CN115203589A - Vector searching method and system based on Trans-dssm model - Google Patents
Vector searching method and system based on Trans-dssm model Download PDFInfo
- Publication number
- CN115203589A CN115203589A CN202210709107.8A CN202210709107A CN115203589A CN 115203589 A CN115203589 A CN 115203589A CN 202210709107 A CN202210709107 A CN 202210709107A CN 115203589 A CN115203589 A CN 115203589A
- Authority
- CN
- China
- Prior art keywords
- search
- vector
- hotel
- model
- trans
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 239000013598 vector Substances 0.000 title claims abstract description 146
- 238000000034 method Methods 0.000 title claims abstract description 28
- 239000011159 matrix material Substances 0.000 claims abstract description 16
- 238000004364 calculation method Methods 0.000 claims abstract description 6
- 238000006243 chemical reaction Methods 0.000 claims abstract description 4
- 238000012549 training Methods 0.000 claims description 36
- 230000011218 segmentation Effects 0.000 claims description 11
- 210000002569 neuron Anatomy 0.000 claims description 10
- 230000008569 process Effects 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 2
- 230000001131 transforming effect Effects 0.000 claims 2
- 230000006870 function Effects 0.000 description 9
- 238000013461 design Methods 0.000 description 6
- 238000012216 screening Methods 0.000 description 6
- 238000011156 evaluation Methods 0.000 description 5
- 230000009466 transformation Effects 0.000 description 4
- 230000006872 improvement Effects 0.000 description 3
- 238000005457 optimization Methods 0.000 description 3
- 230000006835 compression Effects 0.000 description 2
- 238000007906 compression Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9537—Spatial or temporal dependent retrieval, e.g. spatiotemporal queries
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9532—Query formulation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/10—Services
- G06Q50/12—Hotels or restaurants
Landscapes
- Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Tourism & Hospitality (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Economics (AREA)
- General Health & Medical Sciences (AREA)
- Human Resources & Organizations (AREA)
- Marketing (AREA)
- Primary Health Care (AREA)
- Strategic Management (AREA)
- General Business, Economics & Management (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention provides a vector searching method and system based on a Trans-dssm model. The method comprises the following steps: inputting the search condition to a search side DNN of the Trans-dssm model, and compressing the search condition into a search vector through the search side DNN; inputting hotel attribute information to a hotel side DNN of a Trans-dssm model, and compressing the information into hotel vectors through the hotel side DNN; converting the search vector and the hotel vector into the same matrix space through a conversion matrix; calculating cosine similarity between the converted search vector and the hotel vector; sorting different hotel vectors by utilizing the cosine similarity obtained by calculation; and taking the hotels corresponding to the first N hotel vectors in the sequence as search result recall recommendations. The vector searching method and system based on the Trans-dssm model can solve the problem that no result or few results are generated in hotel searching.
Description
Technical Field
The invention relates to the technical field of network search, in particular to a vector search method and system based on a Trans-dssm model.
Background
Aiming at the problem of no result or few results in hotel search, a product strategy of abandoning some search conditions is adopted at present. And when no result or few results appear in hotel search, recommending a result similar to hotel search conditions as a recommending result of no result or few results. The recommended approach is to make the relatively precise search criteria a relatively broad search criteria by discarding some unimportant search criteria. For example, when the user searches for "beijing city + fengtai district + electric bed", because the "electric bed" search is relatively accurate and there is no "electric bed" hotel in the area range, the "electric bed" search can be abandoned to become a "beijing city + fengtai district" search condition, which is relatively wide compared with the original three condition search, and the search result of the wide search condition is taken as the recommendation result of the original accurate condition search, that is, the search result of "beijing city + fengtai district" is taken as the recommendation result of "beijing city + fengtai district + electric bed" with no or few results, and is suggested as "recommended for you".
Although the problem of no or little result can be solved to some extent by formulating a product strategy that rejects the search criteria, there are actually many problems with this strategy. The first order of abandonment of search criteria is difficult to define, often only established according to experience and personal understanding of products, and is not objective. Even if the appropriate search condition rejection order is defined, the effect is less satisfactory. Taking the search condition results of "Beijing City + Chaoyang district +1km Nei +5 star level" as an example of no result: if the search condition is abandoned according to a fixed sequence of star level-distance-my position-destination-city, the search condition is changed into search condition recommendation of Beijing City + sunny district +1km interior, obviously, low-end hotels can be recommended by a large amount of search, and are not ideal recall recommendation results; if the search condition is abandoned in a fixed sequence of distance-star-near-destination-city, the search condition is changed into 'Beijing City + sunny district +5 star-level' search condition recommendation, hotels of 5 star levels beyond 10km are even recalled and hotels beyond 10km are obviously less likely to be booked based on the current position. The ideal result should be a hotel with similar recommendation and search condition, i.e. similar to the search condition, which is a hotel approximately matched with the search condition. If the user search condition is "beijing city + sunny district +1km in +5 star class", the hotel with similar recall recommendation is more ideal recall recommendation result for the user if the hotel is "beijing city + sunny district +1km in +4 star class" or "beijing city + sunny district +2km in +5 star class".
When the hotel search has no result or few results, the hotel which is similar to the search condition is recommended through the product strategy, so that a semantic vector search model which is based on the search condition can be considered. The semantic vector search model compresses complex search conditions into semantic search vectors and compresses hotel related attribute information into hotel vectors. And when no result or few results exist in hotel search, performing similarity calculation by using the semantic search vector and the hotel vector to obtain the hotel vector most similar to the semantic search vector. Because the semantic search vector is the compression of the search condition information, the hotel vector is the compression of the hotel related attribute information, and the hotel vector most similar to the semantic search vector is the hotel most similar to the search condition. The construction of the hotel semantic retrieval model is a key for solving the problem of no result or few result recall recommendation in hotel search.
Disclosure of Invention
The invention aims to solve the technical problem of providing a vector searching method and system based on a Trans-dssm model, which can solve the problem of no result or few results in hotel searching.
In order to solve the technical problem, the invention provides a vector searching method based on a Trans-dssm model, which comprises the following steps: inputting the search condition to a search side DNN of the Trans-dssm model, and compressing the search condition into a search vector through the search side DNN; inputting hotel attribute information to a hotel side DNN of a Trans-dssm model, and compressing the information into hotel vectors through the hotel side DNN; converting the search vector and the hotel vector into the same matrix space through a conversion matrix; calculating cosine similarity between the converted search vector and the hotel vector; sorting different hotel vectors by utilizing the cosine similarity obtained by calculation; and taking the hotels corresponding to the first N hotel vectors in the sequence as search result recall recommendations.
In some embodiments, inputting the search condition to the search side DNN of the Trans-dssm model, compressing the search condition into a search vector by the search side DNN, comprises: converting the search condition into a feature vector through feature engineering; and splicing the feature vectors together to form input features of the search side DNN, and compressing the input search condition features by the search side DNN to output search vectors.
In some embodiments, converting the search criteria into a feature vector by feature engineering comprises: and normalizing the longitude and latitude parameters, and calculating the range of the latitude characteristic value in 400,5400 and the range of the longitude characteristic value in 7300,12500.
In some embodiments, converting the search criteria into a feature vector by feature engineering further comprises: and carrying out word segmentation and normalization on the keyword parameters, using word segmentation, wherein each word is an independent vector, and converting the word sequence into vector representation.
In some embodiments, the training data for the Trans-dssm model is a search log of the user.
In some embodiments, the Margin parameters of the triple loss function are accurately parametrized during training of the Trans-dssm model.
In some embodiments, during training of the Trans-dssm model, model checkpoints are saved.
In some embodiments, the checkpoint maintains model parameters, weights of the model, and a network structure of the model.
In some embodiments, the search-side DNN and the hotel-side DNN both use a three-layer neuron structure, the DNN network uses Dropout parameter settings, and the output values of the DNN neurons use the Relu activation function.
In addition, the invention also provides a vector search system based on the Trans-dssm model, which comprises: one or more processors; storage means for storing one or more programs which, when executed by the one or more processors, cause the one or more processors to implement a method for vector search based on the Trans-dssm model as described above.
After adopting such design, the invention has at least the following advantages:
1) The vector similarity recommendation model output by the Trans-dssm model is only relatively similar and has no strict search boundary, so that the condition of no search result does not exist;
2) Compared with the rule of model improvement and optimization and the double-tower model, the HR index is obviously improved, and the accuracy and the recall rate are better;
3) The vector search model is more intelligent and not simply whether there is a match.
Drawings
The foregoing is only an overview of the technical solutions of the present invention, and in order to make the technical solutions of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and the detailed description.
FIG. 1 is a block diagram of the Trans-dssm model;
fig. 2 is a flow chart of a retrieval prediction process.
Detailed Description
The preferred embodiments of the present invention will be described in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are merely for purposes of illustration and explanation, and are not intended to limit the present invention.
1. Background of model applications
The hotel search has various scenes, 11 types of million-level selectable screening items are contained in the search screening content, the search scenes and the search screening items can be freely combined without quantity limitation, and the traditional search mode of overlapped accurate matching of 'multi-scene and multi-condition screening' can cause few results or no results of hotel search, which is a great harm to user experience and causes the loss of famous names of long-tail traffic.
2. Model design summary
Aiming at the problem of no result or few result of hotel search, the overall structure of the Trans-dssm model is shown in figure 1. A user online subscribes for a hotel, searches for a target hotel, and adds complex and accurate search conditions to cause the problem that the hotel search has no result or few results. The purpose of the model design is to convert the exact search into semantic vector retrieval through the model. The semantic vector retrieval is the similarity retrieval of the search vector and the hotel vector, and the hotel and the search vector are only relatively similar, so when no result or few results are searched in the hotel, the recall recommendation of the search semantic vector is carried out, and the problem of no result or few results cannot be caused.
The Trans-dssm model outputs two vectors, one is a search vector based on search criteria and the other is a hotel vector based on hotel attribute information. When the model is applied, similarity retrieval is carried out on the search vector and all hotel vectors to obtain a relatively similar hotel result list, and the hotel result list is sorted from large to small according to the similarity value with the search vector.
The search vector (QueryEmbedding) is obtained by compressing the search condition by DNN. The search conditions comprise corresponding search scene characteristics and screening item characteristics, and comprise destinations, POIs, nearby coordinates, check-in time, prices, keywords, star levels, brands, house types and the like, and 11 search screening conditions are large, and the search conditions are more than 1 million. The search condition is converted into a feature vector through feature engineering, the feature vector is spliced together to form input features (ConcatQueryFeature) of the search side DNN, and the search side DNN compresses the input features of the search condition to output the search vector.
The hotel vector (HotelEmbedding) is obtained by compressing attribute information of the hotel by the DNN. The wine shop information comprises hotel attribute information such as coordinate position of the hotel, affiliated city business circle, price information, brand information, comment information, score, star level and the like. The hotel related information is converted into a feature vector through feature engineering, and is spliced together to form input features (Concat HotelFeautre) of the hotel side DNN, and the hotel side DNN compresses the input hotel features into the hotel vectors.
The search vector and the hotel vector are generated by two DNNs with unshared network weights, so that the search vector and the hotel vector do not belong to the same matrix space in the matrix space, and if the vectors in different spaces are directly used, similarity calculation is carried out, and similarity values among the vectors are compared, which is unreasonable in theory. The Trans-dssm study converts search vectors for different spaces and hotel vectors to the same space by defining a transformation matrix (TransferMatrix) and multiplying the search vectors and hotel vectors by the matrix at the same time. The actual meaning of multiplying the vector by the Matrix refers to the transformation of the vector in the Matrix space, so that after the search vector and the hotel vector are multiplied by the same transformation Matrix (Transfer Matrix), the search vector and the hotel vector are transformed into the same Matrix space. On the basis of a semantic retrieval model, a transformation matrix is added, so that the method is the core of improved optimization in a Trans-dssm model.
DNN design is the basis of the Trans-dssm model. The DNN uses three layers of neural networks, each layer of which has 512, 256, and 128 neurons. To prevent model overfitting, the DNN network used Dropout parameter settings with random discarding of neurons. The output value of the DNN neuron uses the activation function of Relu, which can prevent the gradient vanishing problem.
The loss function in Trans-dssm uses triple loss (TripletLoss). And calculating a function of similar distance in the triple loss, wherein cosine similarity is adopted, and the similarity between the search vector and the hotel vector is measured. The goal of triplet loss learning is to make the search vector a more similar to the positive sample hotel vector p and less similar to the negative sample vector n, i.e., a dissimilar hotel. The degree of model learning is that the similarity value to positive samples is greater than the similarity value to negative samples by some interval m (margin) value. m (margin) is an adjustable hyper-parameter in the training of the model, and the size of m plays a crucial role in the training of the model. In the model training, the continuous minimization of the loss function is to increase the similarity value of the positive sample and continuously reduce the negative sample, and when the positive sample value is larger than the negative sample, namely the negative sample reaches a certain value m, the loss is 0. The final model makes positive example hotels more and more similar than negative example hotels. The Trans-dssm loss function is as in equation 1.
3. Feature engineering
The characteristics input on the train-dssm model search side is different from that on the hotel side. The characteristics designed at the searching side are mainly based on the characteristics of the searching scene, and the scene characteristics are designed in a targeted manner. Because the proximity search is the largest, the design of the latitude and longitude coordinate features is the first consideration in the model input features. In the design of longitude and latitude coordinate characteristics, a longitude and latitude characteristic conversion formula as a formula 2 is designed. In the formula, a represents the degrees of longitude and latitude, [ ] represents taking an integer, and v is a converted characteristic value. And converting the longitude and latitude characteristics of the continuous numerical type based on the formula 2. For example, longitude 126.3234 is calculated 12632 based on equation 5.2, and latitude 45.5143 is calculated 4551 based on equation 5.2. In practice, on the warp line, the field distance is about 111 kilometers per 1 degree difference in latitude, and one percent latitude is in the range of 1.11 kilometers. On the weft, the actual distance is kilometers per 1 degree difference in longitude, while the actual error is no higher than 1.11 kilometers per one percent longitude. Therefore, based on the calculation method of formula 2, the maximum regional precision expressed by longitude and latitude is in the range of 1.11 kilometers, and the coding is simple, and the coding amount is limited domestically.
V＝[x×100] (2)
Keyword search is also one of the main search scenarios in hotel reservations. The Chinese keywords are complex, and word segmentation is generally performed when processing the Chinese keywords. Common dictionary-based segmentation methods may have OOV problems for long-ended words, i.e. exceeding the defined dictionary range. For example, the word of the electric round bed is not in the initial word stock, and the keyword cannot be extracted from the text, so that the search quality is influenced. If a large lexicon is initially set, this is less realistic and puts more stress on model training. For the search keyword, in order to reduce OOV problem of word segmentation, trans-dssm uses word segmentation, each word is an independent vector, and converts word sequence into vector representation through formula 3. In the formula, n is the number of the divided single words, E refers to the vector of the word, and E is the vector of the calculated whole keyword. Word segmentation has certain advantages: firstly, the word bank is not needed for the word segmentation of the single character, and the number of the segmented words is limited, so that the common Chinese character number and 26 letters are often only needed; secondly, the word segmentation of the single character has better generalization capability for the keywords with wrongly written characters. For example, the electric round bed is wrongly input into the electric round bed, because the original character is wrongly written and the other three characters are correct, the electric round bed is more similar to the retrieval content of the electric round bed in terms of the vector after the words are divided. In addition, in order to increase the representation of text semantic vectors, vectors of Chinese version single word BERT models issued by Google are used as initialization vectors of the text keywords of the Trans-dssm models. Except for the characteristics of the search keywords and the text characteristics of the hotel name, the vector characteristic representation of the hotel name is obtained by using the same characteristic processing method.
Other feature processing of the model input is relatively simple. Category features, such as brands, are encoded and converted into vectors. Numerical characteristics, such as price, are coded after bucket separation and then converted into vector quantity. No matter the characteristics of the search side or the characteristics of the hotel side, the non-category characteristics are converted into category characteristics, then the category characteristics are coded and converted into vectors, and the characteristic vectors are spliced to form a long characteristic vector which is input into a DNN network.
The training data for the model is derived from the user's search logs. The method comprises the steps of collecting 90-day search logs of a user based on a hotel online booking enterprise platform, analyzing search conditions of the user from the logs, and analyzing hotels exposed under the current search conditions and hotels clicked by the user in the search exposed hotels. And generating positive and negative samples required by model training based on the search exposure click log. The positive and negative samples are the training key of the model and are the targets to be optimized by the constructed model, and the scale of the training samples obtained by filtering and analyzing in a 90-day user search log is about 10 hundred million. The positive sample input by the model is a hotel clicked in the hotel exposed under the same search condition, and the negative sample is obtained by randomly sampling in the hotel not clicked in the search exposure. The positive and negative samples have a certain proportion, and the sample proportion is determined according to the actual model training effect and the off-line index evaluation and is an adjustable parameter.
4. Model training
The developed implementation of the Trans-dssm model is based on the open source machine learning framework, tensorflow, published by Google 2015. Training of the Trans-dssm model is mainly to adjust training parameters of the model so as to enable the model to achieve the optimal performance.
The Trans-dssm model uses a TripletLoss loss function, and an important hyper-parameter Margin is arranged in the function during model training, controls the similarity degree of a search vector and a hotel vector, and can adjust the size of the parameter according to actual service requirements. When the Margin is set to be too large, the loss of the model training is relatively large, the model training is difficult to converge at the moment, but the accuracy of the trained model is high; on the contrary, when the Margin setting is too small, the loss of the model training is small, and the model training is relatively easy to converge, but the accuracy of the model training is reduced. Through multiple parameter adjustment, in the training of the Trans-dssm model, the Margin is set to be 0.1, and the performance and convergence of the model are relatively good.
The main module DNN of Trans-dssm is a three-layer network structure, and the number of neurons in each layer is 512, 256 and 128 respectively. The number of layers of the DNN and the number of the neurons are adjustable parameters, when the number of layers and the number of the neurons are larger, the model has the risk of overfitting, and meanwhile the training speed of the model is slower. Reasonable training parameters need to be set for DNN according to experiments and index evaluation conditions. In the DNN, dropout regularization is used in the model, which is also a main parameter in the DNN, and mainly for preventing overfitting, 0.3 is set in the process of training the model, so that the performance is relatively good.
The Trans-dssm model has many other parameter settings when trained. For example, the maximum number of iterations for training the model is set to 400 ten thousand, each training sample is set to 1000, and the learning rate is 0.0001. All hyper-parameter settings are not constant and will vary depending on the scenario, data, features, etc. of the model application.
The Trans-dssm model adopts staged training in the training process. The model is trained every 10 ten thousand times, and check points (checkpoints) of the model are saved. The check points store all information such as model parameters, model weights, and model network structures. When the model is trained in multiple batches of samples in a phased and sequential manner, the model can be recovered from the check point, so that the training can be continued after the network parameters which are trained by the model last time.
5. Model evaluation
TABLE 1
As in the above table: 1. the comparison group A is a Google double-tower model, and the comparison group B is a semantic vector search model of Facebook
2. HR is shorthand for HitRNAte, HR1 is HitRNAte 1, meaning the proportion of the user's clicked item at the 1 st position in the list, HR3, HR10 and so on
3. 1-7 are evaluation indexes of continuous 7-day sampling
TABLE 2
As in the table above: hitRatet test statistics table as above table, in case of confidence of 0.05, because t >0, and p <0.05, i.e. we proposed the Trans-dssm model with significant improvement in HR1, HR3, HR10 compared to control A and control B models
6. Model application
Model application means that after training and evaluation are completed, a model is used for prediction. The prediction flow of the model is shown in fig. 2. After training of the Trans-dssm model is completed, two part vectors are output, wherein one part is a feature vector of a search condition and is used for generating a search vector; the other part is the hotel vector generated directly from the model. During model prediction, a search condition is converted into a characteristic vector through characteristic engineering, the characteristic vector is input into a search side DNN in trained Trans-dssm, and the DNN is compressed into a search vector. The hotel vectors can directly output all hotel vectors after model training is completed, then the ANN vector retrieval service loads the hotel vectors and generates vector retrieval indexes, and therefore rapid searching of large-scale data volume can be supported.
After the output of the model is finished, the model prediction firstly generates search vectors through DNN according to search conditions, the search vectors are based on index data of ANN service of hotel vectors, and n hotels with the nearest search vectors are retrieved, so that the whole process of model retrieval prediction is finished.
The technical scheme of the invention has the following beneficial effects:
1) The vector similarity recommendation model output by the Trans-dssm model is only relatively similar and has no strict search boundary, so that the condition of no search result does not exist;
2) Compared with the rule of model improvement and optimization and the double-tower model, the HR index is remarkably improved, and the accuracy and the recall rate are better;
3) The vector search model is more intelligent and not simply whether there is a match.
It should be noted that the method provided by the present invention can be applied not only to searching hotels, but also to searching other network-related goods.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the present invention in any way, and it will be apparent to those skilled in the art that the above description of the present invention can be applied to various modifications, equivalent variations or modifications without departing from the spirit and scope of the present invention.
Claims (10)
1. A vector searching method based on a Trans-dssm model is characterized by comprising the following steps:
inputting the search condition to a search side DNN of the Trans-dssm model, and compressing the search condition into a search vector through the search side DNN;
inputting hotel attribute information to a hotel side DNN of a Trans-dssm model, and compressing the information into hotel vectors through the hotel side DNN;
converting the search vector and the hotel vector into the same matrix space through a conversion matrix;
calculating cosine similarity between the converted search vector and the hotel vector;
sorting different hotel vectors by utilizing the cosine similarity obtained by calculation;
and taking the hotels corresponding to the first N hotel vectors in the sequence as search result recall recommendations.
2. The method for searching vectors based on the Trans-dssm model as claimed in claim 1, wherein the search condition is inputted to the search side DNN of the Trans-dssm model, and compressed into the search vector by the search side DNN, comprising:
converting the search condition into a feature vector through feature engineering;
and splicing the feature vectors together to form input features of the search side DNN, and compressing the input search condition features by the search side DNN to output search vectors.
3. The method of claim 3, wherein transforming the search criteria into feature vectors by feature engineering comprises:
and normalizing the longitude and latitude parameters, and calculating the range of the latitude characteristic value in 400,5400 and the range of the longitude characteristic value in 7300,12500.
4. The method of claim 3, wherein the transforming the search condition into the feature vector further comprises:
and carrying out word segmentation and normalization on the keyword parameters, using word segmentation, wherein each word is an independent vector, and converting a word sequence into a vector for representation.
5. The Trans-dssm model-based vector search method of claim 1, wherein the training data of the Trans-dssm model is a search log of a user.
6. The method for searching vectors based on the Trans-dssm model as claimed in claim 5, wherein the Margin parameters of the triple loss function are accurately parametered during the training process of the Trans-dssm model.
7. The method of claim 5 wherein model checkpoints are saved during training of the Trans-dssm model.
8. The Trans-dssm model-based vector search method of claim 7, wherein the checkpoint stores model parameters, model weights, and model network structure.
9. The Trans-dssm model-based vector search method of claim 1, wherein the search-side DNN and the hotel-side DNN both use a three-layer neuron structure, the DNN network uses Dropout parameter set, and the output value of the DNN neuron uses the activation function of Relu.
10. A Trans-dssm model-based vector search system, comprising:
one or more processors;
a storage device for storing one or more programs,
when executed by the one or more processors, cause the one or more processors to implement the Trans-dssm model based vector search method of any of claims 1 to 9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210709107.8A CN115203589A (en) | 2022-06-21 | 2022-06-21 | Vector searching method and system based on Trans-dssm model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210709107.8A CN115203589A (en) | 2022-06-21 | 2022-06-21 | Vector searching method and system based on Trans-dssm model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115203589A true CN115203589A (en) | 2022-10-18 |
Family
ID=83576926
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210709107.8A Pending CN115203589A (en) | 2022-06-21 | 2022-06-21 | Vector searching method and system based on Trans-dssm model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115203589A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117522533A (en) * | 2024-01-08 | 2024-02-06 | 江西求是高等研究院 | Hotel intelligent searching method and system |
-
2022
- 2022-06-21 CN CN202210709107.8A patent/CN115203589A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117522533A (en) * | 2024-01-08 | 2024-02-06 | 江西求是高等研究院 | Hotel intelligent searching method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111753060B (en) | Information retrieval method, apparatus, device and computer readable storage medium | |
| CN110298037B (en) | Convolutional neural network matching text recognition method based on enhanced attention mechanism | |
| CN110532554A (en) | A kind of Chinese abstraction generating method, system and storage medium | |
| CN111930887B (en) | Multi-document multi-answer machine reading and understanding system based on joint training mode | |
| CN106295796A (en) | Entity link method based on degree of depth study | |
| CN113505204B (en) | Recall model training method, search recall device and computer equipment | |
| CN110674407A (en) | Hybrid recommendation method based on graph convolution neural network | |
| CN110046250A (en) | Three embedded convolutional neural networks model and its more classification methods of text | |
| CN106844632A (en) | Based on the product review sensibility classification method and device that improve SVMs | |
| CN106776849A (en) | A kind of method and guide system to scheme quick-searching sight spot | |
| WO2023045187A1 (en) | Semantic search method and apparatus based on neural network, device, and storage medium | |
| CN112966091A (en) | Knowledge graph recommendation system fusing entity information and heat | |
| CN113326377A (en) | Name disambiguation method and system based on enterprise incidence relation | |
| CN113515632A (en) | Text classification method based on graph path knowledge extraction | |
| CN113672693A (en) | Label recommendation method of online question and answer platform based on knowledge graph and label association | |
| CN114332519A (en) | Image description generation method based on external triple and abstract relation | |
| CN116662565A (en) | Heterogeneous information network keyword generation method based on contrast learning pre-training | |
| CN115203589A (en) | Vector searching method and system based on Trans-dssm model | |
| CN105183845A (en) | ERVQ image indexing and retrieval method in combination with semantic features | |
| CN104572915A (en) | User event relevance calculation method based on content environment enhancement | |
| CN113934835A (en) | Retrieval type reply dialogue method and system combining keywords and semantic understanding representation | |
| CN113159187A (en) | Classification model training method and device, and target text determining method and device | |
| CN116932736A (en) | Patent recommendation method based on combination of user requirements and inverted list | |
| CN116842934A (en) | Multi-document fusion deep learning title generation method based on continuous learning | |
| CN116957128A (en) | Service index prediction method, device, equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |