Chaveevan Pechsiri

Medicinal property knowledge extraction from herbal documents for supporting question answering system

The aim of this paper is to automatically extract the medicinal properties of an object, especially an herb, from technical documents as knowledge sources for health-care problem solving through the question-answering system, especially What-Question, for disease treatment. The extracted medicinal property knowledge is based on multiple simple sentence or EDUs (Elementary Discourse Units). There are three problems of extracting the medicinal property knowledge: the herbal object identification problem, the medicinal property identification problem for each object and the medicinal property boundary determination problem. We propose using NLP (Natural Language Processing) with statistical based approach to identify the medicinal property and also with machine learning technique as Naive Bayes with verb features for solving the boundary problem. The result shows successfully the medicinal property extraction of the precision and recall of 86% and 77%, respectively, along with 87% correctness of the boundary determination.

Know-Why Extraction from Textual Data for Supporting What Questions

This research aims to automatically extract Know-Why from documents on the website to contribute knowledge sources to support the question-answering system, especially What-Question, for disease treatment. This paper is concerned about extracting Know-Why based on multiple EDUs (Elementary Discourse Units). There are two problems in extracting Know-Why: an identification problem and an effect boundary determination problem. We propose using Naive Bayes with three verb features, a causative-verb-phrase concept set, a supporting causative verb set, and the effect-verb-phrase concept set. The Know-Why extraction results show the success rate of 85.5% precision and 79.8% recall.

Mining Causality for Explanation Knowledge from Text

Mining causality is essential to provide a diagnosis. This research aims at extracting the causality existing within multiple sentences or EDUs (Elementary Discourse Unit). The research emphasizes the use of causality verbs because they make explicit in a certain way the consequent events of a cause, e.g., “Aphidssuckthe sap from rice leaves. Then leaves willshrink. Later, they willbecomeyellow anddry.”. A verb can also be the causal-verb link between cause and effect within EDU(s), e.g., “Aphids suck the sap from rice leavescausingleaves to be shrunk” (“causing” is equivalent to a causal-verb link in Thai). The research confronts two main problems: identifying the interesting causality events from documents and identifying their boundaries. Then, we propose mining on verbs by using two different machine learning techniques, Naïve Bayes classifier and Support Vector Machine. The resulted mining rules will be used for the identification and the causality extraction of the multiple EDUs from text. Our multiple EDUs extraction shows 0.88 precision with 0.75 recall from Naïve Bayes classifier and 0.89 precision with 0.76 recall from Support Vector Machine.

Comparative Study of Using Word Co-occurrence to Extract Disease Symptoms from Web Documents

The research aim is a comparative study of using different word co-occurrence sizes as the two word co-occurrence and the N word co-occurrence on verb phrases to extract disease symptom explanations from downloaded hospital documents. The research results are applied to construct the semantic relations between disease-topic names and symptom explanations for enhancing the automatic problem-solving system. The machine learning technique, Support Vector Machine, and the similarity score determination are proposed to solve the boundary of simple sentences explaining the symptoms for the two word co-occurrence and the N word co-occurrence respectively. The symptom extraction result by the N word co-occurrence provides the higher precision than the two word co-occurrence from the documents.

Collection of HerbalMedicinalProperty relation extracted from texts

This research aims to collect the extracted HerbalMedicinalProperty relations from downloaded herbal-plant documents for creating the herbal-medicinal-property-network based representation. An HerbalMedicinalProperty relation is a semantic relation between one herbal-plant-component concept and several herbal-medicinal-property-concept expressions on texts and vice versa. An herbal-plant-component occurrence is a noun-phrase expression and each herbal-medicinal-property- concept occurrence is an event expression by a verb-phrase of EDU (an Elementary Discourse Unit or a simple sentence). The herbal-medicinal-property-network based representation benefits a recommendation system of solving health-problems on web-boards. The research has two main problems: 1) how to extract HerbalMedicinalProperty relations from the documents, and 2) how to collect the HerbalMedicinalProperty relations for creating the herbal-medicinal-property-network based representation. Therefore, we propose applying a co-occurrence of N-Words (or N-Word-Co) including N-Word-Co size learning on the verb phrase to identify several medicinal-property-concept EDU occurrences over the documents after the linguistic phenomena has been applied to solve the herbal-plant-component concepts. The extracted HerbalMedicinalProperty relations are then collected as a matrix of herbal-plant names, herbal-plant components, and herbal-medicinal properties for creating the herbal-medicinal-property-network based representation. The research results provide the high precision of the HerbalMedicinalProperty-relation extraction from the documents.

Web Board Question Answering System on Problem-Solving Through Problem Clusters

This paper aims to work on the Question Answering (QA) system within online web boards, especially the Why-question, How-question, and Request-Diagnosis-question types approach for solving problems. The research QA system benefits for the online communities in solving their problems, especially on health-care problems of symptoms. Both question and answer expressions are based on multiple EDUs (Elementary Discourse Units) where each EDU is equivalent to a simple sentence or a clause. The research involves two main problems: how to identify the question types of Why, How, and Request-Diagnosis and how to determine the corresponding answer from the knowledge source after solving the question focuses. Thus, the research applies different machine learning techniques, Naive Bayes and Support Vector Machine, to solve the reasoning question type identification. The knowledge source contains several symptom-treatment vector pairs and several cause-effect vector pairs. Therefore, we propose clustering symptoms/problems of the knowledge source before determining an answer based on top-down levels of determining similarity scores between a web board question and the knowledge source. The research achieves 83 % correctness of the answer determination with potentially saving amounts of search time.

Using Extracted Symptom-Treatment Relation from Texts to Construct Problem-Solving Map

This paper aims to extract the relation between the disease symptoms and the treatments (called the symptom-treatment relation), from hospital-web-board documents to construct the problem-solving map which benefits inexpert people to solve their health problems in preliminary. Both symptoms and treatments expressed on documents are based on several EDUs (elementary discourse units). Our research contains three problems: first, how to identify a symptom-concept-EDU and a treatment-concept EDU. Second, how to determine a symptom-concept-EDU boundary and a treatment-concept-EDU boundary. Third, how to determine the symptom-treatment relation from documents. Therefore, we apply a word co-occurrence to identify a disease-symptom-concept/treatment-concept EDU and Naive Bayes to determine a disease-symptom-concept boundary and a treatment-concept boundary. We propose using k-mean and Naive Bayes to determine the symptom-treatment relation from documents with two feature sets, a symptom-concept-EDU group and a treatment-concept-EDU group. Finally, the research achieves 87.5 % precision and 75.4 % recall of the symptom-treatment relation extraction along with the problem-solving map construction.

Multi-Word-Co-occurrence collection from texts for health-problem diagnosis

This research aims to collect multi-word co-occurrences with health-problem/symptom concepts for health-problem diagnosis from wed-board documents. The result of this research is a benefit for assisting the ordinary people in preliminary diagnosis health problems. The multi-Word-Co of the research is based on an event expression by a verb phrase. However, the research contains two main problems; the first problem is how to identify multi-word co-occurrence including the multi-word co-occurrence boundary with the symptom concept after the stop word removal. The second one is the ambiguous multi-word co-occurrence concept. Therefore, the machine learning with Naïve Bayes is applied to solve the consequent words of the verb phrase (after the stop word elimination) as the multi-word co-occurrence with the symptom concept. The results of this research can provide the high precision of the symptom concept determination through multiword co-occurrences on documents.

Touristic Destinations’ Theme Determination for GIS Applications

This research aims to determine touristic destination’s theme (especially tourism activity theme) from the tourism web documents for geographic information system (GIS) applications, i.e. guiding the main interesting tourism activities to tourists. There are two major problems of the theme acquisition; tourism activity extraction and tourism activity generalization. Therefore, this research proposes of using Naïve Bayes Classifier to determine word co-occurrences between verbs and nouns with the tourism activity concept from web documents. Furthermore, this research also applies the fuzzy concept along with the imputation technique, to determine the tourism activity theme by generalizing the extracted tourism activity. The result of the tourism activity extraction shows successfully the precision and recall of 85% and 77%, respectively, with Mean Reciprocal Rank (MRR) of the tourism activity theme is 0.5.