An NLP technique with a syntactic model of a sentence and a semantic model of the words in a sentence has a better chance of understanding relationships among the words in the sentence. For example, a noun phrase comprising an adjective followed by a noun signifies a relationship between the noun, which is the head of the phrase, and the adjective, which is the modifier. Precisely what that relationship is depends on the meaning of the words in the phrases. Consider the phrase "atrial fibrillation.'' The UMLS semantic type for "atrial'' is Body Part, Organ, or Organ Component, and the semantic types of "fibrillation'' are Disease or Syndrome and Sign or Symptom. An NLP application that modeled both the syntactic and the semantic information in this phrase could have a rule that stated: If a syntactic modifier has the semantic type Body Part, Organ, or Organ Component, and the head has the semantic type Disease or Syndrome or Sign or Symptom, the semantic relationship is Head-has-location-Modifier.
An application that validated whether a term mentioned in a report actually is a finding could benefit from modeling semantic and syntactic relationships. For instance, the NLM system FindX (Sneiderman et al., 1996) contains rules based on the semantic type of the words in a modifier-head relation to validate the finding. For example, one rule states: An abnormality or anatomical site modified by a SNOMED adjective is a finding, validating "chest clear to auscultation'' as a finding. Another rule says: A diagnostic or laboratory procedure modified by a SNOMED adjective or a numeric value is a finding. This rule correctly validates arterial blood gas as a finding in (24) and invalidates it in (25).
(24) Arterial blood gas 7.41/42/43/27
(25) We suggest arterial blood gas preoperatively.
Semantic modeling of syntactically related words can also be useful in understanding implicit information in a report. MPLUS (Medical Probabilistic Language Understanding System) is an NLP system that uses Bayesian networks to model the relationship between the words in a report and the ideas or concepts the words represent (Christensen et al., 2002). Figure 17.3 shows a simplified network for radiological findings. The syntactic parse helps determine which words in the sentence should be slotted together into the Bayesian network (i.e., which words are syntactically related). When a new phrase or sentence is slotted into the network, MPLUS can make inferences about the meaning of the words in a sentence in spite of the different combinations of words that can be used to describe the same concept. For example, the phrases "hazy opacity in the left lower lobe'' and "ill-defined densities in the lower lobes'' both indicate localized infiltrates—even though the word "infiltrate'' was not used by the radiologist.
Was this article helpful?