Automobile industry search sug recall word generation method

Abstract

The invention belongs to the technical field of search engines, and discloses an automobile industry search sug recall word generation method, which specifically comprises the following steps of: 1, preparing an automobile corpus, and training a pre-training model of the t5 automobile corpus; compared with a traditional frequency statistical method based on user search query, the search sug recall word generation method has the advantages that some low-frequency user search query can be mined to supplement sug recall, knowledge in the automobile field can be learned through pre-training of the t5 model, and through the fine training sample pair, the search sug recall word generation method can be applied to the automobile field. Therefore, the recall word highly related to the search prefix of the user can be generated, through further optimization of a search engine system, high quality of a recall result is ensured, the generation quality of the recall word is improved, the ability of model learning to generate a high-quality query can be achieved through addition of reinforcement learning, generation of cases not conforming to the fact basis is reduced, and the efficiency of the search engine system is improved. And the correlation and accuracy of the recalled words are further improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of search engines, and specifically relates to a method for generating recall words for automotive industry search sug. Background Art

[0002] The core of search term auto-completion is to provide relevant search suggestions based on the user's input. Among them, "recall" refers to a set of possible search terms or phrases selected from a candidate word library.

[0003] The mining of recall words is mainly achieved through the following steps: data collection, candidate word generation, and statistical analysis. It should be noted that the mining of recall words is a complex process that requires comprehensive consideration of multiple factors, such as data quality, the selection and optimization of statistical algorithms, etc. Common recall word generation methods include: frequency statistics-based, association rule-based, language model-based, and Seq2Seq-based SUG word generation methods. Among them, the frequency statistics method tends to recall words with higher frequencies in the vocabulary, but this does not mean that these words are the most accurate or comprehensive. Some special and low-frequency words may be ignored, resulting in incomplete information coverage. For language model-based methods, a high-quality language model needs to be constructed and trained with a large amount of text data. If there is less available user search history data, it is difficult to obtain enough samples to build an accurate language model. Although a trained language model can generate various possible recall words, including common and uncommon ones, sometimes we hope to generate recall words related to a specific field or task. The language model may be too general in this regard and cannot well meet specific needs. The language model may generate some ambiguous or incorrect recall words, so it needs to be improved. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides a method for generating recall words for automotive industry search sug, which has the advantages of improving the quality and accuracy of recall word generation.

[0005] To achieve the above object, the present invention provides the following technical solution: A method for generating recall words for automotive industry search sug, the specific steps are as follows:

[0006] Step 1: Prepare automotive corpus and train a pre-trained model for t5 automotive corpus.

[0007] Step 2: Construct training sample pairs for the t5 generation model according to the recall word library. For example, for 'BMW X3' in the recall word library, sample pairs composed of prefixes and recall words such as 'bao', 'BMW X3', 'BMW', 'BMW X3', 'BMW X', 'BMW X3' can be segmented.

[0008] Step 3: Fine-tune the sug recall word generation model based on the t5 pre-trained model of the automobile corpus;

[0009] Step 4: Obtain the user search prefix word based on the search completion user log;

[0010] Step 5: The user searches for a prefix word and inputs the recall word generation model, generating the top 50 results based on the beam search algorithm;

[0011] Step 6: The generated recall words are passed through the search engine system to obtain the top 100 non-plaintext recall results;

[0012] Step 7: Based on the trained sug recall word generation model, train the reinforcement model, take the number of materials with non-plaintext recall scores greater than 0.7 / 100 as the reward, and train the reinforcement model;

[0013] Step 8: Use the generative model after intensive training to predict the user's prefix words and generate the top 10 results;

[0014] Step 9: Filter the generated results. The generated words must pass through the search engine and the recalled high-quality materials must meet a certain number. At the same time, the rules filter the generated words with low quality.

[0015] Step 10: The filtered generated words are added to the original recall word library for searching sug recall.

[0016] Preferably, the corpus data described in step 1 needs to be cleaned and preprocessed before model training to remove noise data and duplicate data, and unify the format and coding standards.

[0017] Preferably, the training sample pairs in step 2 need to be diverse, covering prefix and recall word combinations of different lengths and complexities, so as to improve the generalization ability of the model.

[0018] Preferably, when fine-tuning the t5 pre-training model described in step three, it is necessary to perform fine-tuning according to the characteristics of recall word generation.

[0019] Preferably, when processing user logs as described in step 4, it is necessary to strictly comply with privacy protection regulations to ensure user data security.

[0020] Preferably, the search engine described in step six needs to ensure that it can correctly process and parse the recall words.

[0021] Preferably, the top 10 result sorting described in step eight needs to ensure that the recall words most relevant to the user prefix word are ranked first.

[0022] Preferably, the quantity described in step nine can be adjusted according to the actual situation during actual application.

[0023] Preferably, it is necessary to monitor the updated recall word library in step ten to promptly discover and solve possible problems.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] This solution provides a method for generating search sug recall words. Compared with the traditional method based on frequency statistics of user search queries, this method can mine some low-frequency user search queries to supplement sug recall. Through pre-training of the t5 model, knowledge in the automotive field can be learned, and through fine training sample pairs, recall words highly relevant to the user search prefix can be generated. After further optimization by the search engine system, the quality of the recall results is ensured, the quality of recall word generation is improved, and the addition of reinforcement learning enables the model to learn the ability to generate high-quality queries, reducing the generation of cases that do not conform to the factual basis, further enhancing the relevance and accuracy of the recall words, and ultimately providing a more accurate and fast search experience for users. Specific embodiments

[0026] Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0027] The present invention provides a technical solution: a method for generating search sug recall words in the automotive industry, and the specific steps are as follows:

[0028] Step 1: Prepare automotive corpus and train a pre-trained model for t5 automotive corpus.

[0029] Step 2: Construct training sample pairs for the t5 generation model according to the recall word library. For example, for 'BMW X3' in the recall word library, sample pairs composed of prefixes and recall words such as 'bao', 'BMW X3', 'BMW', 'BMW X3', 'BMW X', 'BMW X3' can be segmented.

[0030] Step 3: Fine-tune and train a sug recall word generation model on the basis of the t5 pre-trained model for automotive corpus.

[0031] Step 4: Obtain the user search prefix word according to the user log of search completion.

[0032] Step 5: Input the user search prefix word into the recall word generation model, and generate top 50 results according to the beam search algorithm.

[0033] Step 6: The generated recall words are passed through the search engine system to obtain the top 100 non-plaintext recall results;

[0034] Step 7: Based on the trained sug recall word generation model, train the reinforcement model, take the number of materials with non-plaintext recall scores greater than 0.7 / 100 as the reward, and train the reinforcement model;

[0035] Step 8: Use the generative model after intensive training to predict the user's prefix words and generate the top 10 results;

[0036] Step 9: Filter the generated results. The generated words must pass through the search engine and the recalled high-quality materials must meet a certain number. At the same time, the rules filter the generated words with low quality.

[0037] Step 10: The filtered generated words are added to the original recall word library for searching sug recall.

[0038] The t5 model is pre-trained with automobile corpus, so that the model can learn professional knowledge of the automobile industry. Then, through fine-tuning of the generative model, analysis of user behavior logs and reinforcement learning optimization, efficient generation and accurate recommendation of sug recall words for automobile industry search are achieved, which effectively improves the relevance and accuracy of search results. At the same time, the generated results are filtered to effectively avoid the generation of recall words that are not in line with the facts.

[0039] Among them, the corpus data in step one needs to be cleaned and preprocessed before model training to remove noise data and duplicate data, and unify the format and encoding standards.

[0040] By organizing the data, the training effect of the pre-training model is more significant.

[0041] Among them, the training sample pairs in step 2 need to be diverse, covering prefix and recall word combinations of different lengths and complexities to improve the generalization ability of the model.

[0042] Using data from the actual recall vocabulary can ensure that the model learns real and useful generation patterns, thereby improving the accuracy and practicality of generated recall words.

[0043] Among them, when fine-tuning the t5 pre-training model in step three, it is necessary to fine-tune it according to the characteristics of recall word generation.

[0044] Fine-tuning can enable the model to quickly adapt to recall word generation, improve the efficiency and effectiveness of task completion, and generate recall words that are more in line with user expectations.

[0045] Among them, when processing user logs in step 4, it is necessary to strictly comply with privacy protection regulations to ensure the security of user data.

[0046] By protecting user privacy data, user data leakage can be avoided.

[0047] Among them, the search engine in step six needs to ensure that it can correctly process and parse the recall words.

[0048] Combined with the search engine's indexing and ranking algorithms, it is possible to more accurately assess the quality and value of each recalled word, thereby providing users with more precise and useful search suggestions.

[0049] Among them, the top 10 result sorting in step 8 needs to ensure that the recall words most relevant to the user's prefix words are ranked first.

[0050] By sorting the results, we can ensure that the results are more in line with user needs and expectations.

[0051] Among them, the quantity in step nine can be adjusted according to actual conditions during actual application.

[0052] By filtering high-quality recall words, we ensure that the recall words finally presented to users meet both quality requirements and have a certain quantity guarantee.

[0053] Among them, in step ten, the updated recall vocabulary needs to be monitored so as to promptly discover and solve possible problems.

[0054] Adding newly generated, verified, high-quality recall words to the original recall word library can continuously enrich and improve the content of search suggestions.

[0055] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0056] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for generating search sug recall words in the automotive industry, characterized in that: The specific steps are as follows: Step 1: Prepare automotive corpus and train a pre-trained model for the t5 automotive corpus; Step 2: Construct training sample pairs for the t5 generation model according to the recall word library. For example, for 'BMW X3' in the recall word library, sample pairs composed of prefixes and recall words such as 'bao', 'BMW X3', 'BMW', 'BMW X3', 'BMW X', 'BMW X3' can be segmented; Step 3: Fine-tune and train the sug recall word generation model on the basis of the t5 pre-trained model for automotive corpus; Step 4: Obtain the user's search prefix word according to the user's log of search completion; Step 5: Input the user's search prefix word into the recall word generation model and generate top 50 results according to the beam search algorithm; Step 6: The generated recall words pass through the search engine system to obtain the top 100 non-plaintext recall results; Step 7: Train a reinforcement model on the basis of the trained sug recall word generation model, and take the number of materials with non-plaintext recall scores greater than 0.7 / 100 as the reward to train the reinforcement model; Step 8: Use the generated model after reinforcement training to make predictions on the user's prefix word and generate top 10 results; Step 9: Filter the generated results. The generated words pass through the search engine and need to meet a certain number of high-quality materials for recall; at the same time, filter out low-quality generated words according to the rules; Step 10: Add the filtered generated words to the original recall word library for search sug recall.

2. A method for generating search sug recall words in the automotive industry according to claim 1, characterized in that: The corpus data described in Step 1 needs to be cleaned and preprocessed before training the model to remove noise data and duplicate data, and unify the format and encoding standard.

3. A method for generating search sug recall words in the automotive industry according to claim 1, characterized in that: The training sample pairs described in Step 2 need to be diverse, covering prefix and recall word combinations of different lengths and different complexities to improve the generalization ability of the model.

4. A method for generating search sug recall words in the automotive industry according to claim 1, characterized in that: When fine-tuning the t5 pre-trained model in Step 3, it is necessary to fine-tune according to the characteristics of recall word generation.

5. A method for generating search sug recall words in the automotive industry according to claim 1, characterized in that: When processing the user's log in Step 4, it is necessary to strictly abide by privacy protection regulations to ensure the security of user data.

6. A method for generating search sug recall words in the automotive industry according to claim 1, characterized in that: The search engine described in Step 6 needs to ensure that it can correctly process and parse recall words.

7. A method for generating search sug recall words in the automotive industry according to claim 1, characterized in that: The sorting of the top 10 results described in Step 8 needs to ensure that the recall words most relevant to the user's prefix word are ranked in the front.

8. A method for generating search sug recall words in the automotive industry according to claim 1, characterized in that: The quantity described in step nine can be adjusted according to actual conditions during actual application.

9. A method for generating recall words for searching sug in the automobile industry according to claim 1, Features: As described in step 10, the updated recall vocabulary needs to be monitored so as to promptly discover and resolve possible problems.