A Simple Disaster-Related Knowledge Base for Intelligent Agents
Clark Emmanuel Paulo, Arvin Ken Ramirez, David Clarence Reducindo, Rannie Mark Mateo, Joseph Marvin Imperial
AA Simple Disaster-Related Knowledge Base for Intelligent Agents
Clark Emmanuel Paulo, Arvin Ken Ramirez, David Clarence ReducindoRannie Mark Mateo, Joseph Marvin Imperial
National UniversityManila, Philippines [email protected]
Abstract
In this paper, we describe our efforts in estab-lishing a simple knowledge base by building asemantic network composed of concepts andword relationships in the context of disastersin the Philippines. Our primary source of datais a collection of news articles scraped fromvarious Philippine news websites. Using wordembeddings, we extract semantically similarand co-occurring words from an initial seedwords list. We arrive at an expanded ontol-ogy with a total of 450 word assertions. Welet experts from the fields of linguistics, disas-ters, and weather science evaluate our knowl-edge base and arrived at an agreeability rate of64%. We then perform a time-based analysisof the assertions to identify important seman-tic changes captured by the knowledge basesuch as the (a) trend of roles played by humanentities, (b) memberships of human entities,and (c) common association of disaster-relatedwords. The context-specific knowledge basedeveloped from this study can be adapted byintelligent agents such as chat bots integratedin platforms such as Facebook Messenger foranswering disaster-related queries.
The Philippines is a common ground for natural dis-asters such as typhoons and flooding. According tothe latest statistics of Philippine Atmospheric, Geo-physical and Astronomical Services Administration(PAG-ASA) , there are more tropical cyclones en- Statistics on annual tropical cyclones in the Philippines:http://bagong.pagasa.dost.gov.ph/climate/tropical-cyclone-information tering the vicinity of the Philippines than anywhereelse in the world. In a year, almost 20 tropical cy-clones enter with 70% chance of developing into afull-blown typhoon. As possible result, catastrophicaftermath such as economic failure, destruction ofinfrastructures, and loss of lives may be imminentwithout proper information dissemination and pre-paredness. Thus, directing all research and techno-logical efforts to disaster preparedness and disasterrisk reduction to mitigate the tremendous impacts ofnatural calamities have been prioritized by the coun-try for years.According to Statista , the Philippines has 44 mil-lion active Facebook users in 2019 and is predictedto reach approximately 50 million by 2023. Thus,taking in consideration the disaster-prone situationof the country, we note the importance of establish-ing context-specific knowledge bases where it can beintegrated to commonly used digital platforms suchas Facebook Messenger to answer context-specificquestions on disasters for possible information dis-semination and awareness. To make this possible,we present our initial efforts in building a simpleknowledge base composed of word concepts joinedby semantic word relationships extracted from wordembeddings trained from a large online news corpuson disasters.The overview of the study is as follows: First,we compared and contrasted previous works donein field of ontology learning and mining informa-tion from word embeddings for building knowledge a r X i v : . [ c s . C L ] J a n ases. Next, we discussed the method of extrac-tion of concepts and entities from our news articledataset as well as the semantic labels used for build-ing concept relationships. In the Results section,we detail the outcome of using word embeddingstrained from our collected dataset for extracting se-mantically similar concepts using an initial conceptseedwords list. In addition, we also discussed theintegrity of the disaster ontology from expert valida-tion. We wrap up the study by extensively discussingthe semantic changes exhibited by concepts presentin the knowledge base over time. In this section we highlight significant works in us-ing knowledge bases (KBs) as the main source of in-telligence for virtual agents as well as current trendsusing word embeddings as one way of extracting in-formation and semantic word relationships.
Over the years, the use of powerful knowledge basesare seen on a wide variety of intelligent agents suchas chat bots, storytelling agents, and recommendersystems. The work of (Ong et al., 2018) focusedon building a commonsense knowledge base for astorytelling agent for children using assertions ex-tracted from ConceptNet (Speer et al., 2017). Con-ceptNet is a large semantic network of word rela-tionships that can be adapted by intelligent agentsfor commonsense reasoning and identifying objectrelationships. The study developed a knowledgebase by filtering out concepts and concept relation-ships that are not used in the context of children sto-rytelling from the ConceptNet.Similarly, Han et al. (2015) developed a naturallanguage dialog agent that utilizes a knowledge baseto generate diverse but meaningful responses to theuser. The system extracts related information fromknowledge base, which was adapted from FreeBase(Bollacker et al., 2008), via an information extrac-tion module. The use of a large, external knowledgebase allows the dialog system to expand on the infor-mation from the users response for a more detailedand interactive reply.On the other hand, the work of Wang et al. (2010) focused on the development of a systemcomposed of three ontology-based sub-agents forpersonal knowledge, fuzzy inference, and semanticgeneration for evaluating a person’s health throughhis/her diet. The system makes use of an ontologywith an embedded knowledge base considering thepersons health statistics such as BMI, Caloric Dif-ference, Health Diet Status combined with rules laiddown by domain experts. Results shows that theproposed system exhibits an intelligent behavior inhelping the dietary patterns of users based on theirinformation from the constructed ontology.
The advent of word embeddings as one of the mod-ern approaches in extracting semantic relationshipsof words has fueled research works to use its poten-tial to build more powerful knowledge bases. Sarkaret al. (2018) used a supervised approach, similarto text classification, for predicting the taxonomicrelationship via similarity of two concepts using aWord2Vec embedding (Mikolov et al., 2013b). Re-sults showed that combining the word embeddingwith an SVM classifier outperformed baseline ap-proaches for taxonomic relationship extraction suchas using the Jaccard similarity formula and naivestring matching.The study of Luu et al. (2016) went as far as build-ing a custom neural network architecture with dy-namic weighting to significantly increase the perfor-mance of statistical and linguistic approaches in ex-traction word relationships from word embeddings.The neural network considers not only the word rela-tionship such as hypernym and hyponym but also thecontextual information between the terms. The pro-posed approach exhibits generalizability for unseenword pairs and has obtained 9% to 13% additionalaccuracy score using a general and domain-specificdatasets.Likewise, the work of Pocostales (2016) submit-ted to the SemEval-2016 Task for Taxonomy Extrac-tion Evaluation (Bordea et al., 2016) focused on us-ing GloVe word embedding model (Pennington etal., 2014) with an offset feature to extract hyper-nym candidates from a sample word list. Resultsshowed that a vector offset cannot completely cap-ture the hypernym-hyponym relationship of wordsdue to complexity.
017 Tag family noun fire noun flood noun update verb tricycle noun act verb announced verb ashfall noun police noun lava noun bulletin noun quake noun typhoon noun weakened adjective affected verb
Table 1: Top 5 seed words per year.
For this study, we scraped over 4,500 Filipinodisaster-related news articles from years 2017 to2019 (1,500 articles per year) from Philippine newswebsites using Octoparse Webscraping Tool as ourprimary dataset. The corpus covers a wide rangeof natural disasters that transpired in the Philippinessuch as typhoons, earthquakes, landslides and alsoincludes statistics from damages and casualty re-ports. This large collection of news articles will bethe groundwork of the knowledge base as it containsdisaster-related context words as concepts. We par-titioned the dataset into three by year (2017 to 2019)for the word embedding model generation. The pur-pose of partitioning the dataset will allow us to an-alyze the temporal changes of sematic relationshipsof concepts. More on this is discussed in the suc-ceeding sections. To note, all concepts presented inthis document are translated to English for the inter-national audience.
Building a knowledge base starts with establishingconcepts or words that refer to real world entitiessuch as nouns , adjectives , and verbs that representeveryday objects such as apple, spoon, cake , peo-ple like mother, police, mayor , description of objectssuch as beautiful, red, big , and action words that sig-nify an activity such as walks, eating and jumped (Ong, 2010).To identify the grammatical category of each con-cept to know whether it is a noun, an adjective, ora verb to aid the semantic relationship labelling, weused a Filipino parts-of-speech (POS) tagger devel-oped by Go and Nocon (2017) which is currently github.com/matthewgo/FilipinoStanfordPOSTagger integrated in the Stanford CoreNLP package (Man-ning et al., 2014). We extracted the top 50 high-frequency concepts per year from the collected newsdataset, having a total of 150 initial seed words. Ta-ble 1 shows the top 5 seed words per year from theinitial word list. Both common words such as fam-ily , tricycle , police and update as well as disaster-related words such as flood , ashfall , typhoon , and quake to name a few are present. These conceptsare then paired with other concepts to form a mean-ingful representation of knowledge called a binaryassertion described in the next section. Once the set of context-specific words are obtainedfrom the dataset, in the case of this study, disaster-related concepts, the next process to establish thecorrect semantic relationship of words. These se-mantic relationships can be structured in the formof a binary assertion as previously stated. Ong etal. (2018) stated that the binary assertions of con-cepts are needed by virtual agents such as a story-telling agent or a chatbot to be able to generate re-sponses from a commonsense knowledge base. Se-mantic relationships can also be used for other taskssuch query expansion of words as well as informa-tion retrieval (Attia et al., 2016). For this study, weadapt the binary assertion format used by Concept-Net (Speer et al., 2017) shown below: [concept1 semantic-rel concept2] where semantic-rel stands for the speci-fied semantic relationship of the two concepts ( concept1 and concept2 ) that contains mean-ing. We adapted the six semantic relation labelsfrom the CogALex-2016 Shared Task on Corpus-Based Identification of Semantic Relations (Santuset al., 2016) which are
Synonym (SYN) Antonym(ANT) , Hypernym (HYP) , Membership (PartOf) , elation Tag Rule Synonym SYN If concept1 has the same meaning with concept2
Antonym ANT If concept1 has the opposite meaning with concept2
Hypernym HYP If concept1 has a broader meaning compared to concept2
Performs DO If concept1 is the actor/does of concept2
Membership PartOf If concept1 is a member of concept2
Adjective IS If concept1 describes concept2
Cause CAUSE If concept1 is the cause of event concept2
Effect dueTo If concept1 the resulting effect of event concept2
Random RAND If concept1 has no direct relationship concept2
Table 2: Semantic relations with its corresponding rules and examples. and
Random (RAND) as shown in Table 2. In ad-dition, we also added a few semantic labels of ourown with respect to the concepts that we will beworking on this study which are in the field of dis-asters. Thus, we added
Performs (DO) to signifyaction,
Adjective (IS) to signify description, and
Cause (CAUSE) and
Effect (dueTo) to signify con-sequences of events in a disaster setting.
Word embeddings are representations of an en-tire document vocabulary in a mathematical vectorspace. Each word is represented with a set of real-valued numbers given a specific set of dimensions.Word embedding architectures such as
Word2Vec (Mikolov et al., 2013a; Mikolov et al., 2013b) and
Global Vectors or GloVe (Pennington et al., 2014)capture various relationships of words in a corpussuch as the semantic similarity , syntactic similar-ity , and word co-occurrence to name a few. Thus, iftwo words are commonly used together in the samecontext, such as in disaster-related articles, we canexpect them to be close together when representedin the vector space. For example, finding the mostsimilar terms using the query word warning from aword embedding model trained on a disaster-relateddataset will output words such as typhoon , flood-ing , and signal since the word warning is commonlyused in texts to notify people of possible natural dis-asters.For this study, we generated a word embeddingmodel for each year of partitioned news articledataset from 2017 to 2019. A total of three mod- els were generated using the Word2Vec architecture(Mikolov et al., 2013b). In addition, we used the ini-tial seedwords list which contains 150 concepts (50for each year) as query words to extract semanticallysimilar terms which occur in the context of disaster. Table 3 shows the expanded ontology by queryingfour sample seedwords from the word embeddingmodel. We only filtered the top three semanticallysimilar resulting words from each query word thatfalls under a manually-annotated and qualified se-mantic label discussed in Section 4.2. From this, wearrived at an expanded ontology composed of 450assertions for our disaster-related knowledge base.From the resulting expanded ontology, it alreadyprovides us the knowledge that an intelligent agentcan piece together or form when asked about some-thing in the context of disasters. Take in, for ex-ample, the word tremor . Obtaining the top threeassertions with semantic labels and expanded usingword embeddings informs us that the word tremor issynonymous and interchangeable with the word af-
Figure 1: Hierarchy of hypernyms formed from the ex-panded ontology. eedword Semantic Label Assertions police DO [police DO risk]IS [police IS armed]DO [police DO commitment]tremor SYN [tremor SYN aftershock]dueTo [tremor dueTo quake]dueTo [tremor dueTo earthquake]rescue partOf [rescue partOf operations]HYP [rescue HYP retrieval]HYP [rescue HYP aid]experts IS [experts IS supporting]DO [experts DO recommend]DO [experts DO impose]
Table 3: Expansion of knowledge base using word embeddings. tershock which scientifically means the involuntarymovement of the surface due to breaking of under-ground rocks. Likewise, the words quake and earth-quake are annotated with the semantic label dueTo which denotes effect since earthquakes is root causeof tremors according to scientific definition (Yose,2013).Another observation from the expanded ontologyis the knowledge of understanding general actions tomore specific ones. In the example, the action word rescue is a hypernym , which means it is a generalword that can be possibly specified further, in thiscase, it is the hypernym of the word aid . Conse-quently, the word operation is a hypernym of rescue .Combining the words altogether, we can interpretthe three assertions as a string of successive actionsas shown in Figure 1 where in an operation involvesa rescue and the meaning of rescue may vary suchthat it can be some form of (a) retrieval of missingpeople or (b) aid for the wounded or stranded peo-ple.Lastly, the word embedding model was able toproduce words that imply responsibilities by con-cepts in the form of human entities. In the two exam-ples of human entities shown in Table 3, experts and police , most of the semantic labels are DO whichdenotes action and IS which denotes description .For police , common known actions done in the con-text of disasters are risk and commitment while be-ing armed . For experts , the responsibilities are in theline of supporting a claim as well as recommend- ing and imposing future actions based on scientificknowledge.We observe the significant potential of using wordembeddings for expanding knowledge bases by cap-turing various levels of information such as (a) un-derstanding interchangeable context-specific words,in the case of this study on disaster-related words,and (b) understanding roles played by human enti-ties described in this section. When used by an in-telligent agent, it will have an idea of what responsecan be produced when queried with questions suchas ”What does a police officer do?” or ”What hap-pens after an earthquake?” . To properly gauge the effectivity of using word em-beddings for the expansion of knowledge bases, weperformed a validation process by inviting three ex-perts in the fields of linguistics, disaster response,and meteorology to evaluate the assertions of theknowledge base. Each assertion is evaluated usinga two-scale metric : Agree if the expert deems thatthe assertion observes a correct relationship and se-mantic labelling ([ doctor
SYN medical person ]) or
Disagree if the assertion is not properly labelled toform a correct relationship ([ flood
SYN typhoon ]).Table 4 shows the averaged agreeability rate ofthe expert validation process. Results show that themost accurate model with the highest rating is theWord2Vec model trained on disaster-related newsarticles collected in the year of 2019. This is fol- odel Aggreability Rate
Word2Vec 2019 0.64Word2Vec 2018 0.52Word2Vec 2017 0.49
Table 4: Expert validation for the knowledge base. lowed by word embedding models using the 2018and 2017 dataset. We attribute the semi-low agree-ability scores due to the variation of word usage ofthe experts in their corresponding fields of study.For example, the assertion [ tremor
SYN earthquake ]was evaluated by the linguist and disaster experts as
Agree while the meteorologist contested. The ex-pert on meteorology swears by the scientific defini-tion of which tremors are very different with earth-quakes in a way that tremors are caused by earth-quakes and not an interchangeable term as perceivedby the other two non-technical experts. This patternis also observed with other assertions using the syn-onym labelling such as [ storm
SYN typhoon ] and[ lava
SYN magma ]. In this section, we perform an even more in-depthanalysis by considering the changes in semanticmeaning or semantic information of the assertionsof the knowledge base over time. We break the dis-cussion into three categories we observed from theanalysis.
We observe the changing roles played by three es-sential human entities, governor , experts , and police over time in the setting of a natural disaster as writ-ten in news articles. These entities are expected tobe on full alert and their responsibilities are crucialtowards mitigating the consequences of disasters.The entity governor , or the highest commandingindividual of Philippine province, performs variousroles over time. As seen in Table 5, the entity ismostly connected by the semantic label DO to de-note action with concepts such as assure , explain , oversee , develop , and declare . There are also a fewdescription words connected by the label IS suchas political and mandatory which is obvious sinceholding a gubernatorial position is indeed political and resolutions filed in a governor’s office is in itsessence mandatory .Similarly, the entity experts is also expected tohave various changing roles. For 2017 as seen in theTable, common action words associated are check-ing and verifying which provides one of the mostimportant roles of experts in the field of disasters:validating the integrity of information being publi-cized. For 2018, experts play more of an informationdissemination role with the concept announce as theconnecting action word. In 2019, experts assumeda more stricter role as observed with the connectedconcepts such as recommend and impose .For the entity police , there is consistent associateddescriptor across the years regarding their responsi-bility: risk . This provides us a concrete idea of aconsequence when assuming the role of a police. In2017 and 2018, strong action words such as control , damage , and armed are tied with a policeman’s job.In 2019, however, the entity police assumed a morepassive role as more of an informant with the con-nected action words being study and alerts . We also observed changes in memberships of hu-man entities in context of disasters over time as seenin Table 6. Memberships are denoted by the seman-tic labels
HYP for hypernyms or general words and partOf for hyponyms more specific words. We ob-serve memberships played of three entities: mayor , student , teacher .For the entity mayor , memberships are more spe-cific in 2017 and 2018. The entity is expected topartner with small groups in a municipality such asa sitio or a barangay cite as well as with large gov-ernment agencies such as Division on the Welfare ofthe Urban Poor (DWUP) and Department of Envi-ronment and Natural Resources (DENR) in times ofdisasters. There are also general memberships suchas administration and government to which the en-tity is an obvious member.In the case of the entity student , there is a mix ofspecific and generalized memberships. Certain spe-cific universities such as UP or the University of thePhilippines and PUP or Polytechnic University ofthe Philippines were classified as institutions wherea student may belong. General and obvious conceptssuch as elementary , school , organization and Uni- eed 2017 2018 2019 governor [governor DO assure] [governor IS political] [governor DO declare][governor DO giving] [governor DO oversees] [governor DO resolution][governor DO explain] [governor IS mandatory] [governor DO develop]experts [experts SYN representative] [experts DO work] [experts IS supporting][experts DO checking] [experts IS frantic] [experts DO recommend][experts DO verify] [experts DO announce] [experts DO impose]police [police DO risk] [police DO risk] [police DO risk][police DO control] [police IS armed] [police DO study][police DO damage] [police DO commitment] [police DO alerts]
Table 5: Roles played by human entities.
Seed 2017 2018 2019 mayor [mayor partOf hall] [mayor partOf administration] [mayor RAND workers][mayor partOf DWUP] [mayor DO communication] [mayor partOf government ][mayor partOf sitio] [mayor partOf DENR] [mayor RAND employers]student [student partOf PUP] [student partOf elementary] [student partOf organization][student IS victim] [student partOf school] [student partOf University][student IS resident] [student IS minor] [student partOf UP]teacher [teacher DO education] [teacher DO research] [teacher partOf house][teacher partOf DepEd] [teacher DO education] [teacher SYN employee][teacher partOf school] [teacher partOf school] [teacher partOf school]
Table 6: Memberships of human entities. versity contribute to the commonsense informationof the knowledge base.For the entity teacher , general concepts of mem-berships are more prominent compared to the othertwo entities. The concept school is consistent for allyears. The term
DepEd which means Departmentof Education, the government agency responsiblein shaping the educational landscape of the Philip-pines, is connected to the entity. DepEd oversees el-ementary and intermediate level schools which maymean that the term teacher is commonly tied witheducators from these levels. Interestingly, the con-cept house is also tied with the entity teacher . Al-though it may already be obvious that a teacher as-sumes a different role inside the house maybe as amother or a breadwinner.
For the last category, we observe change in associ-ation of disaster-related words over time as seen in Table 7. We highlight the importance of this analysisto understand how a knowledge base using informa-tion extracted from word embeddings interchangeco-occurring and similar words in the context of dis-asters. For this category, we analyze the four fre-quently occurring natural disasters in the Philippineswhich are earthquakes , eruption , landslide , and typhoon .The first disaster-related word is earthquake .From the formed assertions for all years, earth-quakes are synonymous with the term quakes whichdenote a shortened version of the word. In addition,the term magnitude is also a consistent term con-nected to the seedword earthquake which tells usthat earthquakes have their own corresponding mag-nitudes quantified by some number.In the case of eruption , most assertions formed areused with the semantic label IS to denote descrip-tion or property . Across the years, most of the de-scriptive words associated with the term are negativesuch as amplifying , hazardous , destruction , explo- eed 2017 2018 2019 earthquake [earthquake IS magnitude] [earthquake IS magnitude] [earthquake IS magnitude][earthquake SYN quake] [earthquake SYN quake] [earthquake SYN quake][earthquake RAND drill] [earthquake IS intensity] [earthquake RAND signal]eruption [eruption IS amplifying] [eruption IS happening] [eruption IS hazardous][eruption IS fast] [eruption IS explosive] [eruption IS magmatic][eruption IS confirmed] [eruption CAUSE destruction] [eruption IS happening]landslide [landslide dueTo flood] [landslide dueTo rain] [landslide RAND mountain][landslide HYP mudslide] [landslide IS torrential] [landslide dueTo rains][landslide RAND area] [landslide HYP mudslide] [landslide IS widespread]typhoon [typhoon IS expected] [typhoon SYN storm] [typhoon dueTo Amihan][typhoon partOf calamity] [typhoon IS super] [typhoon SYN hurricane][typhoon SYN onslaught] [typhoon IS powerful] [typhoon SYN cyclone] Table 7: Disaster-related terms and its associated words. sive , and magmatic . These word may denote a senseof urgency compared to the other natural disasterswhen it happens. The term happening has occurredboth in 2018 and 2019 due to two of the most ac-tive volcanoes in the Philippines, Mount Mayon andMount Taal, had shown activity by erupting succes-sively and spewing ash. The eruption caused masspostponement of public activities and over 48,000evacuated locals.For the concept landslide , the common associ-ated word is mudslide . Although by scientific defini-tion, mudslides are as specific type landslides (alsocalled debris flow ). Thus, the landslide term con-forms the the semantic label HYP for hypernyms.The cause of landslides can be tied to floods , whichin turn, caused by rains as joined by the semanticlabel dueTo denoting a consequence or an effect .In addition, landslides are also often described us-ing the words widespread and torrential which givesus a quantifiable idea of the magnitude of landslidesthat occur in the Philippines.Lastly, we have the word typhoon . This disaster-related concept assumes many interchangeable andsynonymous terms such as storm , hurricane , cy-clone , and onslaught . We note the frequent asso-ciation and interchangeability of these words due togeographic locations (Khadka, 2018) but may meanthe same thing—they are all tropical storms . Like-wise, description words tied to the concept typhoon Volcano Bulletin: phivolcs.dost.gov.ph/index.php/volcano-hazard/volcano-bulletins3 are expected , super , and powerful which also pro-vides us an idea of the magnitude of typhoons oc-curring in the country similar to landslides . As a disaster-prone country, the Philippines shouldincrease its efforts in mitigating the effects of nat-ural calamities with the help of technology. Oneway to do this is to consider the potential of in-telligent agents such as chatbots as tools for disas-ter preparedness and information dissemination. Inthis study, we established a simple context-specificknowledge base by doing three important processes:(a) extracting disaster-related concepts from a col-lected news article dataset, (b) building a networkof binary assertions from a curated list of seman-tic labels, and (c) expanding the ontology by query-ing an initial seedwords list from word embeddingsgenerated from the original news dataset. Resultsshow that using word embeddings captured variouslevels of information that may be useful for intelli-gent agents to produce responses such as informa-tion on roles of human entities, generalization andspecification of terms, and common word associa-tion when asked in the topic of disasters. Future di-rections of the study include collection of even moredataset that covers not only news articles but alsoother media for finer-grained assertions. In addition,the study will also benefit from efforts in testing thecapability of the knowledge base in practical appli-cations. eferences [Attia et al.2016] Mohammed Attia, Suraj Maharjan,Younes Samih, Laura Kallmeyer, and Thamar Solorio.2016. Cogalex-v shared task: Ghhh-detecting seman-tic relations via word embeddings. In
Proceedings ofthe 5th Workshop on Cognitive Aspects of the Lexicon(CogALex-V) , pages 86–91.[Bollacker et al.2008] Kurt Bollacker, Colin Evans,Praveen Paritosh, Tim Sturge, and Jamie Taylor.2008. Freebase: a collaboratively created graphdatabase for structuring human knowledge. In
Pro-ceedings of the 2008 ACM SIGMOD internationalconference on Management of data , pages 1247–1250.[Bordea et al.2016] Georgeta Bordea, Els Lefever, andPaul Buitelaar. 2016. Semeval-2016 task 13: Taxon-omy extraction evaluation (texeval-2). In
Proceedingsof the 10th International Workshop on Semantic Eval-uation (SemEval-2016) , pages 1081–1091.[Go and Nocon2017] Matthew Phillip Go and Nicco No-con. 2017. Using Stanford part-of-speech tagger forthe morphologically-rich Filipino language. In
Pro-ceedings of the 31st Pacific Asia Conference on Lan-guage, Information and Computation , pages 81–88.[Han et al.2015] Sangdo Han, Jeesoo Bang, SeonghanRyu, and Gary Geunbae Lee. 2015. Exploiting knowl-edge base to generate responses for natural languagedialog listening agents. In
Proceedings of the 16thAnnual Meeting of the Special Interest Group on Dis-course and Dialogue , pages 129–133.[Khadka2018] Navin Singh Khadka. 2018. Hurricanes,typhoons and cyclones: What’s the difference?, Sep.[Luu et al.2016] Anh Tuan Luu, Yi Tay, Siu Cheung Hui,and See Kiong Ng. 2016. Learning term embed-dings for taxonomic relation identification using dy-namic weighting neural network. In
Proceedings ofthe 2016 Conference on Empirical Methods in NaturalLanguage Processing , pages 403–413.[Manning et al.2014] Christopher D Manning, Mihai Sur-deanu, John Bauer, Jenny Rose Finkel, StevenBethard, and David McClosky. 2014. The stanfordcorenlp natural language processing toolkit. In
Pro-ceedings of 52nd annual meeting of the associationfor computational linguistics: system demonstrations ,pages 55–60.[Mikolov et al.2013a] Tomas Mikolov, Kai Chen, GregCorrado, and Jeffrey Dean. 2013a. Efficient estima-tion of word representations in vector space. arXivpreprint arXiv:1301.3781 .[Mikolov et al.2013b] Tomas Mikolov, Ilya Sutskever,Kai Chen, Greg S Corrado, and Jeff Dean. 2013b.Distributed representations of words and phrases andtheir compositionality. In
Advances in neural infor-mation processing systems , pages 3111–3119. [Ong et al.2018] Dionne Tiffany Ong, Christine RachelDe Jesus, Luisa Katherine Gilig, Junlyn Bryan Al-buro, and Ethel Ong. 2018. Building a common-sense knowledge base for a collaborative storytellingagent. In
Pacific Rim Knowledge Acquisition Work-shop , pages 1–15. Springer.[Ong2010] Ethel ChuaJoy Ong. 2010. A commonsenseknowledge base for generating children’s stories. In .[Pennington et al.2014] Jeffrey Pennington, RichardSocher, and Christopher D Manning. 2014. Glove:Global vectors for word representation. In
Proceed-ings of the 2014 conference on empirical methodsin natural language processing (EMNLP) , pages1532–1543.[Pocostales2016] Joel Pocostales. 2016. Nuig-unlp atsemeval-2016 task 13: A simple word embedding-based approach for taxonomy extraction. In
Proceed-ings of the 10th International Workshop on SemanticEvaluation (SemEval-2016) , pages 1298–1302.[Santus et al.2016] Enrico Santus, Anna Gladkova, StefanEvert, and Alessandro Lenci. 2016. The CogALex-vshared task on the corpus-based identification of se-mantic relations. In
Proceedings of the 5th Workshopon Cognitive Aspects of the Lexicon (CogALex - V) ,pages 69–79, Osaka, Japan, December. The COLING2016 Organizing Committee.[Sarkar et al.2018] Rajdeep Sarkar, John Philip McCrae,and Paul Buitelaar. 2018. A supervised approach totaxonomy extraction using word embeddings. In
Pro-ceedings of the Eleventh International Conference onLanguage Resources and Evaluation (LREC 2018) .[Speer et al.2017] Robyn Speer, Joshua Chin, and Cather-ine Havasi. 2017. Conceptnet 5.5: An open multilin-gual graph of general knowledge. In
Thirty-First AAAIConference on Artificial Intelligence .[Wang et al.2010] Mei-Hui Wang, Chang-Shing Lee,Kuang-Liang Hsieh, Chin-Yuan Hsu, Giovanni Acam-pora, and Chong-Ching Chang. 2010. Ontology-based multi-agents for intelligent healthcare applica-tions.