A perfect match: locating plain text in HTML pages
SciLite is a Europe PMC tool that allows biological terms or relations, such as diseases, chemicals or protein interactions, to be highlighted for readers on abstracts and full text articles. These terms are identified as annotations by text mining algorithms, developed by a variety of text mining groups.
The main challenge for the SciLite tool is locating plain text annotations in HTML pages. The challenges derive from the nature of HTML pages. Below is a list of the major challenges we faced and the solutions adopted to mitigate them.
The pages contain HTML tags, obviously. For example, visit this article, and click on the “Gene Function” checkbox, on the right-hand side of the page, to see the sentence highlighted.
Figure 1: Annotation containing HTML tags
The problem is caused by the
<sub>tag that surrounds the character “v” inside the word “Nav1.7”. Thus if you search for an exact match of the plain sentence in the HTML page, it will not be found. Our solution was to search for a regular expression built to include an optional HTML tag between any two characters of the annotation text. The disadvantage of this approach is that this type of search is much more computationally demanding than an exact match search. Therefore, we decided to adopt this regular expression search only for sentence-based annotations, where the chance of having HTML tags is much higher than for named-entity annotations (usually composed of only one or two words).
HTML encodes special characters. An example is the character “>”. It is encoded as
>inside an HTML page. For example, visit this article, and click on the “Gene Disease Open Targets” checkbox.
Figure 2: Annotation containing HTML-encoded characters
The text of the annotation contains the character “>”. Our solution was to encode the annotation text as it would appear in an HTML page (replacing > with
>), and then perform an exact match search.
Lack of correspondence between the text of the annotation and text inside the HTML. For example, visit this article, and click on the “Gene Function” checkbox.
Figure 3: Annotation containing Greek characters
The original annotation text is “Our results revealed a direct interaction between PRL-3 and integrin beta1 and characterized Y783 of integrin beta1 as a bona fide substrate of PRL-3, which is negatively regulated by integrin alpha1.” The problem is that the Greek letters alpha and beta are represented in two different ways in the page and in the annotation text. A solution to this problem is applying a fuzzy match approach that is discussed later.
Special characters are not properly encoded inside the annotation text. For example, visit this article, click on the “Organism” checkbox, and focus on the annotation “white campion”.
Figure 4: Annotation containing not properly encoded characters
Every annotation comes with prefix and suffix text that help to locate it in the article page. The suffix of this annotation is “is subject to preâdispersal” with the character
ânot properly encoded. In this case as well, our solution was to apply our fuzzy match approach.
Fuzzy Match Strategy
The approach consists of the following steps:
- The fuzzy match algorithm compares the annotation sentences with the list of sentences inside the article. A similarity threshold is defined in order to retrieve only the article sentences that are similar enough to the annotation text. The choice of this threshold is crucial. If it is too restrictive, there is a chance that some true positive matches will be ignored, while if it is too lax, there is a chance that some false positive matches will be found.
- If there is at least one sentence considered similar enough to the annotation, the sentence with the highest similarity score is taken as a valid match.
As has been discussed previously, there are two type of annotations inside the SciLite platform:
- Sentence-based annotations.
- Named-entity annotations (usually consisting of one/two words) with a prefix and a suffix to locate them inside the article.
Because of the nature of the fuzzy match algorithm, it can be directly applied only to sentence-based annotations. Since the fuzzy match search is more computationally demanding than an exact search, we decided to adopt it only if the exact search for the sentence fails.
The fuzzy match has also proven useful for matching the prefix and the suffix of the named entity annotations, as described in problem number 4. Even in this case, we adopt a fuzzy match prefix/suffix search only if the exact search fails, to avoid computational overhead.
We have run some tests to compare the numbers of annotations matched with and without this fuzzy match approach. The sample was made of 8433 full text articles plus 8368 abstracts. The results are as follows:
|Fuzzy Match||Exact Match|
|Sentence based||91.62 %||79.1 %|
|Named Entity||92.11 %||86.93 %|
Table 1: Fuzzy match approach results
As expected, you can see that the benefits of the fuzzy match approach are more significant for sentence-based annotations.
Searching for plain text in HTML pages presents many challenges due to the nature of HTML rendering (tags, characters encoding, mismatched characters). One approach to solve these issues is to introduce techniques that apply some sort of fuzzy matching. However, those techniques can be demanding from a performance point of view, especially if the HTML pages are long and the number of annotations to locate is large. Therefore, it is necessary to carefully balance accuracy of results and performance, deciding when it is appropriate to apply those strategies.