Week 4

Literature survey!

Exploiting Context for Biomedical Entity Recognition: From Syntax to the Web, by Jenny Finkel, Shipra Dingare, et al.

In this paper, they're doing biomedical named entity recognition by using some external resources like the internet, local features like previous and subsequent words and part of speech, and deep syntactic features. They use the Stanford Parser to look at the head and governor of each noun phrase. They say that this was useful for their task, which involved identifying five classes, but is not useful for NER when there are only two classes, such as for Biocreative. There are no data or citations provided to substantiate this claim, though.

Although this paper indicated that deep parsing does have a positive effect, it doesn't evaluate the effects of each individual addition in the data provided. It doesn't examine the effect of using deep parsing, but only the effect of using deep parsing, web resources, part of speech tagging, word shape, word length, etc, combined. So there's no way to tell specifically the advantage of deep parsing alone.

The value of parsing as feature generation for gene mention recognition, by Larry H. Smith, W. John Wilbur.

They test the usefulness of several deep parsers for feature generation. The way they evaluate this is by using only parse features generated by the rest of the sentence and not the features related to the candidate noun phrase itself, after using all the lexical non-parse information (word, part of speech, location) available. They used the following parsers: Charniak, Bikel, Rasp, Minipar Enju, Stanford, Rasp Stanford, and Charniak-Lease. These provided phrase structure trees, grammatical relations, predicate-argument structures, dependency relations, or annotated phrase structure trees.

They found that all nine of the parsers produced a small but statistically significant improvement, but there were no significant differences between the parsers. This paper is limited in that it only evaluates parsing for gene mention recognition, and not named entity recognition as a whole.

Parsing Biomedical Literature, by Matthew Lease and Eugene Charniak

In this paper, they compare a baseline statistical parser (F-measure: 78.3) to various lexically-adapted parsers. Part of speech tagging alone brings parsing accuracy up to 81.5%, although they experienced difficulties with the corpus tag set not matching their own tag set.

They then tested using a biomedical lexicon to increase accuracy. By never letting the parser cross-bracket the compound terms in the biomedical lexicon, it was hoped that it could fix a particular class of errors where boundaries were incorrectly identified. This produced a small but statistically significant result, but it was not significant when paired with part of speech tagging or their entity adaptations. This is attributed to the particular error it was being designed to fix being so infrequent.

They used the GENIA corpus, which has been labeled for named entity recognition, to generate features to identify NP boundaries: a complete reverse of what we're doing. This produced a result roughly equivalent in accuracy to part-of-speech tagging.

Approaches to Text Mining for Clinical Medical Records, by Xiaohua Zhou, et al.

They create a system to perform information extraction tasks on clinical records, making use of a parser. They perform three tasks: extraction of medical terms, relation extraction, and text classification.

For relation extraction (such as identifying "blood pressure" and "144/90" to be related), they used the Link Grammar parser. They viewed the link diagram as a connected graph, and weighted each edge. They then used the "distance" between words as a feature.

The authors only evaluated their system on records of subjects with breast complaints, and all were recorded by the same doctor. This lack of diversity may cause the data to misrepresent the results of using the system on medical records in general.

They noted several relevant limitations associated with using Link Grammar. For one, its lack of training on biomedical produces errors. It's also unable to parse sentence fragments. The latter is a significant problem for clinical records, but may not be very important for parsing biomedical literature. They also mentioned that Link Grammar cannot process multi-word terms.

Tackling the BioCreative 2 Gene mention task with conditional random ﬁelds and syntactic parsing, by Vlachos A.

Evaluated two things: using conditional random fields and using RASP parser to generate features. Determined that lemmas are the most useful, and POS tags and syntactic features are less so. He proposed that the reason for the lack of improvement is because few sentences have the specific error that syntactic features can correct.