results output by different coding channels to obtain the corresponding segment vectors, and forms a set of product segment vectors and a set of query segment vectors respectively. The fragment set retention unit performs filtering on the product fragment vector set and the query fragment vector set, introduces a dynamic fragment filtering mechanism, calculates the semantic importance of each fragment, and removes fragment vectors below the threshold based on the semantic importance, thus obtaining the filtered product fragment vector set and query fragment vector set. The aggregation calculation unit performs vector aggregation processing on the filtered product fragment vector set output and the filtered query fragment vector set output, respectively, to obtain the product set representation vector and the query set representation vector. The product vector sets are stored according to product identifiers to form a product vector set library.
6. The method for rapid e-commerce information retrieval based on deep learning according to claim 1, characterized in that, The process of obtaining the query representative vector and the product representative vector includes: Perform vector aggregation on each query vector in the query vector set to obtain the query representative vector; For each product in the product vector set library, obtain the product vector set corresponding to the product, and perform vector aggregation on each product vector in the product vector set to obtain the product representative vector corresponding to the product.
7. The method for rapid e-commerce information retrieval based on deep learning according to claim 1, characterized in that, The selection of a preset number of candidate products from the product vector set library based on the vector similarity calculation results includes: Based on the query representative vector and the product representative vector, a similarity function is used to calculate the vector similarity score between the query representative vector and each product representative vector; The products in the product vector set library are sorted according to their vector similarity scores, and a preset number of products with the highest similarity scores are selected from the sorting results as a candidate product set.
8. The method for rapid e-commerce information retrieval based on deep learning according to claim 1, characterized in that, The preset number of products in the output sorting results are used as search results, including: Based on the query vector set, obtain the product vector set corresponding to each candidate product in the candidate product set; For each candidate product, the inter-vector distance between each query vector in the query vector set and each product vector in the product vector set corresponding to the candidate product is calculated one by one using a preset inter-vector distance function. The minimum inter-vector distance value corresponding to each query vector is then taken to obtain the set of minimum distance values corresponding to each query vector. The first direction set distance value is obtained by taking the maximum value of the minimum distance value set. The vector distance between each product vector in the product vector set corresponding to the candidate product and each query vector in the query vector set is calculated one by one using a preset vector distance function. The minimum distance value is taken for the vector distance corresponding to each product vector to obtain the minimum distance value set corresponding to each product vector. The second direction set distance value is obtained by taking the maximum value of the minimum distance value set corresponding to each product vector. The larger of the first direction set distance value and the second direction set distance value is determined as the set-level Hausdorff distance between the query vector set and the product vector set corresponding to the candidate product, and the set-level Hausdorff distance is determined as the fine ranking distance score of the candidate product. The candidate products are sorted according to their fine ranking distance scores, and a preset number of candidate products with the smallest fine ranking distance scores are selected from the sorted results and output as the search results.