Apache UIMA ConceptMapper Annotator Documentation

The requirements on the design of the ConceptMapper dictionary were that it be easily extensible and that arbitrary properties could be associated with individual entries. Additionally, the set of properties could not be fixed, but rather customizable for any particular application.

The structure of a ConceptMapper dictionary is quite flexible and is expressed using XML (see Example 2.1, “Sample dictionary entry”). Specifically, it consists of a set of entries, specified by the <token> XML tag, each containing one or more variants (synonyms), the text of which is specified using by the "base" feature of the <variant> XML tag. Entries can have any number of associated properties, as needed. Individual variants (synonyms) inherit features from their parent token (i.e., the canonical form), but can override any or all of them, or even add additional features.

In the following sample dictionary entry, there are 6 variants, and according to the rules described earlier, each inherits the all attributes from the canonical form (canonical, CodeType, CodeValue, and SemClass), though the variants “colonic” and “colic” override the value of the POS (part of speech) attribute:

			<token canonical="colon, nos"
			       CodeType="ICDO" CodeValue="C18.9"
			       SemClass="Site" POS="NN">
				<variant base=”colon, nos”/>
				<variant base=”colon”/>
				<variant base="colonic" POS="JJ" />
				<variant base="colic" POS="JJ" />
				<variant base="large intestine" />
				<variant base="large bowel" />
			</token>

The result of running ConceptMapper are UIMA annotations, and there are two configuration parameters that are used to map the attributes from the dictionary (see AttributeList) to features of UIMA annotations (see FeatureList).

The entire dictionary is loaded into memory, which, in conjunction with an efficient data structure, provides very fast lookups. As stated earlier, dictionaries with millions of entries have been used without any performance issues. The obvious drawback to storing the dictionary in memory is that large dictionaries require large amounts of memory; this is partially mitigated by the fact that the dictionary is implemented as a UIMA shared resource (see DictionaryFile). This means that multiple annotators, such as multiple instances of ConceptMapper that are set up using different parameters, can all access it without having to load it more than once. The dictionary loader is specified in the external resource section of the descriptor, and is expected to implement the interface org.apache.uima.conceptMapper.support.dictionaryResource.DictionaryResource. Two implementations are included in the distribution, org.apache.uima.conceptMapper.support.dictionaryResource.DictionaryResource_impl, the standard implementation, which loads an XML version of a dictionary, and org.apache.uima.conceptMapper.support.dictionaryResource.CompiledDictionaryResource_impl which loads a pre-compiled version, for faster loading. The compiler is supplied as org.apache.uima.conceptMapper.dictionaryCompiler.CompileDictionary, which takes two arguments, a ConceptMapper analysis engine descriptor that loads the dictionary using the standard dictionary loader, and the name of the output file into which to write the compiled dictionary.

Input documents are processed on a token-by-token basis, so it is important that the dictionary entries are tokenized in the same way as the input documents. To accomplish this, ConceptMapper allows any UIMA analysis engine to be specified as the tokenizer for the dictionary entries. See parameter TokenizerDescriptorPath for details.

As stated earlier, input documents are processed on a token-by-token basis. Tokens are processed one span (e.g., a sentence or a noun phrase) at a time. Token annotations are specified by the parameter TokenAnnotation, while span annotations are specified by the parameter SpanFeatureStructure. By default, all tokens within a span are considered, and it is the text associated with each token that is used for lookups. ConceptMapper can also be configured to consider tokens differently:

The ability to skip tokens during lookups based on particular feature values makes it easy to skip, for example, all tokens with particular part of speech tags, or with some previously computed semantic class. For example, given the text below in Example 2.2, “Sample Input Text”:

				Infiltrating mammary carcinoma

Assume each word is a token that has a feature SemanticClass, and that feature for the token “mammary” contains the value “AnatomicalSite”, while the tokens “Infiltrating” and “carcinoma” do not. It is then possible to configure ConceptMapper to indicate that tokens that have a particular feature, in this case SemanticClass, equal to one of a set of values, in this case “AnatomicalSite”, should be excluded when performing dictionary lookups (see parameters ExcludedTokenClasses and ExcludedTokenTypes). By doing this, for the purposes of dictionary lookup, the example text would effectively appear to be:

				Infiltrating carcinoma

In addition to the set of feature values that indicate their associated token are to be excluded during lookup, there are also configuration parameters that can be used to specify a set of feature values for inclusion (see parameters IncludedTokenClasses and IncludedTokenTypes). The algorithm for selecting annotations to include during lookup is as follows:

				if there is an includeList but no excludeList
				include annotation if feature value in includeList

				else if there is an excludeList
				exclude annotation if feature value in excludeList

				else
				include annotation

This provides a simple way to restrict the selection of pre-classified tokens, whether that pre-classification is done via previous instances of ConceptMapper or some altogether different annotator. See TokenTextFeatureName

The actual dictionary lookup algorithm is controlled by three parameters. One specifies token-order independent lookup (OrderIndependentLookup). For example, a dictionary entry that contained the variant:

			<variant base='carcinoma, infiltrating'/>

would also match against any permutation of its tokens. In this case, assuming that punctuation was ignored, it would match against both “infiltrating carcinoma” and “carcinoma, infiltrating”. Clearly, this particular setting must be used with care to prevent incorrect matches from being found, but for some domains it enables the use of a more compact dictionary, as all permutations of a particular entry do not need to be enumerated.

Another parameter that controls the dictionary lookup algorithm toggles between finding only the longest match vs. finding all possible matches (FindAllMatches). For the text:

			... carcinoma, infiltrating ...

If there was a dictionary entry for “carcinoma” as well as the entry for “carcinoma, infiltrating”, this parameter would control whether only the latter was annotated as a result or both would be annotated. Using the setting that indicates all possible matches should be found is useful when subsequent disambiguation is required.

The final parameter that controls the dictionary lookup algorithm specifies the search strategy (SearchStrategy), of which there are three. The default search strategy only considers contiguous tokens (not including tokens from the stop word list or otherwise skipped tokens, as described above), and then begins the subsequent search after the longest match. The second strategy allows for ignoring non-matching tokens, allowing for disjoint matches, so that a dictionary entry of

			    A C

would match against the text

			    A B C

This can be used as alternative method for finding “infiltrating carcinoma” over the text “infiltrating mammary carcinoma”, as opposed to the method described above, wherein the token “mammary” had to have been have been somehow pre-marked with a feature and that feature listed as indicating the token should be skipped. On the other hand, this approach is less precise, potentially finding completely disjoint and unrelated tokens as a dictionary match. As with the default search strategy, the subsequent search begins after the longest match.

The final search strategy is identical to the previous, except that subsequent searches begin one token ahead, instead of after the previous match. This enables overlapped matching. As with the setting that finds all matches instead of the longest match, using this setting is useful when subsequent disambiguation is required.

Output is in the form of new UIMA annotations. As previously discussed, the mapping from dictionary entry attributes to the result annotation features can also be specified. Given the fact that dictionary entries can have multiple variants, and that matches could contain non-contiguous sets of tokens, it can be useful to be able to be able to know exactly what was matched. There are two parameters that can be used to provide this information. One allows the specification of a feature in the output annotation that will be set to the string containing the matched text. The other can be used to indicate a feature to be filled with the list of tokens that were matched. Going back to the example in figure 2, where the token “mammary” was skipped, the matched string would be set to “Infiltrating carcinoma” and the matched tokens would be set to the list of tokens “Infiltrating” and “carcinoma”.

Another output control AE descriptor parameter can be used to specify a feature of the resultant annotation to be set to contain the span annotation enclosing the matched token. Assuming, for example, that the spans being processed are sentences, this provides a convenient way to link the resultant annotation back to its enclosing sentence.

It is also possible to indicate dictionary attributes to store back into each of the matched tokens. This provides the ability for tokens to be marked with information regarding what it was matched against. Going back to the example in figure 2, one way that the SemanticClass feature of the token “mammary” could have been labeled with the value “AnatomicalSite” was using this technique: a previous invocation of ConceptMapper had “mammary” as a dictionary entry, that entry had the SemanticClass feature with the value “AnatomicalSite”, and SemanticClass was listed as an attribute to write back as a token feature. If, instead of “mammary” the match was against a multi-token entry, then each of the multiple tokens would have that feature set.