Text is an important resource for health-related information as it is easy to create and distribute. Furthermore, health literature is widely available in the form of web-based resources for people to obtain information on medical conditions, diseases, and modalities [1]. However, these documents are often written at a level beyond the comprehension of the average reader [2]. This disconnect reflects an overall trend in misinformation regarding health conditions [3,4].

To mitigate this problem, researchers have sought automatic ways to improve the readability of these texts and the resulting reader comprehension. This natural language programming (NLP) task is known as text simplification [5] and has been used to create supervised [6], semisupervized [7], and fully automatic tools [8] to make texts easier for consumers to digest by increasing readability [9]. A central challenge for this research is to develop resources and techniques that enhance the quality and accuracy of these systems. Even though deep neural network approaches and other automated translation algorithms are increasingly being developed, it will take time before they can be applied with sufficient impact and precise simplifications. We intend for our algorithm to supplement existing resources as well as generate useful inputs for other algorithms.

The first step is identifying what makes text difficult. Some of the previous studies have focused on simplifying individual terms, while others have focused on grammatical structures. To identify the difficulty of individual terms, we use term familiarity. For a given term, this measure can be calculated by extracting the likelihood that a term occurs in common language usage [10], which we estimate according to the term’s frequency in the Google Web Corpus [11]. In this work, we add to the body of research that identifies these terms and replaces them with easier synonyms [12]. However, we went beyond existing approaches by generating new explanations for terms that do not exist in the available resources. To do this, we developed and evaluated a new algorithm to generate new explanations. We generated explanations of terms in plain language using word internal parsing and affix dictionaries with SubSimplify.

Ideally, there would be an endless resource of expert-written explanations for difficult terms, optimized for the general public in multiple languages. However, few resources are able to provide appropriate explanations at all and even fewer are able to automatically or semiautomatically produce such explanations.

The resource closest to ideal is the English Consumer Health Vocabulary (CHV) [13], which is included in the Unified Medical Language System (UMLS) [14]. This resource was manually created and provides synonyms as well as definitions for medical terms in a consumer-friendly language. For the purposes of text simplification, these plain language definitions and simple synonyms double as ready-made explanations for difficult terms. However, the number of explanations is low relative to the overall number of difficult terms that occur in a given medical text. The CHV contains 2567 unique definitions and 88,529 synonyms for concepts found in the UMLS. We did not employ the UMLS as a resource because this system focuses on mapping complex medical concepts onto ontologies and is not designed to relate health information to patients or any other person outside the medical domain.

Previous research has shown that the CHV can be used to simplify texts [15-17], but it has also been shown to contain jargon words and not enough consumer-friendly vocabulary when providing summaries for specialized research [18]. Furthermore, while this resource is well tailored to text simplification, it is limited to English terms and explanations. In summary, the CHV provides explanations that can be automatically sourced in a given simplification system. However, CHV is only in English, is for relatively few terms, and can at times contain jargon beyond the reading level of the average reader.

While not being medically focused, WordNet is a useful resource for text simplification. It is an online lexical database containing terms and definitions, as well as interword semantic relations such as hypernyms, hyponyms, synonyms, and antonyms [19]. WordNet provides 128,391 word-sense definitions in English and is also available in Spanish, albeit in a less complete form [20]. Since WordNet is not a medical resource, many of its explanations are not optimal for medical text simplification, and when several senses are provided for a word, it is not always clear which best suits the medical sense. Previously, WordNet has been used to provide synonyms for lexical simplification [21]. For example, hyponym-hypernym relations have been used to generate synonyms that are simpler (more general) for text simplification [22]. In other areas, this resource has been used to simplify texts in the domain of biomolecules [23] and in texts written for non-native English speakers [24]. In summary, while WordNet is larger than CHV and also available in Spanish, the resource is not always optimal for giving the definition for medical terms.

Recent developments using neural networks trained on large bodies of text have produced larger resources such as word embeddings, where words are represented by multidimensional word vectors. The resulting vectors position the word relative to each other in a multidimensional space and have been shown to possess semantic and syntactic relations that allow us to automatically find synonyms and semantically related terms [12]. Given a word, we can use its vector representation to find the word whose vector is nearest to this word. Often, this nearest vector is a synonymous word. One freely available version of this resource is the pretrained Global Vectors for Word Representation (also known as GLoVe) [25]. Prior work has shown that these vectors can prove to be useful to isolate simple yet more frequent terms in the areas of text simplification [26]. However, they can include spurious matches because the approach cannot differentiate antonyms from synonyms. Given that this resource is totally automated, a word vector model can be produced from any language given a relatively large body of text. This means that this resource is also available for Spanish, with pretrained vectors available online [27]. In our study, we employed the GLoVe pretrained vectors for English and for Spanish [27], labeling the approach more generally as Word Embedding Vectors (WEV).

In all, the methods that exist for explanation generation range from specific, and precise, with low coverage to high coverage, with a much lower relative accuracy. In the next section, we describe our approach, which exists on the spectrum between these resources.

We first describe the role that morphological units play in medical terminology and then our algorithm, which extracts information and generates explanations using these morphological units.

The resources described above make use of a word’s definition in isolation without reference to the internal characteristics of that word, (ie, the morphology of the word). While it is not always the case, often romance languages contain morphological units that contain relatively clear semantics, such as the case for the prefix anti- (“against”), or the suffix -s (indicating plural). In certain words in English and in Spanish, these can help one to decipher the meaning of a word. In medicine, many terms, both in English and Spanish, originate from Greek and Latin [28,29]. Greek and Latin affixes have meanings commonly unknown to the average reader, but they nevertheless reflect the overall meaning of a word. While at times the meaning of a word is a direct function of the composition of the meaning of these morphological units, to a large degree in English and Spanish, terms composed of these units tend to have a gestalt effect. On the extreme end, a term may completely differ from the meaning of its morphological units (eg, “ledger” does not mean “ledge”+er). However, this problem of semantic drift is small for medical terms, seemingly because medical terms are less affected by semantic drift than more nonmedical, frequent terms.

Affixes that compose medical terms commonly have clear definitions that reflect a word’s meaning. For example, given the prefix cardio- we know that this term’s meaning relates to “the heart.” Several resources containing these affixes and their definitions are freely available online [30-33]. From these, we created a unique dictionary of affixes along with their definitions for each language. We extracted 586 unique affixes for English and 498 affixes for Spanish. We define an affix as any morphological unit that has some denotation apart from the word itself. Affixes are categorized by their position, with prefixes occurring at the beginning of the words and suffixes occurring at the end of words. A root is any morphological unit that can stand alone as a single word. For example, the term cardiovascular contains the prefix cardio-, and the root vascular. Independently, these morphological units denote the heart and consisting of a vessel or vessels, respectively. Although many resources may not contain a definition for cardiovascular, by parsing these morphological units, we can automatically generate an explanation that reflects the actual denotation of the term: relating to the heart and blood vessels.

In both Spanish and English, words may be composed of multiple suffixes, roots, and prefixes. SubSimplify exploits this fact to generate an explanation for a term. Table 1 shows the examples of affixes and their definitions in both Spanish and English.

In addition to these affix dictionaries, we use word stemming [34] to isolate stemmed, or lemmatized, versions of terms. Stemming and lemmatization are two different methods of reducing a term to something similar to its root, but in a way that does not always reflect the actual root. For example, a resource like WordNet may have a definition for Gastrointestine, but not Gastroinstestinal. By stemming and stripping the affix -al, we increase the ability to find explanations using all resources.

Figure 1 provides an overview of our SubSimplify algorithm. The input to SubSimplify is a term we assume to be difficult, and we recursively lookup affixes and generate an explanation by accumulating the definitions of each affix and root identified. When finished, we align these definitions to provide an explanation of the term.

We use affix dictionaries to identify morphological units programmatically. First, the system identifies affixes and then takes the part of the word that is not an affix and performs a database lookup on stemmed variants of the term. To avoid spurious matches, we work from larger to smaller suffixes, and thus, anti- as in anti-hero would match before a- as in a-symmetry. This process occurs iteratively until no affixes are matched, or until there is no root left. In order to describe this process in sequence, Textbox 1 gives a detailed description of each step.

Since words may contain multiple suffixes, the process occurs multiple times where possible. That is, when we extract a root, it is possible that that root may yet contain another suffix or prefix. To highlight this, we provide an example with the term hyperglycemic in Textbox 2.

This process repeats until there is either no root left, or until the remaining root fails to be identified by any resource. For glycemic, the system will identify -ic and subsequently glyc- before halting at -em-.

If the term contains “-” or any other nonword characters, we split these as well. The parsed affixes and roots are then aligned with their explanations to provide an affix-by-affix breakdown of the term. For any affix that is not identified in the system, as is the case with -em- in hyperglycemic, the definition of the root remains the root itself. Upon presenting the term, these affixes are matched with their definition both by order and by color in order to make identification as easy as possible for a writer. An example explanation for hyperglycemic is shown in Table 2. This table contains explanations for a few different difficult terms to highlight their quality when present. Note that not all resources contain explanations for all terms, so it is extremely rare that all resources can provide an explanation for a single term.

While the CHV [35] and WordNet Summary resources provided full-sentence explanations (semantic simplification), the WEV and WordNet Synonym provide single-word explanations of each term (lexical simplification). SubSimplify provides a hybrid of the two: for the individual parsed subword units, either a synonym or brief description is presented.

Next, we describe 2 studies designed to evaluate the quality, coverage, and usefulness of these explanations in English and Spanish.

To evaluate the quality, coverage, and usefulness of the newly generated explanations and how they compare to existing resources, we conducted two studies: one in English and one in Spanish.

The aim of the English and Spanish studies was to evaluate the efficacy of employing SubSimplify to medical texts, and the results revealed what was expected. Compared with WordNet and CHV, the quality of explanations was, on average, lower. This may be an indication of a few different issues, which we will expand upon in the Limitations.

Surprisingly, the fully automatic system of WEV outperformed our expectations. In creating SubSimplify, we imagined that there would be many spurious matches and synonyms that were unrelated to the difficult terms, but the results showed a better performance than expected. Based upon this, we are motivated to employ WEV as another way to source root synonyms in SubSimplify. That is, in its current form, SubSimplify performs a term lookup in WordNet (and CHV, in English) after parsing each affix. Based upon these results, we are motivated to have the system perform a WEV lookup at this stage as well.

One challenge with SubSimplify is to maximize the understandability of the explanations themselves. While the existing resources contain a single definition of the term, SubSimplify relies upon combining multiple definitions to sew together a single explanation. Currently, the system employs a color-coding scheme that relates the morphological units on one line to their definitions on another. This may make it difficult to read for some people, effectively lowering the quality when people are not used to seeing these definitions. In order to allay this issue, our team plans on implementing a few different formats and testing them out in the near future in an interactive Web program.

SubSimplify is naturally limited by two factors. First, not all difficult medical terms contain subword units, and additionally, not all subword units match or are totally accurate. This is because not all terms contain morphological units, and the system has no way of knowing where the characters within a word are actual examples of suffixes. Beyond that, even if they are correct matches of affixes, there is no guarantee that the actual meaning of the terms is directly reflective of the affixes that are found within. For example, it may provide a meaning for anti- and -bodies in antibodies. But the explanation against bodies does not reflect the actual meaning of the term.

Another possibility is that many of the explanations generated by SubSimplify can be incomplete. For example, it may provide a meaning for -al in distal but no meaning for dist-. The result would then be difficult for anyone to understand. Nevertheless, the system does provide a bridge between hand annotated and automatic texts and, therefore, should be subject to these sorts of exceptions and problematic cases.

Second, SubSimplify, by virtue of using WordNet and CHV, is limited as well to the coverage of those resources. However, we believe that this work presents useful addition to a system aimed at providing explanations for complex terms.

The niche of SubSimplify is to exploit the regularities of morphological units in medical terminology to provide a window into breaking down the jargon of difficult terms into digestible terms. SubSimplify will improve as the resources used to create it do. Furthermore, we want to look at multiword phrases as oftentimes they reveal the contextual meaning that a single-word context cannot provide alone. This approach is intended as an additional resource that one can add to other methods to automatically provide explanations for difficult texts. The explanations generated by this system greatly increase the number of difficult terms for which an easier alternative can be made available and, thereby, present an advance in the area of text simplification in the medical domain.