AliMe KG: Domain Knowledge Graph Construction and Application in E-commerce
Feng-Lin Li, Hehong Chen, Guohai Xu, Tian Qiu, Feng Ji, Ji Zhang, Haiqing Chen
AAliMe KG: Domain Knowledge Graph Construction andApplication in E-commerce
Feng-Lin Li [email protected] Group
Hehong Chen [email protected] Group
Guohai Xu [email protected] Group
Tian Qiu [email protected] Group
Feng Ji [email protected] Group
Ji Zhang [email protected] Group
Haiqing Chen [email protected] Group
ABSTRACT
Pre-sales customer service is of importance to E-commerce plat-forms as it contributes to optimizing customers’ buying process.To better serve users, we propose
AliMe KG , a domain knowledgegraph in E-commerce that captures user problems, points of inter-ests (POI), item information and relations thereof. It helps to under-stand user needs, answer pre-sales questions and generate explana-tion texts. We applied AliMe KG to several online business scenar-ios such as shopping guide, question answering over propertiesand recommendation reason generation, and gained positive re-sults. In the paper, we systematically introduce how we constructdomain knowledge graph from free text, and demonstrate its busi-ness value with several applications. Our experience shows thatmining structured knowledge from free text in vertical domain ispracticable, and can be of substantial value in industrial settings.
CCS CONCEPTS • Computing methodologies → Information extraction ; Se-mantic networks ; •
Information systems → E-commerce in-frastructure . KEYWORDS
E-commerce, Pre-Sales Customer Service, Domain Knowledge Graph
ACM Reference Format:
Feng-Lin Li, Hehong Chen, Guohai Xu, Tian Qiu, Feng Ji, Ji Zhang, and HaiqingChen. 2020. AliMe KG: Domain Knowledge Graph Construction and Appli-cation in E-commerce. In
The 29th ACM International Conference on Infor-mation and Knowledge Management (CIKM ’20), October 19–23, 2020, VirtualEvent, Ireland.
ACM, New York, NY, USA, 8 pages. https://doi.org/10.1145/3340531.3412685
Permission to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor profit or commercial advantage and that copies bear this notice and the full cita-tion on the first page. Copyrights for components of this work owned by others thanACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re-publish, to post on servers or to redistribute to lists, requires prior specific permissionand/or a fee. Request permissions from [email protected].
CIKM ’20, October 19–23, 2020, Virtual Event, Ireland © 2020 Association for Computing Machinery.ACM ISBN 978-1-4503-6859-9/20/10…$15.00https://doi.org/10.1145/3340531.3412685
Understanding customers’ buying process is of key importance toE-commerce platforms as it directly contributes to improving Con-version Rate (CVR). For years, E-commerce sites such as Amazon,Alibaba and Jingdong have been employing search box and catalogas the primary way for customer interaction, teaching customersto search items through product-oriented keywords, and helplingthem to quickly find what they want to buy. To bridge the seman-tic gap between what users want in their mind and how items areorganized in E-commerce platforms, Luo et al. [10] have proposedto capture user needs in product knowledge graph and associatethem with items. Although there have been many efforts to im-prove search engine and recommender system [3, 17–19, 25], thereare still room left for optimizing customers’ buying process.The main reason, in our observation, is that E-commerce sitesusally assume users know exactly what they need and offer searchengine as a tool in support of customer interaction. However, inmany cases, what users have in mind is their problems (e.g., “dryskin”), and they do not have a clear idea about the solution (i.e.,their true needs, e.g., “preserve moisture”), which need to be in-ferred based on domain knowledge. Also, customers often need toseek more information outside E-commerce sites when assessingwhether a product can truly meet their needs or resolve their prob-lems. That is, customers need a service, rather than merely a tool,when making a purchasing decision. The service should be able totalk with customers, infer their needs, provide userful informationabout specific items, give suggestions and reasonable explanations.This is where pre-sales customer service comes in.With the fast development of Deep Learning (DL) techniques,chatbot has become a natural choice for E-commerce customerservice as it is able to automatically answer customers’ questionsand largely improve the efficiency of customer support service. Inlight of this trend, we have launched AliMe [7] for a real-world in-dustrial application that offers pre-sales customer service for hun-dreds of thousands of stores on the Alibaba E-commerce platform .To better understand users, we constructed AliMe KG , a domainknowledge graph in the field of E-commerce that captures user AliMe offers not only pre-sales retail service but also after-sales customer service.We focus on the pre-sales part in this paper. roblems, points of interest (POI), item information, and relationsthereof. The KG serves as the foundation for recognizing user prob-lems (e.g., “dry skin 皮肤干 ”), inferring user needs (e.g., “preservemoisture 保湿 ”), completing item information (e.g., “hyaluronicacid 玻尿酸 ” cause “preserve moisture 保湿 ”), and generating ex-planatory recommendation reasons (e.g., “we recommend this fa-cial cleanser as it contains hyaluronic acid and is able to preservemoisture, which perfectly fits your dry skin problem”).In this paper, we present AliMe KG, introduce our systematicconstruction methodology and semi-automated knowledge min-ing process, and demonstrate its business value through severalrealistic applications in pre-sales conversation scenarios.Our paper makes the following contributions: • We propose AliMe KG, an ongoing domain knowledge graphthat currently supports the top-50 main categories on the Al-ibaba E-commerce platform, and enables our chatbot to un-derstand user problems, infer user needs, answer pre-salesquestions and provide explanation texts. • We present a systematic methodology and semi-automatedprocesses for mining structured knowledge from free texts,and introduce our innovations in the underlying key com-ponents, namely Phrase Mining, Named Entity Recognitionand Relation Extraction. • We applied AliMe KG to several realistic applications anddemonstrated its business value through online A/B tests.The rest of the paper is organized as follows: Sec. 2 introducesour motivation; Sec. 3 presents an overview of AliMe KG; Sec. 4 sys-tematically describes how the KG is constructed; Sec. 5 shows theevaluation results of main building blocks; Sec. 6 demonstrates itsapplication and business value; Sec. 7 reviews related work; Sec. 8discusses future work and concludes the paper.
In general, customers’ buying process includes five stages : needrecognition, information search, evaluation of alternatives, purchasedecision and post-purchase behavior. Over the years, search en-gine and recommender system are the principal means that E-commerceplatforms used to enhance user shopping experience. Customershave been accustomed to finding out solutions to their problemsor needs through general search engines (e.g., Baidu) or verticalwebsites (e.g., Zhihu), and then searching specific items at onlineshopping sites. On observing the semantic gap between user needsand product taxonomy, Luo et al. [10] have proposed to captureuser needs as E-commerce concepts (e.g., “outdoor barbecue”) andthen associate the concepts with product items (e.g., grills and but-ter). In this way, E-commerce search engines are able to under-stand user concepts and accordingly recommend associated items.Desipte the efforts, search engine by itself is insufficient to furtheroptimize customers’ buying process as it is limited to part of thebuying process – namely the information search stage – in cus-tomers’ mental model.Pre-sales customer service is a more natural way for understand-ing users through conversational interaction. Nowadays chatbothas been prevailing in E-commerce customer service, either pre-sales or after-sales, since it is able to reduce large amount of time https://en.wikipedia.org/wiki/Buyer_decision_process spent on customer enquiries. We have already launched AliMe [7]for a real-world industrial application. Although it is able to achievehigh performance through text classification and/or matching overquestion answer (QA) pairs, it is still hard to say our chatbot can“understand” customer questions.To better serve customers, our chatbot need to recognize userproblems, infer user needs, provide useful item information, andgenerate explanation texts based on domain knowledge. Therefore,we propsoed AliMe KG, a domain knowledge graph in E-commerce,and applied it to empower AliMe for a better user experience. We show our core ontology in Fig. 1. Three commonly acceptedconcepts, namely “User”, “Item” and “Scenario”, are adopted fromclassic buying process: a user intent to buy some items at/for acertain scenario. Note that “Scenario” refers to not only shoppingplace (e.g., city, shop) but also various kinds of consumption sce-narios (e.g., Teachers’ Day, outdoor barbecue). The concept “IPV”captures property values of items (Item-Property-Value, e.g., “cleans-ing foam 洁面泡沫 ”- ingredient -“bisabolol 红没药醇 ”). Two newconcepts, Problem and
POI , are our key contributions. “Problem”refers to a problematic state that a user is at (e.g., “pimple 长痘痘 ”),“POI” captures users’ need or solution to user problem (“anti-acne 清痘抑痘 ”). Also, there are two types of newly added links: need ,which relates problem to POI (e.g., “pimple” need “anti-acne”), and cause , which links IPV and POI (e.g., “bisabolol” cause “anti-acne”).These links are established based on domain knowledge.
POI
ItemIntentionUser
Problem
IPVNeed Cause HasHas ScenarioSituation Association
Figure 1: The core ontology of AliMe KG
In general, our KG consists of three layers: User, POI and Item.As shown in Fig 2, the user layer captures users’ problems (e.g.,“pimple 长痘痘 ”). The item layer records items’ properties (e.g.,“ingredient 成分 ”) and property values (e.g., “bisabolol 红没药醇 ”) based on their categories. The POI layer, which serves as thebridge between users and items, on one hand is used to link users’problem (e.g., “pimple” need “anti-acne”), on the other hand is usedto relate items’ property value (e.g., “bisabolol” cause “anti-acne”).That is, the need link can be used to infer users’ need based on theirproblems, and the cause link can be used to retrieve correspondingitems and explain why a product has a feature of concern.There are three points to be noted. First, items in Alibaba (alsoother E-commerce sites) are organized based on Category-Property-Value (CPV): thousands of categories form a hierarchical struc-ture and leaf categories have properties pre-defiend, items are as-sociated with leaf categoris and accordingly instantiate their pre-defined properties. Second, our KG captures domain knowledge atclass level, not instance level. For example, user concern “pimple 长痘痘 ” and product feature “anti-acne 清痘抑痘 ” mentioned above able 1: Types, formats and sources of domain knowledge Type Format Example SourceUser POI Category - Has - POI “clothing” - has_poi - “skin-friendly” item articles, detail pagesUser Probelm User - Has - Problem “user” - has_problem - “pimple” chatlogCPV&IPV Category - Property - Value “cleansing foam” - ingredient - “bisabolol” item articles, detail pagesItem - Property - Value “cleansing foam” - ingredient - “bisabolol”POI Knowledge Problem - Need - POI “pimple” - needs - “anti-acne” chatlogIPV - Cause - POI “bisabolol” - cause - “anti-acne” has_problemUser Pimples (cid:3661)(cid:4850)(cid:4850)
Cleansing Foam (cid:4164)(cid:7566)(cid:4143)(cid:4116)
Anti-Acne (cid:4276)(cid:4850)(cid:3242)(cid:4850)
NeedBisabolol (cid:5441)(cid:4108)(cid:6041)(cid:7144)
IngredientCause
User ItemPOISchema Layer user
Instance Layer
Preference Feature
Figure 2: An excerpt of our domain knowledge graph are class level concepts. When put into use, they will be instanti-ated by specific users and items. Third, dotted links are added orpredicted through KG completion. For instance, if a user concernabout “pimple”, which needs “anti-acne”, we will add a preference link between the user and “anti-acne”.
In this section, we first introduce the types of knowledge in our KG,then describe our methodology about how to acquire knowledge.As shown in Table 1, our KG includes the following types ofknowledge: user problems, POIs, CPV&IPV data, and POI relationalknowledge. Among them, user problems and POIs are captured inphrases. We classify a phrase as a user problem if it describes aproblematic state that users want to find out solutions for (e.g.,“pimple 长痘痘 ”), and as a user POI if it reveals user interests orpotential needs (e.g, “anti-acne 清痘抑痘 ”). CPV and IPV data aremainly imported from the Alibaba product knowledge graph andcomplemented according to our application in pre-sales conversa-tion scenarios. POI knowledge is relational and encodes the asso-ciations of POIs with user problems, and with CPV&IPV data.We present our knowledge mining process in Fig. 3. In gen-eral, it includes two parts: node mining and link prediction . Theprocess takes as input data source, which includes chatlog, itemdetail pages and item articles, firstly extract nodes, then establishlinks, and finally output structured knowledge. During the process,crowdsourcing is employed as the primary way for KG quality in-spection. Each sub-process, including crowdsourcing , has beenautomated in our production environment and is scheduled to runperiodically. We have integrated the Alibaba crowdsourcing platform into our knowledge miningprocess, as such we are able to automatically deploy tasks and recycle labelled data.
The goal of POI mining is to extract potential user interests orneeds. For example, “skin-friendly 亲肤 ”, “anti-acne 清痘抑痘 ”and “safe and non-toxic 安全无毒 ”, are typical POIs in Clothing,Beauty and Tableware. As users are often not aware of or do notexplicitly express their POIs, we choose E-commerce content (itemdetail pages and articles), rather than customer service chatlog, asour data source for POI mining.Given a domain (i.e., first-level categories), we first retrieve theset of leaf categories and associated items from product knowledgegraph, and then obtain detail information and articles for each itemfrom the Alibaba E-commerce content platform. Subsequently, asshown in Fig. 3a, we use heuristic rules and phrase mining to ob-tain phrases, from which we collect both positive and negativesamples through crowdsourcing annotation. After that, we train abinary BERT [4] classifier to predict whether an extracted phraseis a POI. The obtained POI candidates will be further checked bycrowd annotators, and finally those accepted will be added into ourknowledge graph and organized as “Category - has - POI” tuples. Phrase mining is of importance to both POImining and user problem mining (to be detailed in Sec. 4.2). Weadopt the automated phrase mining approach [15], extend it withdeep semantic features for quality phrase classification, and fur-ther employ BERT masked language model (MLM) for pruning.We describe the algorithm flow in Fig. 4. The process takes asinput a corpus from which phrases to be extracted and a lexiconconsists of more than one million words accumulated in practice,and output a set of quality phrases. Specifically, we start collect-ing phrase seeds without human labor: the procedure extracts textsequences that are separated by punctuations and within certainlength as raw phrases, and treats the intersection of frequent rawphrases and lexicon words as phrase seeds. Next, we establish theset of phrase candidates that contains all the n-grams over a certainthreshold (e.g., 3) in the corpus, and label those candidates as posi-tive if they are in the seed pool and negative otherwise. Often, thenumber of positives is rather small (e.g., hundreds) while that ofnegatives is large (e.g, hundreds of thousands). As suggested by Au-toPhrase [15], we train a random forest (RF) classifier by drawing 𝐾 (e.g., 100) phrase candidates with replacement from the positivesand negatives respectively for each base decision tree classifier.For feature design, except statistic features such as frequency,tf-idf, pointwise mutual information and information content , https://en.wikipedia.org/wiki/Pointwise_mutual_information https://en.wikipedia.org/wiki/Information_content (cid:87)(cid:72)(cid:80)(cid:3)(cid:38)(cid:82)(cid:81)(cid:87)(cid:72)(cid:81)(cid:87) (cid:75)(cid:72)(cid:88)(cid:85)(cid:76)(cid:86)(cid:87)(cid:76)(cid:70)(cid:3)(cid:85)(cid:88)(cid:79)(cid:72)(cid:86)(cid:83)(cid:75)(cid:85)(cid:68)(cid:86)(cid:72)(cid:3)(cid:80)(cid:76)(cid:81)(cid:76)(cid:81)(cid:74)(cid:51)(cid:50)(cid:44)(cid:3)(cid:38)(cid:79)(cid:68)(cid:86)(cid:86)(cid:76)(cid:393)(cid:70)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:38)(cid:85)(cid:82)(cid:90)(cid:71)(cid:86)(cid:82)(cid:88)(cid:85)(cid:70)(cid:76)(cid:81)(cid:74) (cid:51)(cid:85)(cid:72)(cid:16)(cid:86)(cid:68)(cid:79)(cid:72)(cid:86)(cid:38)(cid:82)(cid:81)(cid:89)(cid:72)(cid:85)(cid:86)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81) (cid:52)(cid:36)(cid:3)(cid:51)(cid:68)(cid:76)(cid:85)(cid:76)(cid:81)(cid:74) (cid:54)(cid:70)(cid:72)(cid:81)(cid:72)(cid:3)(cid:38)(cid:79)(cid:68)(cid:86)(cid:86)(cid:76)(cid:393)(cid:70)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:44)(cid:51)(cid:57)(cid:3)(cid:48)(cid:76)(cid:81)(cid:76)(cid:81)(cid:74)(cid:51)(cid:82)(cid:79)(cid:68)(cid:85)(cid:76)(cid:87)(cid:92)(cid:3)(cid:45)(cid:88)(cid:71)(cid:74)(cid:72)(cid:80)(cid:72)(cid:81)(cid:87)(cid:44)(cid:51)(cid:57)(cid:3)(cid:49)(cid:82)(cid:85)(cid:80)(cid:68)(cid:79)(cid:76)(cid:93)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:44)(cid:51)(cid:57)(cid:3)(cid:41)(cid:76)(cid:79)(cid:87)(cid:72)(cid:85)(cid:76)(cid:81)(cid:74) (cid:11)(cid:68)(cid:12)(cid:3)(cid:51)(cid:50)(cid:44)(cid:3)(cid:80)(cid:76)(cid:81)(cid:76)(cid:81)(cid:74) (cid:11)(cid:70)(cid:12)(cid:3)(cid:38)(cid:51)(cid:57)(cid:9)(cid:44)(cid:51)(cid:57)(cid:3)(cid:80)(cid:76)(cid:81)(cid:76)(cid:81)(cid:74) (cid:37)(cid:40)(cid:53)(cid:55) (cid:49)(cid:40)(cid:53) (cid:39)(cid:68)(cid:87)(cid:68)(cid:3)(cid:54)(cid:82)(cid:88)(cid:85)(cid:70)(cid:72) Identify POI (cid:15)(cid:3)(cid:51)(cid:57)(cid:15)(cid:3)(cid:56)(cid:86)(cid:72)(cid:85)(cid:3)(cid:51)(cid:85)(cid:82)(cid:69)(cid:79)(cid:72)(cid:80) (cid:38)(cid:85)(cid:82)(cid:90)(cid:71)(cid:86)(cid:82)(cid:88)(cid:85)(cid:70)(cid:76)(cid:81)(cid:74)(cid:53)(cid:72)(cid:79)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:3)(cid:40)(cid:91)(cid:87)(cid:85)(cid:68)(cid:70)(cid:87)(cid:76)(cid:82)(cid:81) (cid:11)(cid:71)(cid:12)(cid:3)(cid:51)(cid:50)(cid:44)(cid:3)(cid:78)(cid:81)(cid:82)(cid:90)(cid:79)(cid:72)(cid:71)(cid:74)(cid:72)(cid:3)(cid:80)(cid:76)(cid:81)(cid:76)(cid:81)(cid:74)(cid:51)(cid:50)(cid:44)(cid:86) (cid:51)(cid:85)(cid:72)(cid:16)(cid:86)(cid:68)(cid:79)(cid:72)(cid:86)(cid:38)(cid:82)(cid:81)(cid:89)(cid:72)(cid:85)(cid:86)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81) (cid:75)(cid:72)(cid:88)(cid:85)(cid:76)(cid:86)(cid:87)(cid:76)(cid:70)(cid:3)(cid:85)(cid:88)(cid:79)(cid:72)(cid:86)(cid:83)(cid:75)(cid:85)(cid:68)(cid:86)(cid:72)(cid:3)(cid:80)(cid:76)(cid:81)(cid:76)(cid:81)(cid:74)(cid:51)(cid:85)(cid:82)(cid:69)(cid:79)(cid:72)(cid:80)(cid:3)(cid:38)(cid:79)(cid:68)(cid:86)(cid:86)(cid:76)(cid:393)(cid:70)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:38)(cid:85)(cid:82)(cid:90)(cid:71)(cid:86)(cid:82)(cid:88)(cid:85)(cid:70)(cid:76)(cid:81)(cid:74) (cid:11)(cid:69)(cid:12)(cid:3)(cid:56)(cid:86)(cid:72)(cid:85)(cid:3)(cid:83)(cid:85)(cid:82)(cid:69)(cid:79)(cid:72)(cid:80)(cid:3)(cid:80)(cid:76)(cid:81)(cid:76)(cid:81)(cid:74) (cid:37)(cid:40)(cid:53)(cid:55) (cid:51)(cid:85)(cid:82)(cid:69)(cid:79)(cid:72)(cid:80)(cid:86) (cid:49)(cid:82)(cid:71)(cid:72)(cid:3)(cid:48)(cid:76)(cid:81)(cid:76)(cid:81)(cid:74)(cid:47)(cid:76)(cid:81)(cid:78)(cid:3)(cid:51)(cid:85)(cid:72)(cid:71)(cid:76)(cid:70)(cid:87)(cid:76)(cid:82)(cid:81) (cid:44)(cid:51)(cid:57)(cid:86)(cid:51)(cid:50)(cid:44)(cid:3)(cid:53)(cid:72)(cid:79)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:68)(cid:79)(cid:3)(cid:46)(cid:81)(cid:82)(cid:90)(cid:79)(cid:72)(cid:71)(cid:74)(cid:72)
Figure 3: Knowledge mining process (cid:39)(cid:76)(cid:70)(cid:87)(cid:76)(cid:82)(cid:81)(cid:68)(cid:85)(cid:92) (cid:51)(cid:75)(cid:85)(cid:68)(cid:86)(cid:72)(cid:3)(cid:54)(cid:72)(cid:72)(cid:71)(cid:86)(cid:51)(cid:75)(cid:85)(cid:68)(cid:86)(cid:72)(cid:3)(cid:38)(cid:68)(cid:81)(cid:71)(cid:76)(cid:71)(cid:68)(cid:87)(cid:72) (cid:11)(cid:41)(cid:85)(cid:72)(cid:84)(cid:88)(cid:72)(cid:81)(cid:87)(cid:3)(cid:81)(cid:16)(cid:74)(cid:85)(cid:68)(cid:80)(cid:86)(cid:12)
Training Dataset (cid:53)(cid:68)(cid:81)(cid:71)(cid:82)(cid:80)(cid:3)(cid:41)(cid:82)(cid:85)(cid:72)(cid:86)(cid:87)(cid:3)(cid:38)(cid:79)(cid:68)(cid:86)(cid:86)(cid:76)(cid:393)(cid:72)(cid:85)(cid:51)(cid:75)(cid:85)(cid:68)(cid:86)(cid:68)(cid:79)(cid:3)(cid:54)(cid:72)(cid:74)(cid:80)(cid:72)(cid:81)(cid:87)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:41)(cid:72)(cid:68)(cid:87)(cid:88)(cid:85)(cid:72)(cid:3)(cid:39)(cid:72)(cid:86)(cid:76)(cid:74)(cid:81)(cid:58)(cid:76)(cid:71)(cid:72)(cid:29)(cid:3)(cid:86)(cid:87)(cid:68)(cid:87)(cid:76)(cid:86)(cid:87)(cid:76)(cid:70)(cid:3)(cid:73)(cid:72)(cid:68)(cid:87)(cid:88)(cid:85)(cid:72)(cid:39)(cid:72)(cid:72)(cid:83)(cid:29)(cid:3)(cid:86)(cid:72)(cid:80)(cid:68)(cid:81)(cid:87)(cid:76)(cid:70)(cid:3)(cid:73)(cid:72)(cid:68)(cid:87)(cid:88)(cid:85)(cid:72) (cid:37)(cid:40)(cid:53)(cid:55)(cid:3)(cid:48)(cid:47)(cid:48) (cid:20)(cid:17)(cid:3)(cid:38)(cid:82)(cid:79)(cid:79)(cid:72)(cid:70)(cid:87)(cid:3)(cid:83)(cid:75)(cid:85)(cid:68)(cid:86)(cid:72)(cid:3)(cid:86)(cid:72)(cid:72)(cid:71)(cid:86)(cid:21)(cid:17)(cid:3)(cid:51)(cid:75)(cid:85)(cid:68)(cid:86)(cid:72)(cid:3)(cid:80)(cid:76)(cid:81)(cid:76)(cid:81)(cid:74) (cid:22)(cid:17)(cid:3)(cid:51)(cid:75)(cid:85)(cid:68)(cid:86)(cid:72)(cid:3)(cid:83)(cid:85)(cid:88)(cid:81)(cid:76)(cid:81)(cid:74) (cid:43)(cid:72)(cid:88)(cid:85)(cid:76)(cid:86)(cid:87)(cid:76)(cid:70)(cid:3)(cid:85)(cid:88)(cid:79)(cid:72)(cid:86) (cid:38)(cid:82)(cid:85)(cid:83)(cid:88)(cid:86) (cid:44)(cid:81)(cid:87)(cid:72)(cid:85)(cid:86)(cid:72)(cid:70)(cid:87)(cid:76)(cid:82)(cid:81) limited positives
Figure 4: Overview of phrase mining model we also consider semantic features that average character embed-dings. As some of the n-grams are invalid phrases and the textsequences that they located in need to be properly segmented, wefurther use POS-guided phrasal segmentation [15] to rectify thefrequency of phrase candidates. At last, we adopt BERT MLM to fil-ter out those incomplete ones. Specifically, we mask the first (resp.last) token of a phrase candidate, and feed it into BERT to check ifthe masked token is in the top- 𝑁 (e.g., 1) predicted token list. Once we have collected training samples,we train a binary BERT classifier for subsequent POI mining. Specif-ically, we continually train BERT on E-commerce corpus, and con-struct samples by concatenating a phrase 𝑝 and its leaf category 𝑙𝑐 in the form of “[CLS] 𝑝 [SEP] 𝑙𝑐 [SEP]” for both training and predic-tion, where “[CLS]” and “[SEP]” are conventional tokens adoptedfrom BERT, and leaf category 𝑙𝑐 is treated as the context of 𝑝 . We treat user problems as problematic states that users are in andextract them from pre-sales conversation between customers andservice staff/chatbot. As shown in Fig. 3b, which is similar to POImining, we extract candidate phrases through heuristic rules and phrase mining, collect training samples through crowdsourcing,and train a binary BERT classifier for user problems classification.One difference, however, is that training and serving samples arein the form of “[CLS] 𝑝 [SEP] 𝑑 𝑝 [SEP]”, where 𝑝 indicates a phrase,and 𝑑 𝑝 refers to one of the sentences that 𝑝 is located in. Notethat user problems in our KG are captured as conceptual knowl-edge and associated with the “User” concept, and are employed forproblem-recognition in online conversations for particular users. Although we are able to import all items from the Alibaba knowl-edge graph, CPV&IPV data can be missing. Therefore, we also needto complement missing properties and property values of concern.For simplicity, we fix the properties and enumerable property val-ues of each leaf category, and focus on mining missing propertyvalues. For example, the color of T-shirt can be red, blue, white andso on. Supposing that the color of a particular T-shirt is missing, wetry to fill the gap with one of the enumerated values through IPVmining. If the missing value of a property is not enumerable, weonly accept it if it exceeds a certain frequency in our data source.The IPV mining process is shown in Fig. 3c. We start with pre-sales conversations in a specific domain, collect QA pairs aboutitems, filter out pairs irrelevant to business scenarios, then extractproperty values through NER. After that, we judge the polarityto check whether an extracted value belongs to a particular item,normalize property value based on synonyms, and at last use pre-defined CPV dictionary for quality control.
In our case, NER is usedto identify property values from a text sequence and then assignthem with property labels. For example, in the question “Can itbe used by pregnant women? 孕妇能用吗 ”, “pregnant women 孕妇 ” is the value of property “target users 适用人群 ”. Understand-ing text from such property-value perspective is foundational inE-commerce as it can be used not only for offline knowledge min-ing but also for online question answering over properties.Adopting the unified Embedder-Encoder-Decoder framework [8],we extend the standard BiLSTM-CRF model [5] with BERT. In ourroposed model (Fig. 5), BERT extracts features and provides con-textual embeddings for input tokens, BiLSTM acts as an encoder,and CRF predicts their final labels. Besides, we incorporate externallexicon (words only) and dictionary (words with type, e.g., wordscan be property values and type can be properties in our case)knowledge through introducing extra features. BERTx x x x x x (cid:2560) (cid:2443) (cid:5766) (cid:4768) (cid:1920) (cid:1176) Pregnant women can use it (cid:34) z e z e z e z e z e z e BiLSTMB-Users I-Users O O O OCRF DecoderBiLSTMEncoderExtra FeatureToken FeatureBERT Embedder
Figure 5: Overview of BERT-BiLSTM-CRF NER model
Given an input sequence of tokens { 𝑥 , 𝑥 , 𝑥 , ..., 𝑥 𝑚 } , we obtainthe representation of 𝑥 𝑖 ( ≤ 𝑖 ≤ 𝑚 ) through concatenating contex-tual feature 𝒛 𝒊 from BERT and extra feature 𝒆 𝒊 constructed basedon lexicon/dictionary, as formulated in Equation 1. 𝒙 𝒊 = 𝒛 𝒊 ⊕ 𝒆 𝒊 (1)For extra features, we first employ a word segmenter with ex-ternal lexicon to segment each input sequence and accordingly ob-tain a sequence of softwords [13, 24], and then use BMES scheme,rather than BIO, to represent the positional information of each to-ken in a softword. For example, for the token “pregnant 孕 ” in thesoftword “pregnant women 孕妇 ”, we give it a label “B” that repre-sent the begining of a softword and assign it a 𝑑 -dimentional (e.g.,256) randomly initialized embedding. For external dict, we also addword type information to each token in a softword. For instance,if we know “pregnant women 孕妇 ” is of type “target users”, thenwe will label the token “pregnant 孕 ” as “B Polarity judgement is a must in IPV min-ing becasue negative polarity would introduce incorrect info. Forexample, in the sentence “The T-shrt is not red”, the extracted value“red” is not the color of the particular T-shirt. We currently useheuristic rules for polarity judgement, and will change to DL mod-els in the near future.
To support user need inference and explain while an item has a fea-ture of concern, we need to relate POI with user problems and IPVs,respectively. The key idea of link prediction between two conceptsis to see whether we are able to find one or more text sentences thatreveal a specific kind of relation between them. We summarize POI relational knowledge mining in Fig.3d. Tak-ing the cause relation for example, we first collect text sentencesfrom E-commerce content, and then identify CPVs (resp. POIs) foreach sentence though NER (resp. dictionary match). We keep thosesentences that include both CPV (e.g., “bisabolol 红没药醇 ”) andPOI (e.g., “anti-acne 清痘抑痘 ”), and let crowd annotators checkwhether a sentence reveal the desired relation (e.g., cause ) betweenthem. At last, we recycle labeled data, and train an classificationmodel for subsequent link prediction. The mining of need triples issimilar, but with the data source extended to Baike corpus.Note that links are established between POIs and CPVs, andinstantiated to IPVs through inheritance when put into use. Forexample, conceptually we have “bisabolol 红没药醇 ” cause “anti-acne 清痘抑痘 ”, where “bisabolol” is the value of property “ingre-dient” of a specific category, if an item at that category has “‘bis-abolol” as its ingredient, the item would also possess that feature. Relation extraction is used to predictwhether two concepts form a specific relation in a certain context(i.e., the sentence they are located in). As in [20], we base our modelon BERT and incorporate information from anchor concepts. For asentence 𝑠 with two anchor concepts 𝑐 and 𝑐 , we insert a specialtoken ‘$’ (resp. ‘ 𝑐 (resp. 𝑐 ) in theinput sentence, and use the “CLS” token embedding together withconcept embeddings for final classification (see Fig. 6). (cid:37)(cid:40)(cid:53)(cid:55) (cid:62)(cid:38)(cid:47)(cid:54)(cid:64) (cid:4053) (cid:5459) (cid:7566) (cid:3535) (cid:1157) (cid:1461) (cid:6415) (cid:3661) (cid:2920) (cid:1742) (cid:7) (cid:7) (cid:6)(cid:6) (cid:4569) (cid:4457) (cid:2920) (cid:3028) (cid:7771)(cid:3664) (cid:19) (cid:20) (cid:21) (cid:22) (cid:23) (cid:24) (cid:20)(cid:20) (cid:20)(cid:21)(cid:25) (cid:26) (cid:27) (cid:28) (cid:20)(cid:19)(cid:20)(cid:22) (cid:20)(cid:23) (cid:20)(cid:24) (cid:20)(cid:25) (cid:20)(cid:26) (cid:20)(cid:27) (cid:20)(cid:28) (cid:21)(cid:19)(cid:19) (cid:20) (cid:21) (cid:22) (cid:23) (cid:24) (cid:25) (cid:26) (cid:27) (cid:28) (cid:20)(cid:19)(cid:25) (cid:26) (cid:27) (cid:28) (cid:20)(cid:19) (cid:20)(cid:20) (cid:20)(cid:21) (cid:20)(cid:22) (cid:20)(cid:23) (cid:20)(cid:24) (cid:62)(cid:38)(cid:47)(cid:54)(cid:64) (cid:4053) (cid:5459) (cid:7566) (cid:3535) (cid:4457) (cid:2920) (cid:3028)(cid:7) (cid:4569) (cid:7771) (cid:1157) (cid:1461) (cid:6415) (cid:3661) (cid:2920) (cid:1742)(cid:7) (cid:6)(cid:6)(cid:3664) (cid:19) (cid:21) (cid:22) (cid:23) (cid:24) (cid:26) (cid:27) (cid:28)(cid:20) (cid:25) (cid:20)(cid:19) (cid:26) (cid:27) (cid:28)(cid:25) (cid:20)(cid:20)(cid:20)(cid:19) (cid:20)(cid:21) (cid:36) (cid:36) (cid:36) (cid:36) (cid:36) (cid:36) (cid:36) (cid:36)(cid:36) (cid:36) (cid:36) (cid:36) (cid:36) (cid:36)(cid:36) (cid:36)(cid:36) (cid:36) (cid:36) (cid:36) (cid:36) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14)(cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:14) (cid:69)(cid:79)(cid:68)(cid:70)(cid:78)(cid:29)(cid:3)(cid:75)(cid:68)(cid:85)(cid:71)(cid:16)(cid:83)(cid:82)(cid:86)(cid:76)(cid:87)(cid:76)(cid:82)(cid:81)(cid:3)(cid:76)(cid:81)(cid:71)(cid:72)(cid:91)(cid:85)(cid:72)(cid:71)(cid:29)(cid:3)(cid:86)(cid:82)(cid:73)(cid:87)(cid:16)(cid:83)(cid:82)(cid:86)(cid:76)(cid:87)(cid:76)(cid:82)(cid:81)(cid:3)(cid:76)(cid:81)(cid:71)(cid:72)(cid:91)(cid:54)(cid:72)(cid:81)(cid:87)(cid:72)(cid:81)(cid:70)(cid:72)(cid:3)(cid:55)(cid:85)(cid:72)(cid:72)(cid:55)(cid:82)(cid:78)(cid:72)(cid:81)(cid:3)(cid:72)(cid:80)(cid:69)(cid:72)(cid:71)(cid:71)(cid:76)(cid:81)(cid:74)(cid:54)(cid:82)(cid:73)(cid:87)(cid:16)(cid:83)(cid:82)(cid:86)(cid:76)(cid:87)(cid:76)(cid:82)(cid:81)(cid:3)(cid:72)(cid:80)(cid:69)(cid:72)(cid:71)(cid:71)(cid:76)(cid:81)(cid:74)(cid:54)(cid:72)(cid:74)(cid:80)(cid:72)(cid:81)(cid:87)(cid:3)(cid:72)(cid:80)(cid:69)(cid:72)(cid:71)(cid:71)(cid:76)(cid:81)(cid:74) (cid:36)(cid:89)(cid:72)(cid:85)(cid:68)(cid:74)(cid:72) (cid:36)(cid:89)(cid:72)(cid:85)(cid:68)(cid:74)(cid:72)(cid:41)(cid:38)(cid:14)(cid:36)(cid:70)(cid:87)(cid:76)(cid:89)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81) (cid:41)(cid:38)(cid:14)(cid:36)(cid:70)(cid:87)(cid:76)(cid:89)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81) (cid:41)(cid:38)(cid:14)(cid:36)(cid:70)(cid:87)(cid:76)(cid:89)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81) (cid:335) (cid:335) (cid:20)(cid:22) (cid:20)(cid:23) (cid:20)(cid:24) (cid:54)(cid:70)(cid:82)(cid:85)(cid:72)(cid:41)(cid:38)(cid:14)(cid:54)(cid:82)(cid:73)(cid:87)(cid:80)(cid:68)(cid:91)(cid:53)(cid:72)(cid:79)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81) (cid:55)(cid:68)(cid:76)(cid:79)(cid:43)(cid:72)(cid:68)(cid:71) Spandex fabric is prized for its elasticity (cid:15)(cid:3)(cid:68)(cid:81)(cid:71)(cid:3)(cid:80)(cid:68)(cid:78)(cid:72)(cid:3)(cid:70)(cid:79)(cid:82)(cid:87)(cid:75)(cid:76)(cid:81)(cid:74)(cid:3)(cid:72)(cid:79)(cid:68)(cid:86)(cid:87)(cid:76)(cid:70)
Figure 6: Overview of relation extraction model
We also borrow the idea of triple knowledge injection from K-BERT [9] to incorporate external knowledge (e.g., HOWNET [14],CNDBpedia [22]). Specifically, we use anchor concepts 𝑐 and 𝑐 toquery triples 𝐸 that with them as heads from a knoweldge graph K ,and inject the triples into the sentence to form a new sentence 𝑠 𝑛𝑒𝑤 through soft position. As in Fig 6, there are two concepts “spandexfabric 氨纶面料 ” and “elastic 有弹力 ” in the given text. The firstconcept is associated with a triple “spandex fabric 氨纶面料 ” - characteristic - “high elasticity 弹性高 ”, which is injected into theoriginal sentence through soft position (red indices). We formualtethis process as Equation 2. = 𝑞𝑢𝑒𝑟𝑦 ({ 𝑐 , 𝑐 } , K) 𝑠 𝑛𝑒𝑤 = 𝑖𝑛 𝑗𝑒𝑐𝑡 ( 𝑠, 𝐸 ) (2) We use an illustrative example to demon-starte our POI knowledge mining process. As shown in Fig. 7, giventhe sentence “Food grade silicone does not contain BPA, is resistantto high temperature sterilization, hence it is a truly reassuring andsafe tableware”, we first identify “Food grade silicone” as a typeof “material” in the Baby Tableware category, and recognize “hightemperature sterilization”, “reassuring” and “safe” as POIs. Afterthat, we establish a 𝑐𝑎𝑢𝑠𝑒 link between the CPV and each POI, cap-ture them as triples as shown in the bottom right corner of Fig. 7 (cid:7651)(cid:2013)(cid:5446)(cid:5073)(cid:5762)(cid:3693)(cid:6710)(cid:1255)(cid:1932)
BPA (cid:1157)(cid:1890)(cid:1382)(cid:7771)(cid:4293)(cid:4222)(cid:4016)(cid:1157)(cid:3598)(cid:4965)(cid:3979)(cid:1381)(cid:2444)(cid:2444)(cid:3505)(cid:2974)(cid:4913)(cid:2590)(cid:1602)(cid:7655)(cid:1613)(cid:41)(cid:82)(cid:82)(cid:71)(cid:3)(cid:74)(cid:85)(cid:68)(cid:71)(cid:72)(cid:3)(cid:86)(cid:76)(cid:79)(cid:76)(cid:70)(cid:82)(cid:81)(cid:72)(cid:3)(cid:71)(cid:82)(cid:72)(cid:86)(cid:3)(cid:81)(cid:82)(cid:87)(cid:3)(cid:70)(cid:82)(cid:81)(cid:87)(cid:68)(cid:76)(cid:81)(cid:3)(cid:37)(cid:51)(cid:36)(cid:15)(cid:3)(cid:76)(cid:86)(cid:3)(cid:85)(cid:72)(cid:86)(cid:76)(cid:86)(cid:87)(cid:68)(cid:81)(cid:87)(cid:3)(cid:87)(cid:82)(cid:3)(cid:75)(cid:76)(cid:74)(cid:75)(cid:3)(cid:87)(cid:72)(cid:80)(cid:83)(cid:72)(cid:85)(cid:68)(cid:87)(cid:88)(cid:85)(cid:72)(cid:3)(cid:86)(cid:87)(cid:72)(cid:85)(cid:76)(cid:79)(cid:76)(cid:93)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:15)(cid:3)(cid:75)(cid:72)(cid:81)(cid:70)(cid:72)(cid:3)(cid:76)(cid:87)(cid:3)(cid:76)(cid:86)(cid:3)(cid:68)(cid:3)(cid:87)(cid:85)(cid:88)(cid:79)(cid:92)(cid:3)(cid:85)(cid:72)(cid:68)(cid:86)(cid:86)(cid:88)(cid:85)(cid:76)(cid:81)(cid:74)(cid:3)(cid:68)(cid:81)(cid:71)(cid:3)(cid:86)(cid:68)(cid:73)(cid:72)(cid:3)(cid:87)(cid:68)(cid:69)(cid:79)(cid:72)(cid:90)(cid:68)(cid:85)(cid:72)(cid:7651)(cid:2013)(cid:5446)(cid:5073)(cid:5762)(cid:3693)(cid:6710)(cid:1255)(cid:1932)
BPA (cid:1157)(cid:1890)(cid:1382)(cid:7771)(cid:4293)(cid:4222)(cid:4016)(cid:1157)(cid:3598)(cid:4965)(cid:3979)(cid:1381)(cid:2444)(cid:2444)(cid:3505)(cid:2974)(cid:4913)(cid:2590)(cid:1602)(cid:7655)(cid:1613)(cid:41)(cid:82)(cid:82)(cid:71)(cid:3)(cid:74)(cid:85)(cid:68)(cid:71)(cid:72)(cid:3)(cid:86)(cid:76)(cid:79)(cid:76)(cid:70)(cid:82)(cid:81)(cid:72)(cid:3)(cid:71)(cid:82)(cid:72)(cid:86)(cid:3)(cid:81)(cid:82)(cid:87)(cid:3)(cid:70)(cid:82)(cid:81)(cid:87)(cid:68)(cid:76)(cid:81)(cid:3)(cid:37)(cid:51)(cid:36)(cid:15)(cid:3)(cid:76)(cid:86)(cid:3)(cid:85)(cid:72)(cid:86)(cid:76)(cid:86)(cid:87)(cid:68)(cid:81)(cid:87)(cid:3)(cid:87)(cid:82)(cid:3)(cid:75)(cid:76)(cid:74)(cid:75)(cid:3)(cid:87)(cid:72)(cid:80)(cid:83)(cid:72)(cid:85)(cid:68)(cid:87)(cid:88)(cid:85)(cid:72)(cid:3)(cid:86)(cid:87)(cid:72)(cid:85)(cid:76)(cid:79)(cid:76)(cid:93)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:15)(cid:3)(cid:75)(cid:72)(cid:81)(cid:70)(cid:72)(cid:3)(cid:76)(cid:87)(cid:3)(cid:76)(cid:86)(cid:3)(cid:68)(cid:3)(cid:87)(cid:85)(cid:88)(cid:79)(cid:92)(cid:3)(cid:85)(cid:72)(cid:68)(cid:86)(cid:86)(cid:88)(cid:85)(cid:76)(cid:81)(cid:74)(cid:3)(cid:68)(cid:81)(cid:71)(cid:3)(cid:86)(cid:68)(cid:73)(cid:72)(cid:3)(cid:87)(cid:68)(cid:69)(cid:79)(cid:72)(cid:90)(cid:68)(cid:85)(cid:72)(cid:53)(cid:72)(cid:79)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:40)(cid:91)(cid:87)(cid:85)(cid:68)(cid:70)(cid:87)(cid:76)(cid:82)(cid:81) (cid:47)(cid:72)(cid:68)(cid:73)(cid:3)(cid:70)(cid:68)(cid:87)(cid:72)(cid:74)(cid:82)(cid:85)(cid:92)(cid:29)(cid:3)(cid:37)(cid:68)(cid:69)(cid:92)(cid:3)(cid:55)(cid:68)(cid:69)(cid:79)(cid:72)(cid:90)(cid:68)(cid:85)(cid:72)(cid:3)(cid:11)(cid:2602)(cid:2602)(cid:7655)(cid:1613)(cid:12)(cid:55)(cid:72)(cid:91)(cid:87)
Category Property Value (cid:7651)(cid:2013)(cid:5446)(cid:5073)(cid:5762)(cid:3693)(cid:6710)(cid:1255)(cid:1932)
BPA (cid:1157)(cid:1890)(cid:1382)(cid:7771)(cid:4293)(cid:4222)(cid:4016)(cid:1157)(cid:3598)(cid:4965)(cid:3979)(cid:1381)(cid:2444)(cid:2444)(cid:3505)(cid:2974)(cid:4913)(cid:2590)(cid:1602)(cid:7655)(cid:1613)(cid:41)(cid:82)(cid:82)(cid:71)(cid:3)(cid:74)(cid:85)(cid:68)(cid:71)(cid:72)(cid:3)(cid:86)(cid:76)(cid:79)(cid:76)(cid:70)(cid:82)(cid:81)(cid:72)(cid:3)(cid:71)(cid:82)(cid:72)(cid:86)(cid:3)(cid:81)(cid:82)(cid:87)(cid:3)(cid:70)(cid:82)(cid:81)(cid:87)(cid:68)(cid:76)(cid:81)(cid:3)(cid:37)(cid:51)(cid:36)(cid:15)(cid:3)(cid:76)(cid:86)(cid:3)(cid:85)(cid:72)(cid:86)(cid:76)(cid:86)(cid:87)(cid:68)(cid:81)(cid:87)(cid:3)(cid:87)(cid:82)(cid:3)(cid:75)(cid:76)(cid:74)(cid:75)(cid:3)(cid:87)(cid:72)(cid:80)(cid:83)(cid:72)(cid:85)(cid:68)(cid:87)(cid:88)(cid:85)(cid:72)(cid:3)(cid:86)(cid:87)(cid:72)(cid:85)(cid:76)(cid:79)(cid:76)(cid:93)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:15)(cid:3)(cid:75)(cid:72)(cid:81)(cid:70)(cid:72)(cid:3)(cid:76)(cid:87)(cid:3)(cid:76)(cid:86)(cid:3)(cid:68)(cid:3)(cid:87)(cid:85)(cid:88)(cid:79)(cid:92)(cid:3)(cid:85)(cid:72)(cid:68)(cid:86)(cid:86)(cid:88)(cid:85)(cid:76)(cid:81)(cid:74)(cid:3)(cid:68)(cid:81)(cid:71)(cid:3)(cid:86)(cid:68)(cid:73)(cid:72)(cid:3)(cid:87)(cid:68)(cid:69)(cid:79)(cid:72)(cid:90)(cid:68)(cid:85)(cid:72)(cid:49)(cid:40)(cid:53) (cid:48)(cid:68)(cid:87)(cid:72)(cid:85)(cid:76)(cid:68)(cid:79) (cid:51)(cid:50)(cid:44)
Head Property Tail (cid:41)(cid:82)(cid:82)(cid:71)(cid:3)(cid:74)(cid:85)(cid:68)(cid:71)(cid:72)(cid:3)(cid:86)(cid:76)(cid:79)(cid:76)(cid:70)(cid:82)(cid:81) (cid:70)(cid:68)(cid:88)(cid:86)(cid:72) (cid:54)(cid:68)(cid:73)(cid:72)(cid:87)(cid:92)(cid:37)(cid:68)(cid:69)(cid:92)(cid:3)(cid:87)(cid:68)(cid:69)(cid:79)(cid:72)(cid:90)(cid:68)(cid:85)(cid:72) (cid:48)(cid:68)(cid:87)(cid:72)(cid:85)(cid:76)(cid:68)(cid:79) (cid:41)(cid:82)(cid:82)(cid:71)(cid:3)(cid:74)(cid:85)(cid:68)(cid:71)(cid:72)(cid:3)(cid:86)(cid:76)(cid:79)(cid:76)(cid:70)(cid:82)(cid:81) (cid:38)(cid:51)(cid:57)(cid:3)(cid:11)(cid:38)(cid:68)(cid:87)(cid:72)(cid:74)(cid:82)(cid:85)(cid:92)(cid:3)(cid:47)(cid:72)(cid:89)(cid:72)(cid:79)(cid:12)(cid:3)
Item Property Value (cid:44)(cid:87)(cid:72)(cid:80)(cid:6)(cid:20) (cid:48)(cid:68)(cid:87)(cid:72)(cid:85)(cid:76)(cid:68)(cid:79) (cid:41)(cid:82)(cid:82)(cid:71)(cid:3)(cid:74)(cid:85)(cid:68)(cid:71)(cid:72)(cid:3)(cid:86)(cid:76)(cid:79)(cid:76)(cid:70)(cid:82)(cid:81) (cid:44)(cid:51)(cid:57)(cid:3)(cid:11)(cid:44)(cid:87)(cid:72)(cid:80)(cid:3)(cid:47)(cid:72)(cid:89)(cid:72)(cid:79)(cid:12) (cid:44)(cid:87)(cid:72)(cid:80)(cid:6)(cid:21) (cid:48)(cid:68)(cid:87)(cid:72)(cid:85)(cid:76)(cid:68)(cid:79) (cid:41)(cid:82)(cid:82)(cid:71)(cid:3)(cid:74)(cid:85)(cid:68)(cid:71)(cid:72)(cid:3)(cid:86)(cid:76)(cid:79)(cid:76)(cid:70)(cid:82)(cid:81)(cid:46)(cid:42)(cid:3)(cid:55)(cid:85)(cid:76)(cid:83)(cid:79)(cid:72) (cid:51)(cid:50)(cid:44) (cid:51)(cid:50)(cid:44)(cid:41)(cid:82)(cid:82)(cid:71)(cid:3)(cid:74)(cid:85)(cid:68)(cid:71)(cid:72)(cid:3)(cid:86)(cid:76)(cid:79)(cid:76)(cid:70)(cid:82)(cid:81) (cid:70)(cid:68)(cid:88)(cid:86)(cid:72) (cid:3)(cid:335)(cid:54)(cid:87)(cid:72)(cid:85)(cid:76)(cid:79)(cid:76)(cid:93)(cid:68)(cid:87)(cid:76)(cid:82)(cid:81)(cid:41)(cid:82)(cid:82)(cid:71)(cid:3)(cid:74)(cid:85)(cid:68)(cid:71)(cid:72)(cid:3)(cid:86)(cid:76)(cid:79)(cid:76)(cid:70)(cid:82)(cid:81) (cid:70)(cid:68)(cid:88)(cid:86)(cid:72) (cid:53)(cid:72)(cid:68)(cid:86)(cid:86)(cid:88)(cid:85)(cid:76)(cid:81)(cid:74)
Figure 7: Example of POI relational knowledge mining
In this section, we first provide a statistical overview of our KG, andthen present detailed evaluations for our key components. Notethat we use precision 𝑝 to evaluate phrase mining (whether an ex-tracted text sequence is a meaningful/quality phrase), 𝐹 to mea-sure NER, and 𝐴𝑈𝐶 to measure binary classification tasks.
We show the statistics of AliMe KG in Table 2. So far we have ac-cumulated 365K + POIs, 1K + user prblems and 29K + CPVs, 8.6K + “User problem - POI” triples, and 113K + C-PV-POI triples. We con-strcuted an active item pool consists of 600K + items, and collected3,300K + IPV triples, 43.95% of which are acquired through our IPVmining. Further, by applying C-PV-POI triples to items, we ob-tained more than 13,590K + I-PV-POI triples via inheritance.
Conceptual level knowledge, including POIs,user problems, CPVs, the assocaitions of POIs with user problems,and with CPVs, has been completely checked by crowdsourcing,
Table 2: The statistics of AliMe KG
Layer Knowledge Numbers Source Qulity ControlUser Layer User problem 1K + Mining (100%) CompleteItem Layer CPV 29K + Import (100%) CompleteItem 600K + Import (100%) -IPV 3,300K + Import (56.05%) SpotMining (43.95%)POI Layer POI 365K + Mining (100%) CompleteUser problem - POI 8.6K + Mining (100%) CompleteC - PV - POI 113K + Mining (100%) CompleteI - PV - POI 13,590K + Mining (100%) Spot hence its quality can be ensured. Instance level knowledge, IPVdata and I-PV-POI triples, was only spot-checked because of itshuge amount. Our spot-check also shows the accuracy of IPV andI-PV-POI triples is of high quality and can be directly used in real-world applications.
We evaluated our phrase mining method withan E-commerce corpus consists of 830K + text sentences. We ran-domly selected 200 phrase candidates obtained from each methodand assess whether a candidate is a quality phrase through crowd-sourcing. We show the results in Table 3, where “Unsupervised”model represents the one use left&right entropy and PMI for phrasecandidate scoring, “RF” indicates random forest, “SEG” means POS-guided phrasal segmentation. One can see that the phrasal seg-mentation and BERT MLM are very helpful in improving the preci-sion of phrase candidates. When combined together, our model isable to achieve a precision of 88% without any manual annotation,which is practically usable in industrial settings. Table 3: Experimental results of phrase mining
Model 𝑝 Unsupervised 28716 58.7%RF 35566 48.5%RF+SEG 13516 70%RF+MLM 12984 80%RF+SEG+MLM 7514 88%Our model is able to support million-scale corpus. Moreover,it is domain independent and does not need human annotation,which are favored by the industry. Besides, as we can also seethat the number of extracted quality phrases decreases as precisionincreases, we will further improve our model to achieve a bettertrade-off.
We evaluated our NER model inthe Beauty domain with 6 property types. We show the results inTable 4. We can see that the incorporation of lexicon knowledge isable to achieve an increase of 1.56% in 𝐹 , and that of dict knowl-edge brings an increase of 3.14%. We assessed our relation extraction modelon 𝑐𝑎𝑢𝑠𝑒 link prediction in the Clothing domain. As in Table 5, al-though baseline model is able to achive high performance on thisbinary classification task (
𝐴𝑈𝐶 = . ), the incorporation of able 4: Experimental results of named entity recognition Model 𝐹 BERT+BiLSTM+CRF 78.17%+Lexicon 79.73%+Dict 81.31%external knowledge still further improves the result: an increaseof . for HOWNET and . for CNDBpedia in 𝐴𝑈𝐶 . Table 5: Experimental results of relation extraction
Model AUCBERT + concept info (Baseline) 95.12%+HOWNET 95.40%+CNDBpedia 95.36%
AliMe KG has been applied to several real-world applications inpre-sales conversation scenarios, including but not limited to shop-ping guide, question answering over properties, and explanatoryrecommendation reason generation. We show how AliMe KG con-tributes to optimizing customers’ buying process in Table 6, andintroduce the applications in the following sections.
Table 6: How AliMe KG optimizes customer buying process
Ends Means ApplicationNeed recognition Infer user needs Shopping guideInformation search - -Evaluation of alternatives Provide product info Question answering over propertiesPurchase decision Provide explanations Recommendation reason generationPost- purchase behavior - -
AliMe [7] is a real-world chatbot application that offers pre-salescustromer service for hundreds of thousands of stores on the Al-ibaba platform. Being similar to offline shopping in physical stores,customers often describe their problems and ask for suggestions orinquire about a particular product. For example, “My skin is a bitdry, what kind of facial cleanser is suitable? 我的皮肤有点干,适合什么洗面奶 ”. If we directly use the extracted keywords “skin isa bit dry” and “facial cleanser” to query search engines, the resultsare often somehow irrelevant.To answer such knowledge-oriented questions, we applied Al-iMe KG for query rewriting and item recall . On one hand, the searchkeywords of the aforementioned query will be re-written as “pre-serve moisture 保湿 ” according to the domain knowledge “dryskin 皮肤干 ” - need - “preserve moisture 保湿 ” captured in our KG,and then used to query the Alibaba product graph. On the otherhand, we maintain an active item pool ourselves and construct in-verted index for items in the form of “ 𝑃𝑂𝐼 − 𝑖𝑡𝑒𝑚 , 𝑖𝑡𝑒𝑚 ... ”based on I-PV-POI knowledge aforehand for item recall in subse-quent recommendation. Our A/B test in the Beauty&Personal Care domain shows that our KG is able to cover 5% of the pre-sales con-versations and bring a relative increase of + in Conver sionRate (CVR). Customers in pre-sales conversations would also ask detailed prod-uct questions to seek more information for making purchasing de-cisions. For example, “Can it be used by pregnant women?” queriesabout the property “target users”. For this purpose, we employedKBQA (question answering over knowledge base) [6] to answerquestions concerning product properties. Specifically, we use NERto recognize properties and/or preoperty values from questions,judge customers’ intention (which property is being queried if mul-tiple ones are mentioned), and then retrieve corresponding prop-erty value as answer. Compared with previous approach that refercustomers to product pages for detailed product questions, we areable to provide accurate answers with AliMe KG.
When customers query about or when we recommend a specificitem, we can also provide a recommendation reason to explain whythe product is suitable based on our domain knowledge graph. Forexample, when customers send an “sweater 卫衣 ” item link in Al-iMe, we can retrieve from our KG its style “round neck 圆领 ” andassociated POIs “cute 可爱 ” and “leisure 休闲 ”, then we are ableto generate an explanatory recommendation reason “This sweaterhas a cute round neck, and brings a feeling of cute and leisure 这件卫衣的领子是圆形的款式,显得非常的可爱和休闲 ” usinggraph-to-sequence generation technique [23]. We have applied ourKG for generating explanatory recommendation reasons mainly inthe Clothing domain. Online tests show that our KG is able to cover30% of the pre-sales conversations and the CVR gains a relative in-crease of 4.8%. There have been many efforts on establishing open domain KGsuch as WordNET [11], Freebase [2] and DBpedia [1]. Unlike earlylexical knowledge bases such as WordNET [11] and HowNET [14]that are mainly established manually by experts, our KG is con-structed semi-automatically, with crowdsourcing involved in theprocess. Being different from Freebase [2] and DBpedia [1] that fo-cus on describing facts with well defined types, we lay emphasison the schema layer, where the POI concepts and relations thereofare extracted from natural language text.NELL [12] tries to automatically extract triples from web withan initial ontology defining categories (e.g., Athlete, Sport) and bi-nary relations (AthletePlaySport), but has a limited precision andscale of concepts. Probase [21] provides a large-scale probabilistic
𝐼𝑠𝐴 concept taxonomy in support of text understanding. Further,ConceptNet [16] inlcudes common sense knowledge by capturinginformal relations between concepts, which are words or phrasesof natural language that conceptualize general human knowledge(e.g., “hot weather”
𝐶𝑎𝑢𝑠𝑒𝐷𝑒𝑠𝑖𝑟𝑒 “turn on air conditioner”). OurKG captures conceptual level knowledge (e.g., “dry skin” 𝑛𝑒𝑒𝑑 “pre-serve moisture”), which is more akin to ConceptNet, but in a ver-tical domain.n the field of E-commerce, Amazon has proposed “Product Knowl-edge Graph (PG)” with the puprose of answering any questionabout products and related questions, but currenly do not focuson user needs. The most similar to us is AliCoCo [10], which en-richs the product taxnomy with E-commerce concepts that cap-ture user needs, and accordingly associates the new concepts withitems for search and recommendation. We differ from their workin two aspects. First, we have different purposes and applications:AliMe KG is mainly used for pre-sales conversation while AliCoCois mainly designed for search and recommendation. Second, moreimportanty, our KG captures domain knowledge which enables toinfer user needs and explain why a product is fit for certain userproblems, which contributes to optimizing customers’ buying pro-cess as we have discussed in Sec. 6.
In this paper, we demonstrated that constructing industrial scaleKG from free text in vertical domain is practicable, and showedhow to reduce manual annotation and utilize external knowledgeto improve model performance. Our approach also has limitationson quality control as concept level knowledge need to be fullychecked at the moment, and need to be further improved.There are three points to be noted when building a knowledgegraph. First of all, set a clear objective. There are some questions toask before diving into technique details: what are the (incremen-tal) value of KG? What are the application scenarios of KG? Is deeplearning (e.g., text classification and matching) sufficient? Second,design a feasible technological path. For example, keep schemasimple as complex schema brings poor scalability, redude humanannotation costs as you can. Third, do not forget data quality con-trol as industrial applications often impose strict quality require-ments on data and knowledge, which are often not easy to satisfy.
In this paper, we propose AliMe KG, a domain knowledge graph de-signed for better connecting customers to items through pre-saleschatbot conversation. We systematically introduce how it is con-structed from free text semi-automatically, and demonstrate how itcontributes to optimizing customers’ buying process through sev-eral real-world applications.On one hand, we will continually enlarge AliMe KG to cover themajority of vertical domains on the Alibaba E-commerce platform.On the other hand, there is a new trend of influencer marketing andlive streaming in E-commerce, for which we need multi-modalityitem content, including but not limited to text, image and video.Multimodal KG is our key topic for the next step.
REFERENCES [1] Sören Auer, Christian Bizer, Georgi Kobilarov, Jens Lehmann, Richard Cyganiak,and Zachary Ives. 2007. Dbpedia: A nucleus for a web of open data. In
Thesemantic web . Springer, 722–735.[2] Kurt Bollacker, Colin Evans, Praveen Paritosh, Tim Sturge, and Jamie Taylor.2008. Freebase: a collaboratively created graph database for structuring humanknowledge. In
Proceedings of the 2008 ACM SIGMOD international conference onManagement of data . 1247–1250.[3] Heng-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra,Hrishi Aradhye, Glen Anderson, Greg Corrado, Wei Chai, Mustafa Ispir, et al. 2016. Wide and deep learning for recommender systems. In
Proceedings of the1st workshop on deep learning for recommender systems . 7–10.[4] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert:Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018).[5] Zhiheng Huang, Wei Xu, and Kai Yu. 2015. Bidirectional LSTM-CRF Models forSequence Tagging.
CoRR abs/1508.01991 (2015). arXiv:1508.01991[6] Feng-Lin Li, Weijia Chen, Qi Huang, and Yikun Guo. 2019. AliMe KBQA: Ques-tion Answering over Structured Knowledge for E-Commerce Customer Service.In
China Conference on Knowledge Graph and Semantic Computing . Springer,136–148.[7] Feng-Lin Li, Minghui Qiu, Haiqing Chen, Xiongwei Wang, Xing Gao, Jun Huang,Juwei Ren, Zhongzhou Zhao, Weipeng Zhao, Lei Wang, et al. 2017. Alime assist:An intelligent assistant for creating an innovative e-commerce experience. In
Proceedings of the 2017 ACM on Conference on Information and Knowledge Man-agement . 2495–2498.[8] Jing Li, Aixin Sun, Jianglei Han, and Chenliang Li. 2020. A survey on deeplearning for named entity recognition.
IEEE Transactions on Knowledge and DataEngineering (2020).[9] Weijie Liu, Peng Zhou, Zhe Zhao, Zhiruo Wang, Qi Ju, Haotang Deng, and PingWang. 2019. K-bert: Enabling language representation with knowledge graph. arXiv preprint arXiv:1909.07606 (2019).[10] Xusheng Luo, Luxin Liu, Yonghua Yang, Le Bo, Yuanpeng Cao, Jinhang Wu,Qiang Li, Keping Yang, and Kenny Q Zhu. 2020. AliCoCo: Alibaba E-commerceCognitive Concept Net. arXiv preprint arXiv:2003.13230 (2020).[11] George A Miller. 1995. WordNet: a lexical database for English.
Commun. ACM
38, 11 (1995), 39–41.[12] Tom Mitchell, William Cohen, Estevam Hruschka, Partha Talukdar, Bishan Yang,Justin Betteridge, Andrew Carlson, Bhanava Dalvi, Matt Gardner, Bryan Kisiel,et al. 2018. Never-ending learning.
Commun. ACM
61, 5 (2018), 103–115.[13] Nanyun Peng and Mark Dredze. 2016. Improving Named Entity Recognitionfor Chinese Social Media with Word Segmentation Representation Learning. In
Proceedings of the 54th Annual Meeting of the Association for Computational Lin-guistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 2: Short Papers .The Association for Computer Linguistics.[14] Fanchao Qi, Chenghao Yang, Zhiyuan Liu, Qiang Dong, Maosong Sun, and Zhen-dong Dong. 2019. Openhownet: An open sememe-based lexical knowledge base. arXiv preprint arXiv:1901.09957 (2019).[15] Jingbo Shang, Jialu Liu, Meng Jiang, Xiang Ren, Clare R Voss, and Jiawei Han.2018. Automated phrase mining from massive text corpora.
IEEE Transactionson Knowledge and Data Engineering
30, 10 (2018), 1825–1837.[16] Robyn Speer, Joshua Chin, and Catherine Havasi. 2017. Conceptnet 5.5: An openmultilingual graph of general knowledge. In
Thirty-First AAAI Conference on Ar-tificial Intelligence .[17] Hongwei Wang, Fuzheng Zhang, Jialin Wang, Miao Zhao, Wenjie Li, Xing Xie,and Minyi Guo. 2018. Ripplenet: Propagating user preferences on the knowledgegraph for recommender systems. In
Proceedings of the 27th ACM InternationalConference on Information and Knowledge Management . 417–426.[18] Hongwei Wang, Fuzheng Zhang, Xing Xie, and Minyi Guo. 2018. DKN: Deepknowledge-aware network for news recommendation. In
Proceedings of the 2018world wide web conference . 1835–1844.[19] Hongwei Wang, Miao Zhao, Xing Xie, Wenjie Li, and Minyi Guo. 2019. Knowl-edge Graph Convolutional Networks for Recommender Systems. In
The WorldWide Web Conference (San Francisco, CA, USA) (WWW ’ . Association forComputing Machinery, New York, NY, USA, 3307 – Proceedings of the 28th ACM In-ternational Conference on Information and Knowledge Management . 2361–2364.[21] Wentao Wu, Hongsong Li, Haixun Wang, and Kenny Q Zhu. 2012. Probase: Aprobabilistic taxonomy for text understanding. In
Proceedings of the 2012 ACMSIGMOD International Conference on Management of Data . 481–492.[22] Bo Xu, Yong Xu, Jiaqing Liang, Chenhao Xie, Bin Liang, Wanyun Cui, andYanghua Xiao. 2017. CN-DBpedia: A never-ending Chinese knowledge extrac-tion system. In
International Conference on Industrial, Engineering and Other Ap-plications of Applied Intelligent Systems . Springer, 428–438.[23] Kun Xu, Lingfei Wu, Zhiguo Wang, Yansong Feng, and Vadim Sheinin.2018. Sql-to-text generation with graph-to-sequence model. arXiv preprintarXiv:1809.05255 (2018).[24] Yue Zhang and Jie Yang. 2018. Chinese NER Using Lattice LSTM. In
Proceed-ings of the 56th Annual Meeting of the Association for Computational Linguistics,ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers , IrynaGurevych and Yusuke Miyao (Eds.). Association for Computational Linguistics,1554–1564.[25] Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma,Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep interest network forclick-through rate prediction. In