Introduction: Knowledge Graph shines in various industries and fields by its ability to describe the relationship between knowledge and everything in the world using graph models. At the same time, knowledge graph technology has also brought new challenges and opportunities to scene search. Against this background, the Meituan team has carried out a new round of technological innovations and applied knowledge graph technology to the cognition of live and tourism scenarios. The title of this sharing is "Application of Knowledge Graph in the Cognition of the Tourism Scene of Meituan Search", which mainly introduces:
- Hotel and Tourism Business Features
- Hotel and Tourism Scene Cognition
- Hotel and Tourism Knowledge Graph
- Hotel and Tourism Knowledge Graph
- search solution based on scene cognition
1
Hotel and tourism business features
report Meituan Dianping , what do you think of? Takeout, food, movies?
In fact, Meituan has more than these, and there are also hotel and tourism businesses that also occupy the top industry! What is the difference between
Hotel and tourism business and other local life services ? The difference is that the travel radius of the hotel and tourism business is larger. Most Meituan’s services, such as script killing, movies, etc., only need to find and meet the needs of users’ merchants in the current area or in the current city. For hotel and tourism business, users have a larger travel radius. Three examples are used to explain the characteristics of the hotel and tourism business better:
A Beijing user wants to ski or watch the sea, and users may choose the higher-quality Chongli Ski Resort more than 200 kilometers away from Beijing; for watching the sea, because there is no corresponding supply in Beijing, users may go to Beidaihe in Qinhuangdao . Therefore, for services that do not provide corresponding services locally, users may go to cities with corresponding services around them or even look for cities with corresponding services across the country. At this time, Meituan needs to find and filter out merchants with suitable scenarios and adapted scenarios for users.
Taking hotels as an example, introduces several commonly used search methods for users:
① Landmark/Business District. Users look for hotels around landmarks or business districts, and this model is more inclined to recommend issues;
② Merchant/Brand. This type of user has a clear goal, looking for a specific hotel or hotel under the corresponding brand;
③Pan scene search. For example, users are looking for a certain type of business, such as a youth hostel, a hotel with certain facilities, or a specific room type such as an e-sports room, etc. Other search methods mainly revolve around a combination of these three searches.
Search for attractions can also be divided into three main methods:
① Administrative region/region. If a user searches for a certain area or administrative region, Meituan needs to find the relevant attractions in the area;
② Merchant/Brand. This type of user also has a clear purpose and will directly search for the names of the corresponding attractions such as Forbidden City , Happy Valley , etc.
③Pan scene search. When users look for a certain type of demand, such as skiing and mountain climbing, at this time, multiple attractions with the same attributes, service items or the same type can meet the user's needs. In this scenario, Meituan needs to make better recommendations for users.
--
2
There are a large number of merchants of different business types in the platform, such as hotels, attractions, food, etc.When a user initiates a search on the platform, we believe that his search behavior is an explicit expression of the demand scenario, that is, he can identify the user's demand scenario through a scenario-based understanding of the user's search behavior. The user demand scenario is expressed through tags and then the merchant’s tag search can be matched with the merchant that can carry their needs.
Why do we need to do scene cognition - I hope to do different experience optimization based on different scenario needs.
Take the search for "Zhongshan Park" as an example: the original search plan will recall text-related results from the perspective of text, but this cannot meet the needs of users. Through the analysis of user behavior data, we found that most users who initiated the "Zhongshan Park" search behavior in Beijing often initiate follow-up actions around the scenes around Zhongshan Park in Beijing.
This essentially represents the scenario needs of a type of users. In order to better meet this scenario needs, we adopted the recommended method shown on the right side of the figure. First, discover the main point that the user is looking for, and then recommend other services to the user around the main point, such as related attractions, nearby food, etc., to better meet the user's diverse needs, and through this main point template style, the user can provide more information that can help make decisions.
This change in demand provides us with new solutions to the problem:
- 1 through product style upgrade to meet the diversified demands of users;
- With the transformation of product styles, it forces a new round of technological upgrades.
For the hotel and tourism business, the scenarios can be roughly divided into two categories: accurate merchant search and pan-scene search.
(1) Accurate merchant search
Accurate merchant search can be specifically divided into:
- Single main point search . For example, Forbidden City and Beijing Kuntai Hotel have only the only corresponding main point, and you can recommend services and merchants to users around this main point.
- Multi-main point search , such as Fantastic. There are many Fantastic Happy Worlds nationwide, and at this time, you can use the multi-main point style to provide information to users.
(2) Pan-scene search
Pan-scene search can be specifically divided into:
- Landmark + X, such as "Youth Hostel near the Bird's Nest", at this time, you need to recommend the main point of the Bird's Nest to the user's green point near the Bird's Nest. 2019 hotel;
- local + X. If you search for "climbing" under the city page in Beijing, you need to recommend merchants who can meet user needs and scenarios in the local and surrounding areas;
- National + X. If you search for three mountains and five mountains, you need to help users recall results that meet user scenario needs in this scenario.
The most basic problem is that the understanding of the demand scenario needs to rely on a large amount of domain knowledge in the underlying layer . The basic structured data does not carry scene-like expressions, so additional mining and understanding is required, for example:
(1) Search for Heshen Mansion. Search for Heshen Mansion, the POI of Gongwang Mansion should appear. At this time, you need to find out that Heshen Mansion is an alias of Gongwang Mansion;
(2) Search for a quiet hotel.This type of search is a pan-scene search. First, you need to explore the hotel knowledge on the B-end to explore whether the hotel has good sound insulation characteristics; second, you need to accurately correlate the online query with the existing B-end knowledge, such as the underlying knowledge label mapped to the "quiet" expression is "good sound insulation".
--
3
(1) Overview
Meituan Travel knowledge graph covers about 800,000 hotel and travel merchants, 100 million user reviews, accumulating 1 million atomic concepts, and the combined demand concepts of 1.5 million, and the relationship between 12 million merchants and concepts.
(2) Four-layer definition
In the process of graph application, the graph is divided into four layers according to the application dimension of user needs:
- Category system layer : combine different business characteristics to divide the category purpose. Taking tourism as an example, 15 first-level categories are defined, and on this basis, the specific category systems of the second-level and third-level are split;
- atomic concept layer : mining and extracting the atomic concept layer from user comments and merchant information;
- Requirements concept layer : filtering the data of the atomic concept layer and combining it in line with semantic dimensions to build concept layer data that is oriented towards search requirements. For example, mountain climbing in the atomic concept layer has a direct mapping label in the demand concept layer. For the two independent semantics of parent-child and hot springs, when there are real user demand scenarios, a more accurate dimension of "parent-child hot springs" label will be combined. This solves the semantic drift problem that occurs when searching online query, if you only use the finest-grained atomic label for recall. Therefore, when costs allow, it will be better to adopt a combination-based demand concept layer.
- POI layer : determine whether the current merchant has the attributes and characteristics represented by the tags or knowledge of the requirements concept layer.
(3) Mining process
- Knowledge extraction
data sources mainly include three categories: structured data, semi-structured data, and unstructured data. Knowledge extraction is performed from these data sources. In the early stage, if there is not enough labeled data, a semi-supervised method is used to perform Auto Phrase, dependent syntax analysis, etc.; after knowledge accumulation, the NER method is used for better extraction.
- Knowledge Classification
Knowledge fragments are classified according to the one, two, and three hierarchical knowledge system mentioned above. As shown in the figure below, the knowledge base is related to the categories related to the Taishan Scenic Area and the knowledge fragments associated at the lowest level. For the Taishan Scenic Area, the lowest data is only natural scenery, but we hope to understand its semantics to a more fine-grained size, such as canyons, waterfalls, mountains, etc. In this case, we need to pay attention to what types of natural scenery there are. Similarly, what activities are suitable for Mount Tai Scenic Area? From the perspective of the crowd, such as suitable for couples to date and graduation trip; the perspective of scene events, suitable for outdoor climbing, sunrise and sunset. These data can lay the foundation for the generation of subsequent query understanding, linking, and recommendation reasons.
(1) Definition of knowledge system
When building a knowledge graph in a vertical field, it is necessary to combine the domain knowledge to define the business. Taking scenic spots as an example, there are about 15 first-level categories, and there are also second- and third-level categories under each first-level category. The third-level classification system can better meet the search-oriented needs.
(2) Knowledge extraction
In the early stage of the business, a semi-supervised learning method, namely Bootstrapped pattern-based learning, accumulate knowledge. Taking mining animal-related entities as an example, the general process is:
- Step1: Construct the seed entity word, in this example, dog;
- Step2: According to the entity word dog, the aforementioned pattern of dog is mined from the corpus: own a, my pet;
- Step3: Use the candidate pattern to further explore the entity knowledge fragments that match the above pattern: house, cat;
- Step4: Evaluate which one house or cat is the same as dog and belongs to the animal type.
After several rounds of iterations, it was found that for animal-related entity words, my pet is a better pattern, so I used this pattern to mine and expand more entities. In the early stage, data can be quickly accumulated.
After the labeled data is available, NER's supervised learning method can be used for better generalization and recognition. The BERT+CRF model was initially used for extraction, but this model easily chopped up semantic fragments of knowledge entities. As shown in the figure below, the comment is "The Four Wonders of Emei in the Golden Pin", and the "Four Wonders of Emei" was drawn, but we hope that the clip can be as complete as possible. Later, the problem was solved by introducing KG-related information. First, the text that needs to be extracted is divided into words, and two layers of vector information of Character-level granularity and Word-level granularity are introduced to assist in determining the boundary of segmentation of segments, which can effectively solve the problem of fragments being chopped.
At first, for efficiency, each of the 15 first-level categories mentioned above was extracted separately. In order to improve the accuracy, the relevant major categories were put together for comprehensive knowledge extraction. In this process, there will be problems of inaccurate knowledge classification. Further optimization is to use a multi-task joint training method, that is, to integrate the knowledge classification tasks and NER tasks.
The general idea is : through the vector information after the first layer encoding (using BERT or BiLSTM encoding), then NER segmentation is used with traditional CRF, and then the vector information is introduced into the classification layer for identification and processing using the idea of casecade cascade. This work has improved the overall accuracy and recall rate relatively well (published in the 2022 DASFAA meeting).
(3) Knowledge Classification
After the initial classification of knowledge for the first-level category, the knowledge needs to be classified in a more fine-grained manner. However, due to business characteristics, many knowledge fragments will belong to multiple nodes in the second-level or third-level categories, so the multi-label classification task is used here to classify the original text fragments. For example, when VR projects "fly over the horizon", when using BERT to encode and directly identify the fragment "fly over the horizon", it is easy to classify it as a roller coaster project, because the expression word "fly over XXX" in more cases does not refer to VR projects but roller coasters or other projects.
In order to better solve this problem, after extracting the fragment, the context information of the fragment is introduced into the model to enrich the context expression. At the same time, feature engineering is carried out in the search log and comment log, manually constructed features are added, and these features are fused and classified uniformly, so that the problem of semantic offsets can be solved in separate texts.
(4) Knowledge is related to merchants
After extracting the knowledge fragment and classifying it, it is necessary to solve the problem of association of knowledge (tags) with platform merchants.
First of all, it should be clear that the above problem is not a closed domain label problem but an open domain label problem, that is, it is not a simply a classification and mount problem of the category system, because tags will continue to accumulate and add with the development of the business and mining. We need to classify these new tags and knowledge, so marking requires a certain generalization ability.
in combination with the platform: the main body of the platform is a merchant, and we need to find merchants with knowledge or tags. Merchants have many user reviews and products listed below. What we can intuitively obtain is the correlation between each evaluation and knowledge and the correlation between each product and knowledge. It feedbacks the correlation between merchants, and there is an additional layer of aggregation process in the middle. The solution to the problem of
The multi-person voting mechanism , that is, every piece of information hanging under the merchant is a feedback from a user. To determine whether it is related or not, or other opinions, the aggregation and voting of this information can make the merchant have this knowledge or tag.
As an example, determine whether a certain attraction can bring pets:
- Step1: Find evidence and find text expressions related to bringing pets;
- Step2: determine the authenticity of the extracted short fragments, which are mainly divided into positive correlation, irrelevant, and negative correlation;
- Step3: Multiple evidence fusion classification. In addition to the evidence correlation obtained earlier, many dimensions of feature information are abstracted based on semantic features and statistical features, such as the text correlation of the POI itself information. This process mainly uses the BERT model to match text correlation.
- Step4: Distribution confidence determination, feed the correlation obtained in the first two steps into the tree model, and finally obtain the classification result, that is, whether it is related.
If knowledge classification is performed online, there are requirements for accuracy, allowing certain recall losses, but it is necessary to ensure that the results are accurate, so that users can experience better online. Therefore, in the last step, the process of determining the distribution confidence is increased. That is, the distribution statistics of the merchant categories in the marking results are filtered to filter the POI of the long-tail category. For example, when a user searches for a mountain climb, after the category statistics, if individual judgments are wrong, the hot spring category accounts for 0.3%, and this type of result is filtered according to the threshold.
(5) MT-BERTh
Whether it is knowledge extraction or knowledge classification, the BERT model is used. Meituan mainly uses the self-developed MT-BERT, which is characterized by introducing a large amount of user comment information in Meituan's business scenario and business information hanging under the merchant to better adapt to the model.
MT-BERT After adding Meituan UGC data, it has made significant improvements in some public data, internal query intention classification and component analysis tasks.
--
4
divides service search into five layers:
- L0 layer : mine relevant knowledge and build index ;
- L1 layer : identify and understand user query and perform structured recall;
- L1 layer : identify and understand user query and perform structured recall;
- L2 layer : Sorting the list in deep learning model based on the recall results;
- L3 layer : Making targeted strategy adjustments in different business scenarios;
- L4 layer : When presenting the list, provide tags, recommendation reasons, lists and other interpretable information to strengthen perception.
Introduction: Knowledge Graph shines in various industries and fields by its ability to describe the relationship between knowledge and everything in the world using graph models. At the same time, knowledge graph technology has also brought new challenges and opportunities to scene search. Against this background, the Meituan team has carried out a new round of technological innovations and applied knowledge graph technology to the cognition of live and tourism scenarios. The title of this sharing is "Application of Knowledge Graph in the Cognition of the Tourism Scene of Meituan Search", which mainly introduces:
- Hotel and Tourism Business Features
- Hotel and Tourism Scene Cognition
- Hotel and Tourism Knowledge Graph
- Hotel and Tourism Knowledge Graph
- search solution based on scene cognition
1
Hotel and tourism business features
report Meituan Dianping , what do you think of? Takeout, food, movies?
In fact, Meituan has more than these, and there are also hotel and tourism businesses that also occupy the top industry! What is the difference between
Hotel and tourism business and other local life services ? The difference is that the travel radius of the hotel and tourism business is larger. Most Meituan’s services, such as script killing, movies, etc., only need to find and meet the needs of users’ merchants in the current area or in the current city. For hotel and tourism business, users have a larger travel radius. Three examples are used to explain the characteristics of the hotel and tourism business better:
A Beijing user wants to ski or watch the sea, and users may choose the higher-quality Chongli Ski Resort more than 200 kilometers away from Beijing; for watching the sea, because there is no corresponding supply in Beijing, users may go to Beidaihe in Qinhuangdao . Therefore, for services that do not provide corresponding services locally, users may go to cities with corresponding services around them or even look for cities with corresponding services across the country. At this time, Meituan needs to find and filter out merchants with suitable scenarios and adapted scenarios for users.
Taking hotels as an example, introduces several commonly used search methods for users:
① Landmark/Business District. Users look for hotels around landmarks or business districts, and this model is more inclined to recommend issues;
② Merchant/Brand. This type of user has a clear goal, looking for a specific hotel or hotel under the corresponding brand;
③Pan scene search. For example, users are looking for a certain type of business, such as a youth hostel, a hotel with certain facilities, or a specific room type such as an e-sports room, etc. Other search methods mainly revolve around a combination of these three searches.
Search for attractions can also be divided into three main methods:
① Administrative region/region. If a user searches for a certain area or administrative region, Meituan needs to find the relevant attractions in the area;
② Merchant/Brand. This type of user also has a clear purpose and will directly search for the names of the corresponding attractions such as Forbidden City , Happy Valley , etc.
③Pan scene search. When users look for a certain type of demand, such as skiing and mountain climbing, at this time, multiple attractions with the same attributes, service items or the same type can meet the user's needs. In this scenario, Meituan needs to make better recommendations for users.
--
2
There are a large number of merchants of different business types in the platform, such as hotels, attractions, food, etc.When a user initiates a search on the platform, we believe that his search behavior is an explicit expression of the demand scenario, that is, he can identify the user's demand scenario through a scenario-based understanding of the user's search behavior. The user demand scenario is expressed through tags and then the merchant’s tag search can be matched with the merchant that can carry their needs.
Why do we need to do scene cognition - I hope to do different experience optimization based on different scenario needs.
Take the search for "Zhongshan Park" as an example: the original search plan will recall text-related results from the perspective of text, but this cannot meet the needs of users. Through the analysis of user behavior data, we found that most users who initiated the "Zhongshan Park" search behavior in Beijing often initiate follow-up actions around the scenes around Zhongshan Park in Beijing.
This essentially represents the scenario needs of a type of users. In order to better meet this scenario needs, we adopted the recommended method shown on the right side of the figure. First, discover the main point that the user is looking for, and then recommend other services to the user around the main point, such as related attractions, nearby food, etc., to better meet the user's diverse needs, and through this main point template style, the user can provide more information that can help make decisions.
This change in demand provides us with new solutions to the problem:
- 1 through product style upgrade to meet the diversified demands of users;
- With the transformation of product styles, it forces a new round of technological upgrades.
For the hotel and tourism business, the scenarios can be roughly divided into two categories: accurate merchant search and pan-scene search.
(1) Accurate merchant search
Accurate merchant search can be specifically divided into:
- Single main point search . For example, Forbidden City and Beijing Kuntai Hotel have only the only corresponding main point, and you can recommend services and merchants to users around this main point.
- Multi-main point search , such as Fantastic. There are many Fantastic Happy Worlds nationwide, and at this time, you can use the multi-main point style to provide information to users.
(2) Pan-scene search
Pan-scene search can be specifically divided into:
- Landmark + X, such as "Youth Hostel near the Bird's Nest", at this time, you need to recommend the main point of the Bird's Nest to the user's green point near the Bird's Nest. 2019 hotel;
- local + X. If you search for "climbing" under the city page in Beijing, you need to recommend merchants who can meet user needs and scenarios in the local and surrounding areas;
- National + X. If you search for three mountains and five mountains, you need to help users recall results that meet user scenario needs in this scenario.
The most basic problem is that the understanding of the demand scenario needs to rely on a large amount of domain knowledge in the underlying layer . The basic structured data does not carry scene-like expressions, so additional mining and understanding is required, for example:
(1) Search for Heshen Mansion. Search for Heshen Mansion, the POI of Gongwang Mansion should appear. At this time, you need to find out that Heshen Mansion is an alias of Gongwang Mansion;
(2) Search for a quiet hotel.This type of search is a pan-scene search. First, you need to explore the hotel knowledge on the B-end to explore whether the hotel has good sound insulation characteristics; second, you need to accurately correlate the online query with the existing B-end knowledge, such as the underlying knowledge label mapped to the "quiet" expression is "good sound insulation".
--
3
(1) Overview
Meituan Travel knowledge graph covers about 800,000 hotel and travel merchants, 100 million user reviews, accumulating 1 million atomic concepts, and the combined demand concepts of 1.5 million, and the relationship between 12 million merchants and concepts.
(2) Four-layer definition
In the process of graph application, the graph is divided into four layers according to the application dimension of user needs:
- Category system layer : combine different business characteristics to divide the category purpose. Taking tourism as an example, 15 first-level categories are defined, and on this basis, the specific category systems of the second-level and third-level are split;
- atomic concept layer : mining and extracting the atomic concept layer from user comments and merchant information;
- Requirements concept layer : filtering the data of the atomic concept layer and combining it in line with semantic dimensions to build concept layer data that is oriented towards search requirements. For example, mountain climbing in the atomic concept layer has a direct mapping label in the demand concept layer. For the two independent semantics of parent-child and hot springs, when there are real user demand scenarios, a more accurate dimension of "parent-child hot springs" label will be combined. This solves the semantic drift problem that occurs when searching online query, if you only use the finest-grained atomic label for recall. Therefore, when costs allow, it will be better to adopt a combination-based demand concept layer.
- POI layer : determine whether the current merchant has the attributes and characteristics represented by the tags or knowledge of the requirements concept layer.
(3) Mining process
- Knowledge extraction
data sources mainly include three categories: structured data, semi-structured data, and unstructured data. Knowledge extraction is performed from these data sources. In the early stage, if there is not enough labeled data, a semi-supervised method is used to perform Auto Phrase, dependent syntax analysis, etc.; after knowledge accumulation, the NER method is used for better extraction.
- Knowledge Classification
Knowledge fragments are classified according to the one, two, and three hierarchical knowledge system mentioned above. As shown in the figure below, the knowledge base is related to the categories related to the Taishan Scenic Area and the knowledge fragments associated at the lowest level. For the Taishan Scenic Area, the lowest data is only natural scenery, but we hope to understand its semantics to a more fine-grained size, such as canyons, waterfalls, mountains, etc. In this case, we need to pay attention to what types of natural scenery there are. Similarly, what activities are suitable for Mount Tai Scenic Area? From the perspective of the crowd, such as suitable for couples to date and graduation trip; the perspective of scene events, suitable for outdoor climbing, sunrise and sunset. These data can lay the foundation for the generation of subsequent query understanding, linking, and recommendation reasons.
(1) Definition of knowledge system
When building a knowledge graph in a vertical field, it is necessary to combine the domain knowledge to define the business. Taking scenic spots as an example, there are about 15 first-level categories, and there are also second- and third-level categories under each first-level category. The third-level classification system can better meet the search-oriented needs.
(2) Knowledge extraction
In the early stage of the business, a semi-supervised learning method, namely Bootstrapped pattern-based learning, accumulate knowledge. Taking mining animal-related entities as an example, the general process is:
- Step1: Construct the seed entity word, in this example, dog;
- Step2: According to the entity word dog, the aforementioned pattern of dog is mined from the corpus: own a, my pet;
- Step3: Use the candidate pattern to further explore the entity knowledge fragments that match the above pattern: house, cat;
- Step4: Evaluate which one house or cat is the same as dog and belongs to the animal type.
After several rounds of iterations, it was found that for animal-related entity words, my pet is a better pattern, so I used this pattern to mine and expand more entities. In the early stage, data can be quickly accumulated.
After the labeled data is available, NER's supervised learning method can be used for better generalization and recognition. The BERT+CRF model was initially used for extraction, but this model easily chopped up semantic fragments of knowledge entities. As shown in the figure below, the comment is "The Four Wonders of Emei in the Golden Pin", and the "Four Wonders of Emei" was drawn, but we hope that the clip can be as complete as possible. Later, the problem was solved by introducing KG-related information. First, the text that needs to be extracted is divided into words, and two layers of vector information of Character-level granularity and Word-level granularity are introduced to assist in determining the boundary of segmentation of segments, which can effectively solve the problem of fragments being chopped.
At first, for efficiency, each of the 15 first-level categories mentioned above was extracted separately. In order to improve the accuracy, the relevant major categories were put together for comprehensive knowledge extraction. In this process, there will be problems of inaccurate knowledge classification. Further optimization is to use a multi-task joint training method, that is, to integrate the knowledge classification tasks and NER tasks.
The general idea is : through the vector information after the first layer encoding (using BERT or BiLSTM encoding), then NER segmentation is used with traditional CRF, and then the vector information is introduced into the classification layer for identification and processing using the idea of casecade cascade. This work has improved the overall accuracy and recall rate relatively well (published in the 2022 DASFAA meeting).
(3) Knowledge Classification
After the initial classification of knowledge for the first-level category, the knowledge needs to be classified in a more fine-grained manner. However, due to business characteristics, many knowledge fragments will belong to multiple nodes in the second-level or third-level categories, so the multi-label classification task is used here to classify the original text fragments. For example, when VR projects "fly over the horizon", when using BERT to encode and directly identify the fragment "fly over the horizon", it is easy to classify it as a roller coaster project, because the expression word "fly over XXX" in more cases does not refer to VR projects but roller coasters or other projects.
In order to better solve this problem, after extracting the fragment, the context information of the fragment is introduced into the model to enrich the context expression. At the same time, feature engineering is carried out in the search log and comment log, manually constructed features are added, and these features are fused and classified uniformly, so that the problem of semantic offsets can be solved in separate texts.
(4) Knowledge is related to merchants
After extracting the knowledge fragment and classifying it, it is necessary to solve the problem of association of knowledge (tags) with platform merchants.
First of all, it should be clear that the above problem is not a closed domain label problem but an open domain label problem, that is, it is not a simply a classification and mount problem of the category system, because tags will continue to accumulate and add with the development of the business and mining. We need to classify these new tags and knowledge, so marking requires a certain generalization ability.
in combination with the platform: the main body of the platform is a merchant, and we need to find merchants with knowledge or tags. Merchants have many user reviews and products listed below. What we can intuitively obtain is the correlation between each evaluation and knowledge and the correlation between each product and knowledge. It feedbacks the correlation between merchants, and there is an additional layer of aggregation process in the middle. The solution to the problem of
The multi-person voting mechanism , that is, every piece of information hanging under the merchant is a feedback from a user. To determine whether it is related or not, or other opinions, the aggregation and voting of this information can make the merchant have this knowledge or tag.
As an example, determine whether a certain attraction can bring pets:
- Step1: Find evidence and find text expressions related to bringing pets;
- Step2: determine the authenticity of the extracted short fragments, which are mainly divided into positive correlation, irrelevant, and negative correlation;
- Step3: Multiple evidence fusion classification. In addition to the evidence correlation obtained earlier, many dimensions of feature information are abstracted based on semantic features and statistical features, such as the text correlation of the POI itself information. This process mainly uses the BERT model to match text correlation.
- Step4: Distribution confidence determination, feed the correlation obtained in the first two steps into the tree model, and finally obtain the classification result, that is, whether it is related.
If knowledge classification is performed online, there are requirements for accuracy, allowing certain recall losses, but it is necessary to ensure that the results are accurate, so that users can experience better online. Therefore, in the last step, the process of determining the distribution confidence is increased. That is, the distribution statistics of the merchant categories in the marking results are filtered to filter the POI of the long-tail category. For example, when a user searches for a mountain climb, after the category statistics, if individual judgments are wrong, the hot spring category accounts for 0.3%, and this type of result is filtered according to the threshold.
(5) MT-BERTh
Whether it is knowledge extraction or knowledge classification, the BERT model is used. Meituan mainly uses the self-developed MT-BERT, which is characterized by introducing a large amount of user comment information in Meituan's business scenario and business information hanging under the merchant to better adapt to the model.
MT-BERT After adding Meituan UGC data, it has made significant improvements in some public data, internal query intention classification and component analysis tasks.
--
4
divides service search into five layers:
- L0 layer : mine relevant knowledge and build index ;
- L1 layer : identify and understand user query and perform structured recall;
- L1 layer : identify and understand user query and perform structured recall;
- L2 layer : Sorting the list in deep learning model based on the recall results;
- L3 layer : Making targeted strategy adjustments in different business scenarios;
- L4 layer : When presenting the list, provide tags, recommendation reasons, lists and other interpretable information to strengthen perception.
(1) Accurate merchant search
model the problem: first, identify the main point; secondly, recommend related merchants around the main point, including attractions and nearby hotels, etc. The second step is to subdivide in detail: When the user is in the planning decision-making process, the user can recommend merchants that can replace the main point; when the user has determined to consume at a certain main point, the user can recommend merchants that can match itineraries
There are two technical points mentioned above:
- How to identify the main point, it is called the merchant link or main point link, that is, a variant of the entity link. How to better recommend
- How to revolve around the main point.
①Technical point 1: How to identify the main point
adopts the entity linking scheme based on context information . The reason for using this solution is: in Meituan's business, the required results may be different in different geographical locations. For example, when searching for Dragon Dream Hotel near Dragon Dream Hotel, No. 492 Anhua Road, Shanghai, the user is most likely looking for Dragon Dream Hotel near him.
The specific strategy is mainly divided into two fractions : the first score is a word sequence, that is, the probability prediction score between the sequence fragment that can be linked to a certain entity and the probability prediction score between the entity, and the alias expression of the merchant by search logs and knowledge graphs is obtained; the second score is the information score based on the context, and the score is divided into two parts based on the business characteristics. First, the semantic score of the context of the text semantics itself; second, the geographical context score, that is, the score is calculated based on the distance between the user and the merchant.
②Technical point 2: How to better recall
Two methods are used for recall:
- Based on the vector correlation of the user behavior sequence
Constructs all POIs clicked by the user to form a sequence, and encodes it based on skip-gram to obtain a POI vector, and then calculates it with the POI vector of the main point to obtain a head POI similar to it. It is worth noting that we have made targeted adjustments based on the business characteristics: First of all, the hotel and tourism business is a relatively low-frequency business, so we will aggregate users' behavior sequences over a longer period of time and focus on solving the low-frequency problem. Secondly, with the differences in the behavioral characteristics of different cities, we have introduced side-information such as cities, time, categories, prices, etc. to better calculate the correlation of vectors.
- Based on GCN vector correlation
Each POI has relevant knowledge. We constructed a heterogeneous graph of User-Query-POI-Item (knowledge) and obtained the vector of POI through the graph learning method.
(2) Pan-scene search
Model the problem: First, identify which scene needs are there; second, search merchants based on the scene and sort them in a personalized manner; finally, explain why the results can meet the scene needs.
gives an example. If you search for "a park suitable for walking your children" in Beijing, the area you are looking for is Beijing. Different text fragments in the query will be linked to different tags, and there is also a concept of master-slave between tags. Therefore, you can infer the superior scene, homoslovakia, triggered lists, and tags that need to be displayed, and finally form a recall syntax for subsequent processing.
pan-scene search involves three main technical points:
①Technical point 1: Pan-scene link
0 1 Recognize query, that is, link the scene tag.This process is mainly carried out online and is divided into six steps:
- Step1: Trigger judgment to determine whether the current query is the type of pan-scene search. For example, the Palace Museum is an accurate search rather than a pan-scene search, and climbing a mountain is a pan-scene search. At the same time, this step also requires identifying relevant intentions, such as judging whether the search intention is a scenic spot, hotel or catering, etc.;
- Step2: Preprocessing the query, including word segmentation, Non-Link and identification of the target area;
- Step3: Generate candidate sequences based on the processed fragments, perform multiple combinations, and also use the technology of skipping links. For example, the existing three fragments A, B, and C may generate the following sequences of ABC, AB_C, A_B_C, and AC;
- Step4: Use the combined sequence to perform inverted index recall, including whitelist matching, pattern matching, vector recall, etc. to expand related tags;
- Step5: Tag sorting, sorting the above recall results. This process has several important features, including the correlation between the current entity and sequence, the correlation between query and entity, mention information, and the statistical characteristics after clicking aggregation and attribution. These are combined and classified, and the topN tags are selected in different businesses for application;
- Step6: The final judgment process, disambiguating and reasoning the different tags linked by multiple fragments identified.
②Technical point 2: Sort
This technical point considers how to better perform personalized sorting based on scene search. The expression of the pan-search class lacks semantic correlation in the POI name dimension, so information from the knowledge dimension needs to be supplemented in the model.
First of all, at the feature level, to enrich merchants' KG-based semantic expression vectors, the following methods are mainly adopted:
- is based on a multi-domain structure and introduces the text information of the tag;
- uses the GCN structure to train the vector between POI and query, and introduce the vector into the subsequent model.
Secondly, in terms of model structure, sequence modeling of user interest scenarios and merchant scenarios. Innovations in this work:
- currently POIs and the attention of the user's long-time sequence and shortt-time sequence. Among them, long-time sequence and short-time sequence refer to the sequence encoding generated by the behavior list of POIs that the user has clicked over a longer period of time and the short-time list of behaviors;
- introduces the sequence of tags. The identified tags of the demand scenarios that the current user is looking for are aggregated with the merchant's tags that the user has had previous behavior and become sequences; the knowledge information of the tags mounted on the POI itself is also a sequence. Encoding these two sequences and doing attention work can better capture the correlation between the user's demand scenario tags and the merchant's scenario sequence and user interest sequence under pan-scene search.
③Technical point 3: Result interpretability
This technical point considers how to better explain to users why the current results and search scenarios are related. Part of this problem is achieved through recommendation reasons. There are two ways to implement recommendation reasons: extracted recommendation reasons and generative recommendation reasons.
- Extraction recommendation reason
This method mines relevant recommendation reasons through extraction. It is mainly divided into two major scenarios during extraction:
first category, such as searching for "climbing". This type of query belongs to a specific scenario. In this specific scenario, the recommendation words we hope to provide to users are directly related to the scene, that is, the expression of the place suitable for climbing when other users come to the scenic spot.For this kind of recommendation reason, text matching is used and BERT-MRC is used to recall candidate sentences.
second category, such as searching for "Shanghai Tourism". The scope of this type of query is relatively broad, so in this scenario, the default feature will be recommended to users, namely the characteristics description of the scenic spot itself. For this type of recommendation words that can directly represent the characteristics of the merchant, a combination of short sentences and the idea of pointer-generator network extraction is used to generate candidate sentences.
When there is a candidate recommendation, whether it is the user's recommendation or the merchant's own recommendation, it will be uniformly entered into the judgment module to make a series of quality judgments on the candidate sentences, including rewriting of the sentence, expressing smoothness, expressing emotions, etc. Through these modules, multi-dimensional scores are obtained, and finally feeding them into the tree model as features for overall quality judgment, and obtaining the final judgment score.
- generated recommendation reason
extracted recommendation reason can solve most problems, but there are some problems:
sentence expression is correct but long, and there are sentence length restrictions when the front end shows it to users. Therefore, in order to better use longer sentences or sentences with no problem but express relatively awkward and bright spots, some sentences are compressed by a transformer-based network structure to make them meet the limitations of sentence length.
In addition, many POIs have fewer comments themselves, and the user's expression quality is poor, and we have strict overall control over the recommendations, so we cannot dig out recommended words related to the scene.
At this time, we consider generating recommendations based on the merchant's existing knowledge and expanding the data. This idea is implemented based on the KG and scenario keyword generation scheme. There are two key points, giving examples: First, "Sima Tai Great Wall" has tags related to POI in the Word embedding layer; second, control tags to control relevant recommendations based on the "Spring Outing" tag. Through the two dimensions, the generated fragment is finally obtained. After encoder, I got "the spring flowers are romantic and the summer is full of green". Based on this sentence, we make the sentence quality judgment mentioned above. Use this method to supplement the recall and unify the offline candidates for the final recommendation.
The above introduces relevant candidates for generating recommendation reasons in offline links. Whether it is extraction or generation, after being deployed online, it will also involve the distribution of specific online traffic. That is, when the list is full of content that meets the needs of the user's scenario, further considerations need to be taken into account: first, recommendations and query are related; second, recommendations in the list should not be too homogenized, and diverse expressions should be interspersed with; finally, the content should be kept novel, etc.
Overall, from the underlying data layer to the upstream display layer, the overall architecture will be divided into many layers, the specific structure is as follows:
Today's sharing ends, thank you everyone.
share guest: Chen Qi Meituan Advanced algorithm expert
Edited and organized by: Mao Jiahao Ping An Zhejiang Branch of China (Internship)
Production Platform: DataFunTalk
1/ Share Guest
Chen Qi｜ Meituan Search and NLP Department Advanced algorithm expert
2/ About us
DataFun: focuses on sharing and communication of big data and artificial intelligence technology applications. Initiated in 2017, it held more than 100 offline and 100+ online salons, forums and summits in Beijing, Shanghai, Shenzhen, Hangzhou and other cities, and has invited more than 2,000 experts and scholars to participate in the sharing. Its official account DataFunTalk has produced a total of 800 original articles, millions of views, and 150,000+ precise fans.
This process is mainly carried out online and is divided into six steps:- Step1: Trigger judgment to determine whether the current query is the type of pan-scene search. For example, the Palace Museum is an accurate search rather than a pan-scene search, and climbing a mountain is a pan-scene search. At the same time, this step also requires identifying relevant intentions, such as judging whether the search intention is a scenic spot, hotel or catering, etc.;
- Step2: Preprocessing the query, including word segmentation, Non-Link and identification of the target area;
- Step3: Generate candidate sequences based on the processed fragments, perform multiple combinations, and also use the technology of skipping links. For example, the existing three fragments A, B, and C may generate the following sequences of ABC, AB_C, A_B_C, and AC;
- Step4: Use the combined sequence to perform inverted index recall, including whitelist matching, pattern matching, vector recall, etc. to expand related tags;
- Step5: Tag sorting, sorting the above recall results. This process has several important features, including the correlation between the current entity and sequence, the correlation between query and entity, mention information, and the statistical characteristics after clicking aggregation and attribution. These are combined and classified, and the topN tags are selected in different businesses for application;
- Step6: The final judgment process, disambiguating and reasoning the different tags linked by multiple fragments identified.
②Technical point 2: Sort
This technical point considers how to better perform personalized sorting based on scene search. The expression of the pan-search class lacks semantic correlation in the POI name dimension, so information from the knowledge dimension needs to be supplemented in the model.
First of all, at the feature level, to enrich merchants' KG-based semantic expression vectors, the following methods are mainly adopted:
- is based on a multi-domain structure and introduces the text information of the tag;
- uses the GCN structure to train the vector between POI and query, and introduce the vector into the subsequent model.
Secondly, in terms of model structure, sequence modeling of user interest scenarios and merchant scenarios. Innovations in this work:
- currently POIs and the attention of the user's long-time sequence and shortt-time sequence. Among them, long-time sequence and short-time sequence refer to the sequence encoding generated by the behavior list of POIs that the user has clicked over a longer period of time and the short-time list of behaviors;
- introduces the sequence of tags. The identified tags of the demand scenarios that the current user is looking for are aggregated with the merchant's tags that the user has had previous behavior and become sequences; the knowledge information of the tags mounted on the POI itself is also a sequence. Encoding these two sequences and doing attention work can better capture the correlation between the user's demand scenario tags and the merchant's scenario sequence and user interest sequence under pan-scene search.
③Technical point 3: Result interpretability
This technical point considers how to better explain to users why the current results and search scenarios are related. Part of this problem is achieved through recommendation reasons. There are two ways to implement recommendation reasons: extracted recommendation reasons and generative recommendation reasons.
- Extraction recommendation reason
This method mines relevant recommendation reasons through extraction. It is mainly divided into two major scenarios during extraction:
first category, such as searching for "climbing". This type of query belongs to a specific scenario. In this specific scenario, the recommendation words we hope to provide to users are directly related to the scene, that is, the expression of the place suitable for climbing when other users come to the scenic spot. Introduction: Knowledge Graph shines in various industries and fields by its ability to describe the relationship between knowledge and everything in the world using graph models. At the same time, knowledge graph technology has also brought new challenges and opportunities to scene search. Against this background, the Meituan team has carried out a new round of technological innovations and applied knowledge graph technology to the cognition of live and tourism scenarios. The title of this sharing is "Application of Knowledge Graph in the Cognition of the Tourism Scene of Meituan Search", which mainly introduces:
- Hotel and Tourism Business Features
- Hotel and Tourism Scene Cognition
- Hotel and Tourism Knowledge Graph
- Hotel and Tourism Knowledge Graph
- search solution based on scene cognition
1
Hotel and tourism business features
report Meituan Dianping , what do you think of? Takeout, food, movies?
In fact, Meituan has more than these, and there are also hotel and tourism businesses that also occupy the top industry! What is the difference between
Hotel and tourism business and other local life services ? The difference is that the travel radius of the hotel and tourism business is larger. Most Meituan’s services, such as script killing, movies, etc., only need to find and meet the needs of users’ merchants in the current area or in the current city. For hotel and tourism business, users have a larger travel radius. Three examples are used to explain the characteristics of the hotel and tourism business better:
A Beijing user wants to ski or watch the sea, and users may choose the higher-quality Chongli Ski Resort more than 200 kilometers away from Beijing; for watching the sea, because there is no corresponding supply in Beijing, users may go to Beidaihe in Qinhuangdao . Therefore, for services that do not provide corresponding services locally, users may go to cities with corresponding services around them or even look for cities with corresponding services across the country. At this time, Meituan needs to find and filter out merchants with suitable scenarios and adapted scenarios for users.
Taking hotels as an example, introduces several commonly used search methods for users:
① Landmark/Business District. Users look for hotels around landmarks or business districts, and this model is more inclined to recommend issues;
② Merchant/Brand. This type of user has a clear goal, looking for a specific hotel or hotel under the corresponding brand;
③Pan scene search. For example, users are looking for a certain type of business, such as a youth hostel, a hotel with certain facilities, or a specific room type such as an e-sports room, etc. Other search methods mainly revolve around a combination of these three searches.
Search for attractions can also be divided into three main methods:
① Administrative region/region. If a user searches for a certain area or administrative region, Meituan needs to find the relevant attractions in the area;
② Merchant/Brand. This type of user also has a clear purpose and will directly search for the names of the corresponding attractions such as Forbidden City , Happy Valley , etc.
③Pan scene search. When users look for a certain type of demand, such as skiing and mountain climbing, at this time, multiple attractions with the same attributes, service items or the same type can meet the user's needs. In this scenario, Meituan needs to make better recommendations for users.
--
2
There are a large number of merchants of different business types in the platform, such as hotels, attractions, food, etc.When a user initiates a search on the platform, we believe that his search behavior is an explicit expression of the demand scenario, that is, he can identify the user's demand scenario through a scenario-based understanding of the user's search behavior. The user demand scenario is expressed through tags and then the merchant’s tag search can be matched with the merchant that can carry their needs.
Why do we need to do scene cognition - I hope to do different experience optimization based on different scenario needs.
Take the search for "Zhongshan Park" as an example: the original search plan will recall text-related results from the perspective of text, but this cannot meet the needs of users. Through the analysis of user behavior data, we found that most users who initiated the "Zhongshan Park" search behavior in Beijing often initiate follow-up actions around the scenes around Zhongshan Park in Beijing.
This essentially represents the scenario needs of a type of users. In order to better meet this scenario needs, we adopted the recommended method shown on the right side of the figure. First, discover the main point that the user is looking for, and then recommend other services to the user around the main point, such as related attractions, nearby food, etc., to better meet the user's diverse needs, and through this main point template style, the user can provide more information that can help make decisions.
This change in demand provides us with new solutions to the problem:
- 1 through product style upgrade to meet the diversified demands of users;
- With the transformation of product styles, it forces a new round of technological upgrades.
For the hotel and tourism business, the scenarios can be roughly divided into two categories: accurate merchant search and pan-scene search.
(1) Accurate merchant search
Accurate merchant search can be specifically divided into:
- Single main point search . For example, Forbidden City and Beijing Kuntai Hotel have only the only corresponding main point, and you can recommend services and merchants to users around this main point.
- Multi-main point search , such as Fantastic. There are many Fantastic Happy Worlds nationwide, and at this time, you can use the multi-main point style to provide information to users.
(2) Pan-scene search
Pan-scene search can be specifically divided into:
- Landmark + X, such as "Youth Hostel near the Bird's Nest", at this time, you need to recommend the main point of the Bird's Nest to the user's green point near the Bird's Nest. 2019 hotel;
- local + X. If you search for "climbing" under the city page in Beijing, you need to recommend merchants who can meet user needs and scenarios in the local and surrounding areas;
- National + X. If you search for three mountains and five mountains, you need to help users recall results that meet user scenario needs in this scenario.
The most basic problem is that the understanding of the demand scenario needs to rely on a large amount of domain knowledge in the underlying layer . The basic structured data does not carry scene-like expressions, so additional mining and understanding is required, for example:
(1) Search for Heshen Mansion. Search for Heshen Mansion, the POI of Gongwang Mansion should appear. At this time, you need to find out that Heshen Mansion is an alias of Gongwang Mansion;
(2) Search for a quiet hotel.This type of search is a pan-scene search. First, you need to explore the hotel knowledge on the B-end to explore whether the hotel has good sound insulation characteristics; second, you need to accurately correlate the online query with the existing B-end knowledge, such as the underlying knowledge label mapped to the "quiet" expression is "good sound insulation".
--
3
(1) Overview
Meituan Travel knowledge graph covers about 800,000 hotel and travel merchants, 100 million user reviews, accumulating 1 million atomic concepts, and the combined demand concepts of 1.5 million, and the relationship between 12 million merchants and concepts.
(2) Four-layer definition
In the process of graph application, the graph is divided into four layers according to the application dimension of user needs:
- Category system layer : combine different business characteristics to divide the category purpose. Taking tourism as an example, 15 first-level categories are defined, and on this basis, the specific category systems of the second-level and third-level are split;
- atomic concept layer : mining and extracting the atomic concept layer from user comments and merchant information;
- Requirements concept layer : filtering the data of the atomic concept layer and combining it in line with semantic dimensions to build concept layer data that is oriented towards search requirements. For example, mountain climbing in the atomic concept layer has a direct mapping label in the demand concept layer. For the two independent semantics of parent-child and hot springs, when there are real user demand scenarios, a more accurate dimension of "parent-child hot springs" label will be combined. This solves the semantic drift problem that occurs when searching online query, if you only use the finest-grained atomic label for recall. Therefore, when costs allow, it will be better to adopt a combination-based demand concept layer.
- POI layer : determine whether the current merchant has the attributes and characteristics represented by the tags or knowledge of the requirements concept layer.
(3) Mining process
- Knowledge extraction
data sources mainly include three categories: structured data, semi-structured data, and unstructured data. Knowledge extraction is performed from these data sources. In the early stage, if there is not enough labeled data, a semi-supervised method is used to perform Auto Phrase, dependent syntax analysis, etc.; after knowledge accumulation, the NER method is used for better extraction.
- Knowledge Classification
Knowledge fragments are classified according to the one, two, and three hierarchical knowledge system mentioned above. As shown in the figure below, the knowledge base is related to the categories related to the Taishan Scenic Area and the knowledge fragments associated at the lowest level. For the Taishan Scenic Area, the lowest data is only natural scenery, but we hope to understand its semantics to a more fine-grained size, such as canyons, waterfalls, mountains, etc. In this case, we need to pay attention to what types of natural scenery there are. Similarly, what activities are suitable for Mount Tai Scenic Area? From the perspective of the crowd, such as suitable for couples to date and graduation trip; the perspective of scene events, suitable for outdoor climbing, sunrise and sunset. These data can lay the foundation for the generation of subsequent query understanding, linking, and recommendation reasons.
(1) Definition of knowledge system
When building a knowledge graph in a vertical field, it is necessary to combine the domain knowledge to define the business. Taking scenic spots as an example, there are about 15 first-level categories, and there are also second- and third-level categories under each first-level category. The third-level classification system can better meet the search-oriented needs.
(2) Knowledge extraction
In the early stage of the business, a semi-supervised learning method, namely Bootstrapped pattern-based learning, accumulate knowledge. Taking mining animal-related entities as an example, the general process is:
- Step1: Construct the seed entity word, in this example, dog;
- Step2: According to the entity word dog, the aforementioned pattern of dog is mined from the corpus: own a, my pet;
- Step3: Use the candidate pattern to further explore the entity knowledge fragments that match the above pattern: house, cat;
- Step4: Evaluate which one house or cat is the same as dog and belongs to the animal type.
After several rounds of iterations, it was found that for animal-related entity words, my pet is a better pattern, so I used this pattern to mine and expand more entities. In the early stage, data can be quickly accumulated.
After the labeled data is available, NER's supervised learning method can be used for better generalization and recognition. The BERT+CRF model was initially used for extraction, but this model easily chopped up semantic fragments of knowledge entities. As shown in the figure below, the comment is "The Four Wonders of Emei in the Golden Pin", and the "Four Wonders of Emei" was drawn, but we hope that the clip can be as complete as possible. Later, the problem was solved by introducing KG-related information. First, the text that needs to be extracted is divided into words, and two layers of vector information of Character-level granularity and Word-level granularity are introduced to assist in determining the boundary of segmentation of segments, which can effectively solve the problem of fragments being chopped.
At first, for efficiency, each of the 15 first-level categories mentioned above was extracted separately. In order to improve the accuracy, the relevant major categories were put together for comprehensive knowledge extraction. In this process, there will be problems of inaccurate knowledge classification. Further optimization is to use a multi-task joint training method, that is, to integrate the knowledge classification tasks and NER tasks.
The general idea is : through the vector information after the first layer encoding (using BERT or BiLSTM encoding), then NER segmentation is used with traditional CRF, and then the vector information is introduced into the classification layer for identification and processing using the idea of casecade cascade. This work has improved the overall accuracy and recall rate relatively well (published in the 2022 DASFAA meeting).
(3) Knowledge Classification
After the initial classification of knowledge for the first-level category, the knowledge needs to be classified in a more fine-grained manner. However, due to business characteristics, many knowledge fragments will belong to multiple nodes in the second-level or third-level categories, so the multi-label classification task is used here to classify the original text fragments. For example, when VR projects "fly over the horizon", when using BERT to encode and directly identify the fragment "fly over the horizon", it is easy to classify it as a roller coaster project, because the expression word "fly over XXX" in more cases does not refer to VR projects but roller coasters or other projects.
In order to better solve this problem, after extracting the fragment, the context information of the fragment is introduced into the model to enrich the context expression. At the same time, feature engineering is carried out in the search log and comment log, manually constructed features are added, and these features are fused and classified uniformly, so that the problem of semantic offsets can be solved in separate texts.
(4) Knowledge is related to merchants
After extracting the knowledge fragment and classifying it, it is necessary to solve the problem of association of knowledge (tags) with platform merchants.
First of all, it should be clear that the above problem is not a closed domain label problem but an open domain label problem, that is, it is not a simply a classification and mount problem of the category system, because tags will continue to accumulate and add with the development of the business and mining. We need to classify these new tags and knowledge, so marking requires a certain generalization ability.
in combination with the platform: the main body of the platform is a merchant, and we need to find merchants with knowledge or tags. Merchants have many user reviews and products listed below. What we can intuitively obtain is the correlation between each evaluation and knowledge and the correlation between each product and knowledge. It feedbacks the correlation between merchants, and there is an additional layer of aggregation process in the middle. The solution to the problem of
The multi-person voting mechanism , that is, every piece of information hanging under the merchant is a feedback from a user. To determine whether it is related or not, or other opinions, the aggregation and voting of this information can make the merchant have this knowledge or tag.
As an example, determine whether a certain attraction can bring pets:
- Step1: Find evidence and find text expressions related to bringing pets;
- Step2: determine the authenticity of the extracted short fragments, which are mainly divided into positive correlation, irrelevant, and negative correlation;
- Step3: Multiple evidence fusion classification. In addition to the evidence correlation obtained earlier, many dimensions of feature information are abstracted based on semantic features and statistical features, such as the text correlation of the POI itself information. This process mainly uses the BERT model to match text correlation.
- Step4: Distribution confidence determination, feed the correlation obtained in the first two steps into the tree model, and finally obtain the classification result, that is, whether it is related.
If knowledge classification is performed online, there are requirements for accuracy, allowing certain recall losses, but it is necessary to ensure that the results are accurate, so that users can experience better online. Therefore, in the last step, the process of determining the distribution confidence is increased. That is, the distribution statistics of the merchant categories in the marking results are filtered to filter the POI of the long-tail category. For example, when a user searches for a mountain climb, after the category statistics, if individual judgments are wrong, the hot spring category accounts for 0.3%, and this type of result is filtered according to the threshold.
(5) MT-BERTh
Whether it is knowledge extraction or knowledge classification, the BERT model is used. Meituan mainly uses the self-developed MT-BERT, which is characterized by introducing a large amount of user comment information in Meituan's business scenario and business information hanging under the merchant to better adapt to the model.
MT-BERT After adding Meituan UGC data, it has made significant improvements in some public data, internal query intention classification and component analysis tasks.
--
4
divides service search into five layers:
- L0 layer : mine relevant knowledge and build index ;
- L1 layer : identify and understand user query and perform structured recall;
- L1 layer : identify and understand user query and perform structured recall;
- L2 layer : Sorting the list in deep learning model based on the recall results;
- L3 layer : Making targeted strategy adjustments in different business scenarios;
- L4 layer : When presenting the list, provide tags, recommendation reasons, lists and other interpretable information to strengthen perception.
(1) Accurate merchant search
model the problem: first, identify the main point; secondly, recommend related merchants around the main point, including attractions and nearby hotels, etc. The second step is to subdivide in detail: When the user is in the planning decision-making process, the user can recommend merchants that can replace the main point; when the user has determined to consume at a certain main point, the user can recommend merchants that can match itineraries
There are two technical points mentioned above:
- How to identify the main point, it is called the merchant link or main point link, that is, a variant of the entity link. How to better recommend
- How to revolve around the main point.
①Technical point 1: How to identify the main point
adopts the entity linking scheme based on context information . The reason for using this solution is: in Meituan's business, the required results may be different in different geographical locations. For example, when searching for Dragon Dream Hotel near Dragon Dream Hotel, No. 492 Anhua Road, Shanghai, the user is most likely looking for Dragon Dream Hotel near him.
The specific strategy is mainly divided into two fractions : the first score is a word sequence, that is, the probability prediction score between the sequence fragment that can be linked to a certain entity and the probability prediction score between the entity, and the alias expression of the merchant by search logs and knowledge graphs is obtained; the second score is the information score based on the context, and the score is divided into two parts based on the business characteristics. First, the semantic score of the context of the text semantics itself; second, the geographical context score, that is, the score is calculated based on the distance between the user and the merchant.
②Technical point 2: How to better recall
Two methods are used for recall:
- Based on the vector correlation of the user behavior sequence
Constructs all POIs clicked by the user to form a sequence, and encodes it based on skip-gram to obtain a POI vector, and then calculates it with the POI vector of the main point to obtain a head POI similar to it. It is worth noting that we have made targeted adjustments based on the business characteristics: First of all, the hotel and tourism business is a relatively low-frequency business, so we will aggregate users' behavior sequences over a longer period of time and focus on solving the low-frequency problem. Secondly, with the differences in the behavioral characteristics of different cities, we have introduced side-information such as cities, time, categories, prices, etc. to better calculate the correlation of vectors.
- Based on GCN vector correlation
Each POI has relevant knowledge. We constructed a heterogeneous graph of User-Query-POI-Item (knowledge) and obtained the vector of POI through the graph learning method.
(2) Pan-scene search
Model the problem: First, identify which scene needs are there; second, search merchants based on the scene and sort them in a personalized manner; finally, explain why the results can meet the scene needs.
gives an example. If you search for "a park suitable for walking your children" in Beijing, the area you are looking for is Beijing. Different text fragments in the query will be linked to different tags, and there is also a concept of master-slave between tags. Therefore, you can infer the superior scene, homoslovakia, triggered lists, and tags that need to be displayed, and finally form a recall syntax for subsequent processing.
pan-scene search involves three main technical points:
①Technical point 1: Pan-scene link
0 1 Recognize query, that is, link the scene tag.This process is mainly carried out online and is divided into six steps:
- Step1: Trigger judgment to determine whether the current query is the type of pan-scene search. For example, the Palace Museum is an accurate search rather than a pan-scene search, and climbing a mountain is a pan-scene search. At the same time, this step also requires identifying relevant intentions, such as judging whether the search intention is a scenic spot, hotel or catering, etc.;
- Step2: Preprocessing the query, including word segmentation, Non-Link and identification of the target area;
- Step3: Generate candidate sequences based on the processed fragments, perform multiple combinations, and also use the technology of skipping links. For example, the existing three fragments A, B, and C may generate the following sequences of ABC, AB_C, A_B_C, and AC;
- Step4: Use the combined sequence to perform inverted index recall, including whitelist matching, pattern matching, vector recall, etc. to expand related tags;
- Step5: Tag sorting, sorting the above recall results. This process has several important features, including the correlation between the current entity and sequence, the correlation between query and entity, mention information, and the statistical characteristics after clicking aggregation and attribution. These are combined and classified, and the topN tags are selected in different businesses for application;
- Step6: The final judgment process, disambiguating and reasoning the different tags linked by multiple fragments identified.
②Technical point 2: Sort
This technical point considers how to better perform personalized sorting based on scene search. The expression of the pan-search class lacks semantic correlation in the POI name dimension, so information from the knowledge dimension needs to be supplemented in the model.
First of all, at the feature level, to enrich merchants' KG-based semantic expression vectors, the following methods are mainly adopted:
- is based on a multi-domain structure and introduces the text information of the tag;
- uses the GCN structure to train the vector between POI and query, and introduce the vector into the subsequent model.
Secondly, in terms of model structure, sequence modeling of user interest scenarios and merchant scenarios. Innovations in this work:
- currently POIs and the attention of the user's long-time sequence and shortt-time sequence. Among them, long-time sequence and short-time sequence refer to the sequence encoding generated by the behavior list of POIs that the user has clicked over a longer period of time and the short-time list of behaviors;
- introduces the sequence of tags. The identified tags of the demand scenarios that the current user is looking for are aggregated with the merchant's tags that the user has had previous behavior and become sequences; the knowledge information of the tags mounted on the POI itself is also a sequence. Encoding these two sequences and doing attention work can better capture the correlation between the user's demand scenario tags and the merchant's scenario sequence and user interest sequence under pan-scene search.
③Technical point 3: Result interpretability
This technical point considers how to better explain to users why the current results and search scenarios are related. Part of this problem is achieved through recommendation reasons. There are two ways to implement recommendation reasons: extracted recommendation reasons and generative recommendation reasons.
- Extraction recommendation reason
This method mines relevant recommendation reasons through extraction. It is mainly divided into two major scenarios during extraction:
first category, such as searching for "climbing". This type of query belongs to a specific scenario. In this specific scenario, the recommendation words we hope to provide to users are directly related to the scene, that is, the expression of the place suitable for climbing when other users come to the scenic spot.For this kind of recommendation reason, text matching is used and BERT-MRC is used to recall candidate sentences.
second category, such as searching for "Shanghai Tourism". The scope of this type of query is relatively broad, so in this scenario, the default feature will be recommended to users, namely the characteristics description of the scenic spot itself. For this type of recommendation words that can directly represent the characteristics of the merchant, a combination of short sentences and the idea of pointer-generator network extraction is used to generate candidate sentences.
When there is a candidate recommendation, whether it is the user's recommendation or the merchant's own recommendation, it will be uniformly entered into the judgment module to make a series of quality judgments on the candidate sentences, including rewriting of the sentence, expressing smoothness, expressing emotions, etc. Through these modules, multi-dimensional scores are obtained, and finally feeding them into the tree model as features for overall quality judgment, and obtaining the final judgment score.
- generated recommendation reason
extracted recommendation reason can solve most problems, but there are some problems:
sentence expression is correct but long, and there are sentence length restrictions when the front end shows it to users. Therefore, in order to better use longer sentences or sentences with no problem but express relatively awkward and bright spots, some sentences are compressed by a transformer-based network structure to make them meet the limitations of sentence length.
In addition, many POIs have fewer comments themselves, and the user's expression quality is poor, and we have strict overall control over the recommendations, so we cannot dig out recommended words related to the scene.
At this time, we consider generating recommendations based on the merchant's existing knowledge and expanding the data. This idea is implemented based on the KG and scenario keyword generation scheme. There are two key points, giving examples: First, "Sima Tai Great Wall" has tags related to POI in the Word embedding layer; second, control tags to control relevant recommendations based on the "Spring Outing" tag. Through the two dimensions, the generated fragment is finally obtained. After encoder, I got "the spring flowers are romantic and the summer is full of green". Based on this sentence, we make the sentence quality judgment mentioned above. Use this method to supplement the recall and unify the offline candidates for the final recommendation.
The above introduces relevant candidates for generating recommendation reasons in offline links. Whether it is extraction or generation, after being deployed online, it will also involve the distribution of specific online traffic. That is, when the list is full of content that meets the needs of the user's scenario, further considerations need to be taken into account: first, recommendations and query are related; second, recommendations in the list should not be too homogenized, and diverse expressions should be interspersed with; finally, the content should be kept novel, etc.
Overall, from the underlying data layer to the upstream display layer, the overall architecture will be divided into many layers, the specific structure is as follows:
Today's sharing ends, thank you everyone.
share guest: Chen Qi Meituan Advanced algorithm expert
Edited and organized by: Mao Jiahao Ping An Zhejiang Branch of China (Internship)
Production Platform: DataFunTalk
1/ Share Guest
Chen Qi｜ Meituan Search and NLP Department Advanced algorithm expert
2/ About us
DataFun: focuses on sharing and communication of big data and artificial intelligence technology applications. Initiated in 2017, it held more than 100 offline and 100+ online salons, forums and summits in Beijing, Shanghai, Shenzhen, Hangzhou and other cities, and has invited more than 2,000 experts and scholars to participate in the sharing. Its official account DataFunTalk has produced a total of 800 original articles, millions of views, and 150,000+ precise fans.
2/ About us
DataFun: focuses on sharing and communication of big data and artificial intelligence technology applications. Initiated in 2017, it held more than 100 offline and 100+ online salons, forums and summits in Beijing, Shanghai, Shenzhen, Hangzhou and other cities, and has invited more than 2,000 experts and scholars to participate in the sharing. Its official account DataFunTalk has produced a total of 800 original articles, millions of views, and 150,000+ precise fans.
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-