As the cornerstone of the digitalization of the new retail industry, the Product knowledge map provides a precise structured understanding of the product and plays a crucial role in business applications. Compared with the original merchant mapping in Meituan Brain, the commodity mapping needs to deal with more dispersed, complex and massive data and business scenarios, and faces challenges such as low quality information sources, multiple data dimensions, and reliance on common sense and professional knowledge. This paper will focus on the retail product knowledge map, introduce meituan in the construction of commodity hierarchy, attribute system construction, map construction of human efficiency improvement and other aspects of exploration, hoping to help or inspire you.

background

Meituan brain

In recent years, artificial intelligence is rapidly changing people’s lives. There are actually two technological driving forces behind it: deep learning and knowledge mapping. Deep learning is classified as an implicit model, which is usually oriented to a specific task, such as playing go, cat recognition, face recognition, speech recognition and so on. Generally speaking, it can achieve excellent results on many tasks, but it also has some limitations, for example, it requires massive training data and powerful computing power, it is difficult to carry out cross-task migration, and it does not have good interpretability. On the other hand, knowledge graphs, as explicit models, are also one of the technical drivers of ARTIFICIAL intelligence, and can be applied to a wide range of tasks. Compared with deep learning, knowledge in knowledge graph can be precipitated, which is more explicable and closer to human thinking. It supplements human knowledge accumulation for implicit deep model and is complementary to deep learning. Therefore, many large Internet companies in the world are actively layout in the field of knowledge graph.

Meituan connects hundreds of millions of users and tens of millions of merchants, and there is a wealth of knowledge about daily life behind it. In 2018, meituan Knowledge Graph team started to build Meituan Brain, focusing on using knowledge graph technology to empower business and further improve user experience. Specifically, Meituan brain in Meituan business involves the do level of merchants, the level of food/commodity, billions of user comments, and millions of levels of the scene behind the in-depth understanding and structured knowledge modeling, building, shop, commodity and scene correlation between knowledge, thus forming a large-scale knowledge in the field of life service. At present, Meituan Brain has covered billions of entities and tens of billions of triples, proving the effectiveness of knowledge mapping in catering, takeout, hotel, finance and other scenarios.

Exploration in the new retail field

Meituan has gradually broken through the original boundaries and explored new businesses in the life service field. It is not limited to helping people “eat better” through takeout and catering, but in recent years, it has also gradually expanded to retail, travel and other fields to help people “live better”. In the retail field, Meituan has successively implemented a series of corresponding businesses such as Meituan flash purchase, Meituan food purchase, Meituan selection, and Tuan good goods, gradually realizing the vision of “everything goes home”. In order to better support the new retail business of Meituan, we need to establish the knowledge map of the retail products behind, accumulate structured data, and deeply understand the commodities, users, attributes and scenarios in the retail field, so as to better provide services in the retail product field for users.

Compared with restaurants, takeout and hotels, the construction and application of knowledge mapping in retail goods field is more challenging. On the one hand, the quantity of goods is larger and the scope of coverage is wider. On the other hand, the display information of the commodity itself is often sparse, and it is largely necessary to combine common knowledge in life for reasoning, so as to complete the properties hidden in the tens of dimensions and complete the understanding of the commodity. In the example below, simple product description like “Lay’s Cucumber flavor” actually corresponds to rich implicit information. Only after structural extraction of these knowledge and corresponding knowledge inference, can the optimization of downstream search and recommendation modules be better supported.

Objectives of commodity atlas construction

According to the characteristics of Meituan retail business, we have developed a multi-level, multi-dimensional and cross-business retail product knowledge map system.

A multi-level

In different application scenarios of different services, the definition of “goods” may vary, and the goods with different granularity need to be understood. Therefore, in our retail product knowledge map, we have established a five-level hierarchy system, including:

  • L1- Commodity SKU/SPU: the granularity of commodities sold in the corresponding business is the object of users’ transactions, which are usually commodities hung by merchants, such as “Mengniu low-fat and high-calcium milk 250ml box sold by Wangjing Carrefour”. This level is also the cornerstone of the lowest level of the commodity map, linking the business commodity repository with the knowledge of the map.
  • L2- Standard goods: describe the objective granularity of the goods themselves, for example, “Mengniu low-fat and high-calcium milk 250ml box”. No matter what channels and merchants buy the goods, there is no difference in the goods themselves. The bar code is the objective basis at the level of standard goods. At this level, we can model objective knowledge surrounding a standard product, such as a standard product having the same brand, taste, packaging, and other attributes.
  • L3- Abstract commodities: Further, we will abstract the standard commodities upward, such as “Mengniu Low fat and High calcium milk”. At this level, we no longer pay attention to the specific packaging and specifications of commodities, but aggregate commodities of the same series into abstract commodities, bearing users’ subjective cognition of commodities, including nicknames, brand cognition and subjective evaluation of commodity series.
  • L4- Main product category: describe the essence of the main product category, such as “eggs”, “cream of strawberry”, “desktop sausage” and so on. As the backstage category system of the commodity atlas, this layer models the categories in the commodity field in an objective way, carrying users’ demands for commodities. For example, eggs of various brands and origins can meet users’ demands for this category.
  • L5- Business category: Compared with the background category system of the main category, business category as the foreground category system will be manually defined and adjusted according to the current development stage of the business. Each business will establish the corresponding foreground category system according to the characteristics and requirements of the current business stage.

multidimensional

  • Commodity attribute perspective: Around the commodity itself, we need a large number of attribute dimensions to describe the commodity. The dimension of commodity attribute is mainly divided into two categories: one is general attribute dimension, including brand, specification, packaging, origin, etc.; For example, for milk, we will pay attention to fat content (full fat/low fat/skim milk) and storage mode (normal temperature milk, frozen milk). Commodity attribute mainly describes the objective knowledge of commodities, which is often built on the level of standard commodities.
  • User cognitive perspective: Besides objective commodity attribute dimension, users will have a series of subjective cognition for commodity, such as commodity alias, commonly known as (” the little black bottle “, “happy water”), the evaluation of the product (” sweet “, “melt in your mouth,” “cost-effective”), the listing of the goods/list (” imported food list “, “summer summer stock”) and other dimensions. These subjective perceptions tend to be based on the level of abstract goods.
  • Category/category perspective: From the category/category perspective, different categories/categories may have different concerns. At this level, we will model what typical brands are under each category/category, what typical attributes users care about, and how long the repurchase cycle is for different categories.

Across business

The goal of meituan Brain commodity knowledge Atlas is to model commodity knowledge in the objective world, instead of being limited to a single business. In the five-level system of commodity atlas, standard commodity, abstract commodity and category system are decoupled from business and built around objective commodity, and all dimension data built around these levels also depict objective knowledge of commodity field.

When applied to each business, objective atlas knowledge can be associated upward to the business foreground category and downward to the business commodity SPU/SKU, so as to complete the access of each business data, realize the connection between each business data and objective knowledge, and provide a more comprehensive cross-business panoramic data perspective. Using such data, we can more comprehensively model and analyze users’ preference for business and category, as well as their sensitivity to price and quality. In terms of commodities, we can more accurately model the repurchase cycle, region/season/festival preference of each category.

The challenges of commodity atlas construction

The challenges to the construction of the commodity knowledge graph mainly come from the following three aspects:

  1. Low quality of information source: the product itself has a lack of information, often in the title and pictures. Especially in the context of LBS electronic shopping mall like Meituan flash purchase, merchants need to upload a large amount of commodity data, and there are many incomplete information for commodity information input. In addition to titles and pictures, commodity details also contain a lot of knowledge information, but their quality is often uneven and their structures are different, so it is extremely difficult to mine knowledge from them.
  2. Multiple data dimensions: There are many data dimensions that need to be built in the commodity field. Taking commodity attributes as an example, we not only need to build general attributes, such as brand, specification, packaging, taste and other dimensions, but also cover specific attribute dimensions under each category/category, such as fat content, sugar content, battery capacity, etc. The whole will involve attribute dimensions of hundreds of dimensions. Therefore, the efficiency of data construction is also a big challenge.
  3. Rely on common sense/professional knowledge: people in daily life because there are lot of common sense knowledge accumulation, can be very short description for its hidden information, for example, in see “treat cucumber” knows when such a product is actually treat cucumber flavor of potato chips, see “tang’s monk meat” know when in fact this is not a kind of meat, but a snack. Therefore, we also need to explore semantic understanding methods combined with common knowledge. At the same time, in the field of medicine, such as a nurse, the construction of the graph depends on the strong professional knowledge, such as the relationship between diseases and drugs, and this kind of relations to the requirement of accuracy is extremely high, all knowledge need to be accurate, so also need a good combination of expert and algorithm for efficient map construction.

Commodity Map construction

After understanding the goals and challenges of atlas construction, we will introduce specific solutions of commodity atlas data construction.

Hierarchical system construction

Category system construction

Essential category describes the most detailed category to which the essence of a commodity belongs. It aggregates a category of commodities and bears the ultimate consumer demand of users, such as “high calcium milk” and “beef jerky”. There is also a certain difference between essential category and category. Category is a collection of several categories, which is an abstract concept of category and cannot be specified to a specific commodity category, such as “dairy products” and “fruit”.

Category marking: For the construction of commodity atlas, the key step is to establish the association between commodity and category, that is, to label the commodity category. Through the association between commodities and categories, we can establish the association between commodities in the commodity database and user needs, and then show specific commodities to users. The following is a brief introduction to the category marking method:

  1. Category word list construction: Category marking first needs to build a preliminary commodity category word list. First of all, we obtain preliminary product candidate words through word segmentation, NER, new word discovery and other operations on commodity database, search log, merchant label and other data sources of each e-commerce business of Meituan. Then, the dichotomous model is trained (to judge whether a word is a category) by labeling a small number of samples. In addition, we combined the method of active learning to pick out the indistinguishable samples from the predicted results and label them again, and continue to iterate the model until the model converges.
  2. Category marking: First of all, we identify the named entity of the commodity title, and combine with the category word table in the previous step to obtain the candidate category of the commodity, such as identifying “skimmed milk” and “milk” in “Mengniu Skimmed Milk 500ml”. Then, after we have the product and its category, we use the monitoring data to train the binary classification model for category marking. We input the Pair of SPU_ID and TAG (

    ) to predict whether it will match. Specifically, on the one hand, we use the rich semi-structured corpus in the business to construct statistical features around tag words; on the other hand, we use named entity recognition, Bert-based semantic matching and other models to produce high-order correlation features. On this basis, we input the above features into the final judgment model for model training.
    ,>
  3. Category label post-processing: In this step, we carry out some post-processing strategies for categories typed on the model, such as category cleaning strategies based on picture correlation and entity recognition results of commodity titles.

Through the above three steps, we can establish the relationship between the product and the category.

Category system: Category system consists of category and relationship between categories. Common category relationships include synonyms and superposition, etc. In the process of building category system, the following methods are commonly used to complete the relationship. We mainly use the following methods:

  1. Rule-based category relationship mining. In the encyclopedia and other general corpus data, some categories have fixed pattern description, such as “corn is also known as corn, corn stick, maize, pearl rice, etc.”, “durian is one of the famous tropical fruits”, so synonyms and upper and lower positions can be extracted by using rules.
  2. Category relation mining based on classification. Similar to the category marking method mentioned above, synonyms and upper and lower bits are constructed as samples of

    , and the binary classification model is used to judge the validity of category relations through statistical features mined from commodity databases, search logs, encyclopedia data, UGC and semantic features obtained based on sentence-bert. For the trained classification model, we also select difficult samples in the results through active learning and carry out secondary labeling, so as to continuously iterate the data and improve the performance of the model.
    ,>
  3. Category relation inference based on graph. After obtaining the preliminary synonym and up-down relation, we use the existing relation to construct the network, and use GAE, VGAE and other methods to predict the network link, so as to complete the edge relation of the atlas.

Standard/abstract goods

Standard goods is to describe the objective fact of the particle size of the goods itself, and sales channels and merchants have nothing to do, and the bar code of goods is the objective basis of the standard goods. Standard item association means that all the business SKU/SPU that belong to the bar code of a certain commodity are correctly associated with the bar code of the commodity, so as to model the corresponding objective knowledge on the standard commodity level, such as the corresponding attributes of the brand, taste and packaging of the standard commodity. The following is a case to illustrate the specific tasks and programs associated with the standard.

** Case: ** The picture below is a standard commodity of a bull three meter cord board. Merchants input information, will be directly related to the commodity bar code. Through merchant data entry, part of the bid association is completed, but the proportion of this part is relatively small, and there are a lot of link missing, link error problems. In addition, different businesses for the same standard products, the description of the title of goods is strange. Our goal is to fill in the missing links and link the product to the correct standard.

Aiming at the task of bidding association, we construct a synonym discrimination model in the commodity domain: using a small amount of related data already provided by merchants in the way of remote supervision, as the existing knowledge graph to construct training samples of remote supervision. In the model, positive examples are standard codes with high confidence. Negative examples are SPUs with similar commodity names or images in the original data that do not belong to the same standard. After the training samples with high accuracy were constructed, the synonym model was trained by BERT model. Finally, the final accuracy rate can reach more than 99% through the model’s independent denoising method. Overall can achieve brand, specifications, packaging and other dimensions sensitive.

Abstract goods are the level of users’ cognition. As the object of users’ comments, this layer is more effective in modeling users’ preferences. At the same time, in the display of decision-making information, abstract commodity granularity is also more consistent with user cognition. For example, in the list of ice cream as shown in the following figure, skUs corresponding to abstract goods in users’ cognition are listed, and then characteristics and recommendation reasons of different abstract goods are displayed accordingly. The overall construction method of the abstract commodity layer is similar to that of the standard commodity layer, which adopts the model flow of the standard commodity association and adjusts the rules in the data construction part.

Attribute dimension construction

A comprehensive understanding of a commodity needs to cover all attribute dimensions. For example, “Lay’s cucumber flavor potato chips” needs to explore its corresponding brand, category, taste, packaging specifications, label, origin and user comment characteristics and other attributes, so as to accurately reach users in product search and recommendation scenes. The source data of commodity attribute mining mainly includes commodity title, commodity picture and semi-structured data.

The product title contains the most important information dimension for the product. At the same time, the product title analysis model can be applied in the understanding of query, which can quickly and deeply understand and split the user, and can also provide high-order features for the downstream recall ranking. Therefore, here we focus on the use of commodity title attribute extraction method.

Item title parsing as a whole can be modeled as a text sequence annotation task. For example, for the commodity title “Lay’s Cucumber potato chips”, the goal is to understand the various components in the text sequence of the title, such as Lay’s corresponding brand, cucumber corresponding taste, and potato chip is a category. Therefore, we use named Entity Recognition (NER) model to analyze the commodity title. However, there are three major challenges in product title analysis :(1) less context information; (2) Rely on common knowledge; (3) Annotation data usually have a lot of noise. To solve the first two challenges, we first tried to introduce atlas information into the model, which mainly includes the following three dimensions:

  • Node information: Map entities are used as dictionaries and soft-lexicon access to alleviate the problem of NER boundary segmentation errors.
  • Related information: Commodity title analysis relies on common knowledge. For example, in the absence of common knowledge, we cannot confirm whether “cucumber” is a commodity category or taste attribute only from the title “Lay’s cucumber potato chips”. Therefore, we introduce the knowledge map of associated data relieves the problem of lack of common sense knowledge: in the knowledge map, pleasure and chips exist between “brand – sales – category” relationship, but there is no direct relationship between treat with cucumber, so you can use graph structure to relieve the shortage of common sense knowledge NER model. Specifically, we use Graph Embedding technology for embedded characterization of map, using the map Graph structure information to map the words, words, said and then contains a Graph structure information of embedded said and text semantic representation for fusion splicing, access to the NER model, that half of the model can take into account the semantics, Common sense information is also taken into account.
  • Node type information: The same word can represent different attributes. For example, “cucumber” can be used as both category and attribute. Therefore, when Graph Embedding modeling for the Graph, we split the entity nodes according to different types. When the node representation of the atlas is added into NER model, the attention mechanism is used to select the representation of entity types that are more consistent with semantics according to the context, so as to alleviate the problem of different meanings of words under different types and realize the integration of different types of entities.

Next, we discuss how to mitigate the noise of labeling. In the labeling process, the problem of missing or mismarking is unavoidable, especially in the labeling of commodity title NER, which is more complicated. For the noise problem in the annotation data, the following methods are adopted to optimize the noise annotation: the original Hard training method is no longer adopted, but the Soft training method based on confidence data. Then, the iterative cross-verification is carried out by Bootstrapping, and adjustments are made according to the confidence of the current training set. Through experimental verification, Soft training +Bootstrapping multi-round iteration method is used to significantly improve the model effect on the data set with large noise ratio. For detailed methods, please refer to our paper Iterative Strategy for Named Entity Recognition with Imperfect Annotations in NLPCC 2020.

efficiency

The construction of knowledge graph is usually a separate mining method for the data of each domain dimension. This mining method is manual and inefficient. For each different field and data dimension, we need to customize the construction of task-related feature and annotation data. In the commodity scenario, there are so many dimensions to mining that efficiency gains are critical. We first model the knowledge mining task into three classification tasks, including node modeling, relationship modeling and node association. In the training process of the whole model, the two steps mentioned above are the most needed for efficiency optimization :(1) feature extraction for tasks; (2) Data annotation for tasks.

In view of the feature extraction part, we avoided for different mining tasks do customization features mining way, but try to decoupling characteristics and tasks, build a cross task graph mining characteristics of general system, using the vast amounts of characteristic library to characterization of target node/relationship/connection, and use the supervision and training data for the combination of the characteristics and selection. Specifically, the map feature system constructed by us mainly consists of four types of feature groups:

  1. The feature of rule template is to integrate the ability of rule model with the prior knowledge of human.
  2. The feature of statistical distribution can make full use of all kinds of corpus and make statistics based on different levels of different corpus.
  3. The syntactic analysis features utilize the NLP domain’s modeling ability to introduce dimensional features such as participle, part of speech, and syntax.
  4. The embedded representation feature is the ability to introduce BERT and other semantic understanding models by using higher-order model capabilities.

In terms of data annotation, we mainly improve efficiency from three perspectives.

  1. Through semi-supervised learning, full use of unlabeled data for pre-training.
  2. Through the active learning technique, the samples that can provide the most information gain for the model are selected for annotation.
  3. The remote supervision method is used to construct the remote supervision samples through the existing knowledge to train the model, and the value of the existing knowledge can be brought into play as much as possible.

Man-machine integration – professional atlas construction

At present, the structure of the healthcare industry is changing, and consumers are more inclined to use online healthcare solutions and drug delivery services, so the pharmaceutical business has gradually become one of the important businesses of Meituan. Compared with the construction of general commodity knowledge map, drug domain knowledge has the following two characteristics :(1) it is highly professional and requires relevant background knowledge to judge the corresponding attribute dimensions, such as the applicable symptoms of drugs. (2) High accuracy requirements, for strong professional knowledge is not allowed to make mistakes, otherwise it is more likely to lead to serious consequences. Therefore, we combined intelligent model and expert knowledge to construct drug knowledge map.

The knowledge in drug atlas can be divided into weak professional knowledge and strong professional knowledge. Weak professional knowledge refers to the knowledge that can be easily acquired and understood by ordinary people, such as the method of drug use and applicable population. And strong professional knowledge is the knowledge that needs to be judged by talents with professional background, such as the indications and symptoms of drugs. Since these two types of data rely on experts to different degrees, different mining links are adopted respectively:

  • Weak professional knowledge: for drug atlas weak professional knowledge mining, we from the manual, encyclopedic knowledge, such as the data source to extract the corresponding information, and combining with the rules of expert knowledge to precipitate out strategy, with the help of general semantic model to extract the corresponding knowledge, and through expert batch sampling, completed the construction of data.
  • Strong professional knowledge: For the mining of strong professional knowledge of drug atlas, in order to ensure 100% accuracy of relevant knowledge, we extracted candidates of drug related attribute dimensions through the model, and then gave these candidate knowledge to experts for full quality inspection. Here, we mainly reduce the energy expenditure of professional pharmacists on basic data as much as possible through the ability of algorithms, and improve the efficiency of experts in extracting professional knowledge from semi-structured corpus.

In specialized fields such as pharmaceuticals, there are often differences in the presentation of expertise and user habits. Therefore, in addition to mining strong and weak professional knowledge, we also need to fill in the difference between professional knowledge and users, so as to better combine drug atlas with downstream applications. To this end, we mined alias data of diseases, symptoms and efficacy, as well as commonly known data of generic drug names, from data sources such as user behavior logs and daily conversations in the field, to open up the path between user habits and professional expression.

Application of commodity atlas

Ever since Google incorporated the knowledge Graph into its search engine and dramatically improved search quality and user experience, the knowledge graph has played an important role in all verticals. In the field of Meituan products, we have also applied the product atlas effectively in the search, recommendation, merchant side, user side and other downstream scenarios of commodity business. Next, we will introduce several typical cases.

Structured recall

Data from the commodity atlas is very helpful for understanding commodities. For example, in the product search, when the user searches for headache and back pain, he can know which medicine has the effect of relieving pain through the structured knowledge graph. When searching for cute strawberry and cucumber potato chips, users need to rely on the common knowledge of the graph to understand what they really want is ice cream and potato chips, not strawberries and cucumbers.

Generalization of ordering model

On the one hand, the category information, category information and attribute information of the atlas can be used as a relatively strong correlation judgment method and intervention means, and on the other hand, it can provide the commodity aggregation ability of different coarse and fine granularity, and provide the ranking model with generalization features, which can effectively improve the generalization ability of the ranking model. It’s even more valuable in the commodity space, where user behavior is sparse. Specific features include:

  1. Commodity aggregation is carried out through the granularity, and the ranking model is connected to the ID feature.
  2. Statistical characteristics were constructed after the aggregation of each particle size.
  3. The high dimensional vector representation of commodities is combined with the ranking model by means of graph embedding representation.

Multimodal map embedding

Existing research work has proved in many fields that embedding the data of the knowledge graph and combining it with the ranking model in the form of high-dimensional vector representation can effectively alleviate the problem of data sparseness and cold start in the ranking/recommendation scenario by introducing external knowledge. , however, the traditional graph embedding work, often ignored the knowledge map of multimodal information, such as the commodities we have in the field of image, the title of the goods, the introduction of businesses such as the simple mapping of the node type knowledge, to further improve the introduction of this information can also be embedded map to recommend/sort of information gain.

There are some problems when the existing map embedding methods are applied to multi-modal map representation, because in multi-modal scenes, the meaning of edges in the map is no longer a simple semantic inference relationship, but a multi-modal information supplement relationship. Therefore, we also aimed at the characteristics of multi-modal maps. MKG Entity Encoder and MKG Attention Layer are proposed to better model the multi-modal knowledge map, and their representation is effectively integrated into the recommendation/ranking model. For the specific method, please refer to our paper multi-modal Knowledge Graphs for Recommender Systems published in CIKM 2020.

User/merchant side optimization

The product atlas provides explicit explanatory information on the client side to assist users in making decisions. Specific forms of presentation include screening items, feature labels, lists, recommendation reasons, etc. The dimension of the filter item is determined by the attribute category concerned by the user under the corresponding category of the current query word. For example, when the user searches for potato chips, the user usually focuses on its taste, packaging, net content, etc. We will display the filter item according to the enumeration value of the supplied data under these dimensions. The characteristic label of the product comes from the extraction of the title, the detailed page information of the product and the comment data, showing the characteristics of the product with concise and clear structured data. The reasons for recommending products can be obtained through comment extraction and text generation, which are linked with query words to give the reasons why products are worth buying from the perspective of users, while the list data is more objective, reflecting the quality of products with real data such as sales volume.

On the merchant side, namely the merchant publishing side, the commodity atlas provides real-time prediction ability based on the commodity title, helping merchants to mount the category and improve the attribute information. Businesses, for example, fill in the title “Germany import deaton skim milk box of 12”, after the goods map provides online category forecast service can be mounted to the “food and beverage, dairy products – pure milk” category, and through the entity recognition service, get goods “origin – Germany”, “whether imports – the import”, “brand – patient”, “fat – skim”, The attribute information of “specification -12 boxes” will be confirmed and released by the merchants after the prediction is completed, which can reduce the maintenance cost of the merchants’ commodity information and improve the information quality of the released commodities.

Author’s brief introduction

Xue Zhi, Feng Jiao, Zi Wen, Kuang Jun, Lin Sen, Wu Wei, etc., all from Meituan platform search and NLP department NLP center.

Recruitment information

Meituan brain Knowledge Map team is recruiting a large number of positions, internship, campus recruitment, social recruitment, coordinate in Beijing/Shanghai, welcome interested students to join us, using natural language and knowledge map technology to help you eat better, better life. Resumes can be sent to: [email protected].

Read more technical articles from meituan’s technical team

Front end | | algorithm back-end | | | data security operations | iOS | Android | test

| in the public bar menu dialog reply goodies for [2020], [2019] special purchases, goodies for [2018], [2017] special purchases such as keywords, to view Meituan technology team calendar year essay collection.

| this paper Meituan produced by the technical team, the copyright ownership Meituan. You are welcome to reprint or use the content of this article for non-commercial purposes such as sharing and communication. Please mark “Content reprinted from Meituan Technical team”. This article shall not be reproduced or used commercially without permission. For any commercial activity, please send an email to [email protected] for authorization.