RDF data extracted automatically often contain long textual literals. This paper shows how to use natural language processing techniques to automatically generate specific RDF triples from the information in the literals. We look specifically at drug indications found in the DailyMed dataset. We develop knowledge schemas to capture its information as well as precise syntactic-based methods of knowledge extraction to automatically generate instances of these schemas from textual data.
predicate. The indication for each drug is rather lengthy with its size varying from 1 word to 1338 words (average of 127 words).
The texts found in the labels contain sentences like Fluticasone propionate ointment is a medium potency corticosteroid indicated for the relief of the inflammatory and pruritic manifestations of corticosteroid-responsive dermatoses in adult patients. The word indicated appears to be a strong linguistic pattern for expressing a drug’s indication. As we will present in more details in section 2 (related work), linguistic patterns are often used for knowledge extraction. But, in our case, for an in-depth analysis of these indication predicates, they are not sufficient. Considering the complexity of language and the complexity of the knowledge to be represented, we chose to explore deep syntactical language analysis approaches and to define a knowledge schema for knowledge representation for drug therapy.
This paper is structured as follow. After a brief overview of related work, we define a general method for knowledge extraction and a general framework for knowledge representation as applied to the DailyMed indication predicate. Then we look more deeply at the resource and refine both our knowledge schema and our knowledge extraction method. Finally, we present an evaluation of our extraction approach. 2
Numerous proposals have been made to facilitate ontological engineering through
automatic discovery from domain data or domain-specific natural language texts.
Early algorithms for relation extraction, like DIPRE [
At least three dominating machine-learning related paradigms have been
applied to the task of extracting relational facts from non-labelled or
structured text [
Recently, [
Some specific Information Extraction techniques are investigated for the
Biomedical field[
Our first task is to define a knowledge schema for the representation of the knowledge expressed in the indication predicate. We start with a general schema which we will refine after (see section 4). Associated with such schema, we define a general extraction method to automatically extract information from the indication object (the literal). 3.1
Knowledge schema We first devise a general Treatment logical structure, that can be expressed in RDF, and that includes the concepts of drug, disease, and treatment as well as some refinements on the type of treatment. Below is an indication sentence and the knowledge structure we wish to automatically extract.
Restoril is indicated for the short-term treatment of insomnia. dailymed_drug:3239 rdfs:label "Restoril". loc:T234 loc:drugInvolved dailymed_drug:3239; rdf:type loc:Treatment, loc:DrugTherapy, loc:ShortTermTreatment; loc:target "insomnia";
Within the semantic web paradigm, we use the RDF format, and we express concepts via resources (URIs) and relations between concepts via predicates.
Our vocabulary for the RDF descriptions consists in a set of classes that are used to specify treatment types. To clearly indicate that the treatments involve some drug, we attribute to all of them the type loc:DrugTherapy. We use the predicate loc:target to link the treatment with its target disease or disorder. Predicate loc:drugInvolved identifies the drug that is used for the treatment4.
We do not make attempts in this research to link the extracted information to existing URIs. This is a research problem in itself outside the scope of the present research. 3.2
Knowledge extraction For precise knowledge extraction, we use a syntactic approach. As mentioned earlier in section 2, this is in line with recent work on knowledge extraction in the biomedical domain. The text is parsed with Stanford Parser5, a statistical parser that provides a dependency tree output. In a dependency tree, every word, except the root word (which usually is the main verb of the sentence) is linked to another word by a dependency relation. In a dependency relation, one word (the head word) is dominating the other one. Figure 1 shows the dependency tree returned by Stanford parser for our example sentence Restoril is indicated for the short-term treatment of insomnia (note that the root node is indicated in boldface). There is a subject relation (nsubjpass) between Restoril and indicated, and an indirect object relation (prep) between indicated and for. Note that the actual object is treatment, the dependent of for.
nsubjpass auxpass pobj det prep amod prep pobj Restoril is indicated for the short-term treatment of insomnia
Let T be a dependency tree for sentence S and sub(Head) the sub-tree of T whose root node is the word Head in S.
The extractor is based on a set of rules hP attern; Extractioni, where P attern is a syntactic pattern, in fact a dependency tree where some nodes are variables to be substituted by a sub-tree, and Extraction is a specification of how the RDF
and route of administration. In this paper, we leave these aspects out, but the method that is presented could be extended to embrace them.
triples are to be generated from the information found in the dependency tree that will be matched with the pattern. We define functions Drug(U1; U2; subT ree), Disease(URI; subT ree) and T reatment(URI; subT ree), that are used in the extraction part of a rule, to extract information for drug, disease and treatment, respectively. The function for Drug for the example above would be: Drug(dailymed_drug:3239; loc:T234; sub(Restoril)) = loc:T234 loc:drugInvolved dailymed_drug:3239 .
dailymed_drug:3239 rdfs:label "Restoril" .
An additional function, Qualif (U RI; sub), is used to process the following qualifiers: short-term, long-term, first-line, second-line, initial and acute. These qualifiers are not processed in the T reatment() function because they do not strictly subcategorize a treatment, as opposed to a palliative treatment which is a specific type of treatment. One important consequence of this separation is that a qualifier may be combined with any type of treatment. We will see subcategorization of treatments in section 4 as we refine our schema. nsubpass
prep pobj preppobj
In this section, we see how a simple frequency analysis on the corpus made of all drug indications show lexical and syntactical variations that help us refine our representation schema and consequently our extraction rules.
Corpus statistics All literals, object of the indication predicate in DailyMed have been joined to form a corpus of 3683 indications. Frequencies of 5-grams were counted to find variation in expression. Table 1 shows the most frequent 5-gram. We see clearly that the relation of interest is almost always expressed by the passive form is indicated for/in/as. Note that there are many occurrences of negation, where it is specified that some drug is not indicated for some disease. It is important to recognize these negated forms, otherwise we would extract an indication description where we should not.
5-gram # occ. indicated for the management of 163 indicated for the treatment of 1311 indicated as an adjunct to 142 indicated in the treatment of 354 indicated for the prevention of 90 indicated for the relief of 346 to reduce the risk of 76 reduce the development of drug-resistant 284 for the topical treatment of 72 to treat or prevent infections 278 is not indicated for the 69 alone or in combination with 173 indicated as adjunctive therapy in 68 the treatment of patients with 164 adjunctive therapy in the treatment 55
It is also clear that the interaction between a drug and a disease is not limited to a generic treatment relation. Reality is more complex: treatments can be subcategorized (prophylactic, palliative, short-term, symptomatic, management, etc), some treatments involve drug combination and, finally, some are adjunctive therapies, that is, they are given in addition to an initial treatment. From our corpus analysis, we identified more than 30 kinds of treatment. Examples of these are listed in Table 2. In each row, we indicate the category of the treatment, a short definition and the number of occurrences found in the corpus. Although unspecified treatments are most commonly found in Dailymed descriptions, more than half of indications lead to more specific treatments. That number is not negligible and confirms the relevance of recognizing these specificities. 4.2
Refinement of the representation schema Based on Table 2, different variations need to be included in the representation schema. We show this refinement process with one example: adjunctive therapy.
An adjunctive treatment is a treatment used in conjunction with another to increase the chance of cure or to augment the efficacy of another initial treatment. A new URI for the primary treatment must be introduced to which is associated the adjunctive drug therapy. The indicated drug is linked to this adjunctive therapy, whereas the disease is linked to the primary treatment. A new predicate loc:associatedWith establishes the link with the primary treatment. Treatment Treatment Management Prophylactic therapy Symptomatic treatment Topical treatment Combination therapy Short-term treatment Control teatment Adjunctive treatment Definition No specificity indicated.
Helps the patient to manage a chronic disease.
Used for prevention Relief of symptoms Medication applied to the skin surface.
Simultaneous use of a variety of drugs.
Used on a short periord Control of physiologic process or disease Used in conjunction with another to increase chance of cure or augment first treatment’s efficacy.
Below is an example with its corresponding RDF description.
Entacapone is used as an adjunct to levodopa/carbidopa to treat patients with idiopathic Parkinson’s Disease.
dailymed_drugs:DB00494 rdf:label "Entacapone". loc:DT00494A rdf:type loc:DrugTherapy, loc:AdjunctiveTreatment; loc:drugInvolved dailymed_drugs:DB00494; loc:associatedWith loc:A0023. loc:A0023 rdf:type loc:Treatment; loc:drugInvolved "levodopa/carbidopa"; loc:target "idiopathic Parkinson’s Disease".
nsubpass prep pobj aod prep pobj prep pobj M indicated as adjunctive therapy in T of D
T 2 ftreatment, prevention,
{UT rdf:type loc:DrugTherapy, UT rdf:type loc:AdjunctiveTreatment, UT loc:associatedWith UP } [ Drug(UD; UT ; subtree(M )) [ T reatment(UP ; sub(T )) [ Disease(UP ; sub(D)) [ Qualif (UP ; sub(T )) managementg The main challenge in the extraction process is to develop appropriate extraction rules and extraction functions for different types of expressions of the same knowledge. New extraction rules must be defined if the variation affects directly the dependency tree. Otherwise, if the variation occurs within the subtrees, the extraction functions must be refined. The same two examples, adjunctive therapy and combination therapy, are used to explain this refinement process.
For adjunctive therapy, the variations are at the dependency tree level. Figure 3 shows a new Rule Adj_0 to process sentences such as Gabapentin is indicated as adjunctive therapy in the treatment of partial seizures. Note that an additional URI must be provided to represent the primary treatment.
Name Tr_1 Tr_2 Tr_3 Tr_4 Sympt_0 Adj_0 Adj_1 Adj_2 Cont_0 Red_0 Neg_0
Example Freq. Univasc is indicated for treatment of patients with hypertension. (prepositional 111 phrase attached to patients) Univasc is indicated for treatment of patients with hypertension. (prepositional 140 phrase attached to treatment) Camptosar is also indicated for patients with metastatic carcinoma of the colon 43 or rectum.
Citalopram is indicated for the treatment of depression. 2729 Surmontil is indicated for the relief of symptoms of depression. 80 Gabapentin is indicated as adjunctive therapy in the treatment of partial 44 seizures.
Trihexyphenidyl HCl tablets are indicated as an adjunct in the treatment of 18 all forms of parkinsonism.
Glipizide tablets are indicated as an adjunct to diet for the control of hyper- 77 glycemia.
Provigil is indicated to improve wakefulness in adult patients with excessive 120 sleepiness associated with narcolepsy Inapsine is indicated to reduce the incidence of nausea and vomiting associated 49 with surgical and diagnostic procedures.
Kemstro is not indicated in the treatment of skeletal muscle spasm. 237 Following our corpus analysis, we developed 15 extraction rules grouped in 7 categories. Table 3 gives example sentences for the most frequently used rules. Some rules are mostly variations due to the fact that there are many ways to express the same thing and that syntactic parsers tend to make mistake with prepositional attachment. According to the semantics of the sentence, the phrase with D should not be attached to the main verb, but it is frequently interpreted that way by the parser (115 occurrences in our data). We have no choice but to add an additional rule to catch these cases.
The order in which rules are applied is important. Negation is the first pattern that must be tested, since in this case no description must be extracted. For example, when sentence Flovent Diskus is not indicated for the relief of acute bronchospasm is processed, it must be recognized that it is a negation and an empty set must be returned. Then rules are applied in order of specificity, with Rule Tr_4 (the most general rule) triggered by default when no other rule applies.
We wanted to evaluate the capacities of our system to discover a maximum of drug-disease facts. We also wished to evaluate if the rules actually allow to extract the desired information with accuracy. According to this, our evaluation is intended to measure two criteria, first coverage and second precision, according to a selected set of predefined rules applied to a reference corpus. To build our experimental protocol, we have manually devised 15 rules (as shown in table 3, inspired by statistics on our corpus (frequent patterns as presented in section 1) and measured how many examples of the reference corpus we correctly found with these rules. 5.1
Coverage evaluation To get an idea of the overall coverage of our extractor, we selected from Dailymed all sentences that contain the verb indicated. This represents 5325 sentences. For 3580 of these sentences (67%), a pattern has been recognized. With this small manual effort (only 15 rules), we cover most of the relevant patterns found in the corpus. We look further to find out what sentences were left out.
First, we find that many sentences are actually irrelevant for our purpose, such as the following ones: The routes of administration and indicated concentrations for mepivacaine are... Amiodarone also can be used to treat patients with VT/ VF for whom oral amiodarone is indicated, but who are unable to take oral medication.
Renal function studies should be performed when indicated.
Also, we see that some cases require anaphora resolution, which is outside the scope of our research: they are also indicated for use in secondary amenorrhea. But mostly, which might be surprising within such a restricted corpus, the syntactic variations are incredibly high. Consider for example the following sentences, which are not correctly recognized by our patterns: INDOCIN I.V. is indicated to close a hemodynamically significant patent ductus arteriosus in premature infants weighing between 500 and 1750 g when after 48 hours usual medical management (e.g., fluid restriction, diuretics, digitalis, respiratory support, etc.) is ineffective. Vinorelbine is indicated as a single agent or in combination with cisplatin for the first-line treatment of ambulatory patients with unresectable, advanced nonsmall cell lung cancer (NSCLC).
Then, within the 3580 sentences covered, we look at the distribution among our rules. Last column of table 3 provides, for each rule of our extractor, the number of times it has been used when processing the whole corpus. Note that the total of rules used is higher than the total of sentences for which a description has been extracted. The reason for this is that, as we have seen, some rules require the recursive application of other rules. An important observation is the necessity of detecting negative forms: the second most used rule is the one that recognizes this form. Also, considering the distribution, we see that even if the default rule is by far the most used (about 77%), the importance of other rules is not negligible. And, as we will see later, there are many cases where Rule Tr_4 was activated erroneously, that is, a specific case has not been recognized by one of the other rules and should have been. Now, considering rule groups, we see that none seem to be irrelevant, with the exception of rules for combination therapy. In this last case, it appears that almost all instances are detected by the Drug() extraction function (62 occurrences). 5.2
Precision evaluation To evaluate the precision of our extractor, we manually looked at negative and positive instances. First, we randomly selected a set of 100 sentences for which no description have been generated (negative instances) by our extractor. For each sentence, we determined whether it should have been covered by our extractor (false negatives), that is, whether it does not fall out of the extractor’s vocabulary and describes the kind of treatment that is extracted by our rules. We did the same with sentences for which an RDF description has been generated (positive instances): we randomly selected 150 of them and determined, for each one, if the extracted RDF description is correct (true positives). We computed precision (P) and true negative ratio (TNR), using the following formulas (TP = true positives, FP = false positives, TN = true negatives and FN = false negatives): P = T PT+PF P
T NR = T NT+NF N
Results are given at Table 4. Note that for precision we provide two measures called Strict and Relaxed. In the strict evaluation, a description is considered correct if it contains all the expected triples. In the relaxed evaluation, we accept partial descriptions. For example, if the sentence refers to a palliative treatment and the extractor returns the description of a generic treatment, it would be accepted. It may also be the case that the drug is indicated for two diseases (denoted by a coordination in the sentence), and only one appears in the description. Table 4b shows the results for each rules individually (values do not sum up to 150 because more than one rule may be applied to the same sentence).
TP FP Precision 53 22 0.67 (strict evaluation) 64 11 0.85 (relaxed evaluation) TN FN True negative ratio 79 21 0.79 (a)
Rule group # occ. Precision Tr 122 0.64 (0.83) Adj 9 0.89 (0.89) Sympt 8 1,00 (1.00) Cont 8 0.8 (0.89) Red 6 0.83 (1.00) Comb 5 0.80 (1.00) (b)
The results show that precision is high for both strict and relaxed measures, with the exception of Tr and Cont rules, where values are low for strict evaluation. It is important to remember that cases rejected by strict evaluation are valid representations. Their only problem is that some information is missing. 6
We presented a syntactic-based method for knowledge extraction. We looked at one type of information, the object literal of the indication predicate in DailyMed dataset. Analysing this data with simple 5-gram frequency analysis, we discovered variations in types of treatment and drug therapy which we captured with variations in our knowledge schema. We developed rules based on dependency trees to extract the required information from the long literal providing this information in natural language text. We showed high precision results on the extracted knowledge.
This present work is in contrast to our previous work investigating shallow
textual analysis methods applied to the Web at large, in search of drug
indications[
Spending the time to refine a RDF structure of a dataset in any domain can be valuable to the Semantic Web community. Especially if such dataset contains information found from trusted sources and is only partially RDFized. The value of the Semantic Web is in its capability of sharing and linking information. This assumes breaking down the information to the notion of concepts (URIs) and labels for such concepts. We see our research as giving a method for analysing textual data, promoting the idea of taking the time to define specialized schema and specialized extraction rules which can then accelerate largely the extraction of knowledge from the dataset.
There are many paths for future work. One is to use the same approach on other predicates in other domains. Another is to go back to knowledge extraction in larger, noisier data, and exploit the benefit of knowledge learned from smaller focused data to help the knowledge extraction process.