In 2015, a survey of Chinese internet users showed that medicine and health care are the two most popular searched topics and accounted for 55.15% of all searches [1]. However, it is difficult for most users to express medical concepts using professional terms such as bronchus, brain, and extracellular space [2-4], and online forums and news media explain such professional medical terms with very little detail. The gap between consumer language and medical terminology makes searching and retrieving information difficult and biases the understanding of health information [5-7].

Development of the consumer health vocabularies, which map languages of consumers and medical experts, is a potential way to bridge the gap. Several commercial and noncommercial groups, such as Apelon and Public Health Terminology by Intelligent Medical Objects, and Open Access and Collaborative Consumer Health Vocabulary [8], tried to bind consumer health vocabularies with the Unified Medical Language System (UMLS) or the International Classification of Diseases (ICD). Several factors dominate the expansion of a quality consumer health vocabulary: a comprehensive search to identify related nonstandard expressions, abbreviations and common typos, a consensus between consumers’ point of view and professional classification, and periodic updating for new terms. These factors make the expansion process complicated, costly, and time-consuming.

To accelerate the expansion process, researchers developed many approaches to extract and map consumer terms automatically or semiautomatically, including the n–gram-based approach [9], pattern-based approach [7], co-occurrence analysis [10], and machine learning methods [9,11]. Although the consumer health vocabularies mined through these hand-crafted heuristic approaches are more accurate, many relevant pairs could be missing. Recent theoretical and experimental results from Wang et al [12] showed that matching professional-consumer concept pairs through text embedding approaches can capture the semantic similarities between professional concepts and consumer concepts, thus yielding a high recall. However, with only unsupervised algorithms, many irrelevant pairs could be generated. In order to retain the advantage of text embedding and improve precision as much as possible, we propose a semisupervised representation learning method to make the concept embedding specific in the consumer vocabulary mining process. With the knowledge introduced by the reviewer, concept embeddings can continuously improve themselves. Our approach provides a related consumer term list sorted by their semantic distance to a particular medical term and helps reviewers identify synonym pairs efficiently. We extracted consumer health terms from one of the most popular health forums in China and manually evaluated the performance of our approach. The experimental results are promising, showing performance improvement of up to 16% with a small amount of seed pairs.

Two mainstream approaches for identifying synonyms are rule-based algorithm and word similarity measurement. A rule-based algorithm identifies synonyms by semantic patterns. For example, Vydiswaran et al took advantage of common linking phrases such as “also called,” “also known as,” and “also referred to as” to extract synonyms from Wikipedia [7]. There are many ways to calculate word similarities, including n-gram, edit-distance (Levenshtein distance), WordNet-distance, and cosine-distance between word vectors [13-16]. Among them, training distributed word vectors and extracting synonyms from top closest words are the most popular ways. Henriksson et al created word vectors using latent semantic analysis with random indexing and permutation to identify medical synonyms and abbreviation-expansion pairs [17,18]. He et al created word vectors including linguistic, contextual, and statistical features and used K-means to gather new consumer health terms on social media [19]. Elhadad et al created word vectors combining both contextual and semantic features and cluster terms from breast cancer forums into predefined semantic categories [20]. Wang et al created word vectors using word2vec, an open-source natural language processing tool released by Google, to extract symptoms from UMLS [12].

As early as 1998, Marshall [21] mapped the consumer health terminology from WellMed (a health care website) to SNOMED (Systematized Nomenclature of Medicine) and UMLS, which helped patients search information using nonprofessional expressions. In 2001, Patrick expanded UMLS, the Eurodicautom of the European Commission’s Translation Service, and the European Commission Glossary of popular and technical medical terms, by adding words from the Dictionary of American Regional English, but only focused on diabetes-related terms [22]. Both Marshall and Patrick constructed their consumer health vocabularies manually, which is inefficient and unscalable. In 2005, Zeng developed a two-step approach, which combined corpus-based text analysis and manual review, to build an open-source consumer health vocabulary [23]. To reduce the labor in term mapping, Zeng improved the two-step approach by adding n-gram, logistic regression, and even natural language processing and machine learning algorithm (parts of speech, noun phrases, and named entities recognition) [9-11]. Since then, the two-step semiautomatic approach—term identification algorithms followed by manual review—evolved into a common practice in many consumer health vocabulary researches [7,10,24].

To alleviate the problems mentioned above, we propose a consumer health term–finding framework based on a distributed word vector space model. The overview of the framework is shown in Figure 1. The workflow can be interpreted as a feed-forward neural network. The first step of our approach requires a corpus of raw text for the unsupervised pretraining of the embedding matrix E as the embedding layer (Figure 1). Text embedding approaches have proven to be very effective in capturing the similarities between words and phrases, which can yield professional-consumer concept pairs that cannot be found by feature- and pattern-based methods.

We used THULAC (Tsinghua University - Lexical Analyzer for Chinese) [25], a Chinese segmentation tool, to change Chinese text into words. Thereafter, word embedding tools were used to compute the vector space, as described above. We collected professional concepts from the official Chinese version of ICD-10, calculated the frequency of words in the corpus, and extracted those with a count of over 1000 and their corresponding consumer concepts with context as seed pairs. All the weights were initialized uniformly at random. In order to obtain a model that takes professional concepts as input and consumer concept list as output after training, we introduced an embedding space–adapting process consisting of an embedding projection and a supervised ranking method. The embedding projection contains a projection layer, a hidden layer, and a target projection layer (Figure 1) to achieve a “smaller” embedding space that preserves more supervisory signal.

The output concept list is not necessarily the most similar word of the input concept in the word vector space, but with the supervised signal of seed terms, their similarity is more of a measure of professional-consumer concept pair similarity. A special ranking loss function is used in the ranking loss layer to calculate the similarity of professional-consumer concept pairs. After manually selecting the output professional-consumer concept pairs, we input the selection results into the training data. New professional-consumer concept pairs were discovered through iterations.

Word embeddings are generally trained to reconstruct linguistic contexts of words by optimizing an objective function that can be measured without annotations. One popular approach is to estimate the embeddings by maximizing the probability that the words within a given window size are predicted correctly. Word embedding takes a large corpus of text as its input and produces a vector space, with each unique word in the corpus assigned a corresponding vector in the space. In word embedding training, one of the key issues is the formulation of the training objective function, minimization or maximization of which may produce meaningful word vector representations. Ideally, the training objective function should reflect the fact that the semantic word similarities measured on learned word vectors are consistent with human cognition. Recently Wang et al [26] performed a comprehensive comparative study on the different word embedding techniques in biomedical texts. Of those, we chose three popular word embedding methods: Word2Vec [27], Global Vectors (GloVe) [28], and FastText [29].

As mentioned in the Introduction, word embedding is a useful unsupervised technique to capture the similarity between words and phrases, which can yield high recall of professional-consumer concept pairs. It can also be used as a pretraining phase prior to supervised training. However, even if the embeddings provide compact, real, valued representations of each word in a vocabulary, it only indicates that word embeddings produce a semantic space that models synonymy to a certain degree. Current methods use pretrained embedding to initialize model parameters and then use the labeled data to guide them for the intended task (eg, we use professional-consumer concept pairs that already exist as the supervision to produce a semantic space dedicated to finding such pairs). If, as in our case, only a small amount of supervised data are available, this can lead to severe overfitting. Furthermore, rare words will receive very few updates and their embedding will be poorly adapted for our task. We propose two solutions to avoid these problems.

To evaluate the effectiveness of the proposed model, three groups of experiments were designed. The three kinds of word embeddings with different vector size were further trained by the proposed model and evaluated. The baselines were the original word embeddings described above. The effect of the projection layer was studied in the second group experiments. Two comparison groups were involved, one that used the standard structure without the projection layer and another that used the proposed projection layer.

We tested our methods with the corpus obtained from two different Chinese communities to cover different perspectives. The Tianya community is one of the most popular online forums in China, and the data are open and easy to retrieve. The health care sector of the Tianya community—Tianya Hospital—has a large number of disease consulting posts initiated by consumers, and about 180 Mb of data are used in our experiment. Haodf is the largest Chinese medical question-and-answer website where all questions are created by patients and answered by doctors, and about 2 Gb of data are used in our experiment. We considered these to be ideal sources of consumer health corpus, used Scrapy [30] for a full-text crawling from those corpora, and removed user information before further processing. The messages from the consumer forums were preprocessed as follows: URLs were replaced with a token URL and words occurring less than 30 times in the corpus were replaced by a special UNKNOW symbol. We collected professional concepts from the Chinese official version of ICD-10, calculated the frequency of words in these two corpora, extracted them with a count of over 1000, and manually collected their corresponding consumer concepts with context as seed pairs. The annotation process was performed by one medical professional and reviewed by three medical professionals. We finally obtained 224 seed pairs that all three reviewers consistently agreed upon.

We use the mean reciprocal rank to evaluate the quality of word embeddings. Mean reciprocal rank is a statistic measure for evaluating any process that produces a list of possible response to a sample of queries and orders them by probability of correctness. The mean reciprocal rank is the average of the reciprocal ranks of results for a sample of queries Q: where rank_i refers to the rank position of the first relevant synonym for the i-th query.

We use a large collection of candidate medical concepts and build a small set of ground truth professional-consumer concept pairs. We randomly select 100 pairs from our seed pairs for evaluation.

In general, the performance of the proposed method is detailed in Table 1. All word embeddings are significantly enhanced after fine tuning. The performance of the best word embedding is FastText, with a 400-dimensional vector with Haodf and projection matrix size set to 40, and it is also significantly improved in all datasets. The rich n-gram features used in FastText are important in Chinese synonym finding and have much higher performance than others. These remarkable improvements demonstrate that our method may transfer the complementary knowledge from the weak embeddings into the strong embeddings.

Table 2 shows the system performance with no projection matrix and different projection matrix size. As baselines, we considered a simple log-linear approach, which uses the unsupervised embeddings directly as features in a log-linear classifier. We tested the model performance. Furthermore, we observed that updating the embeddings always led to inferior results. This suggests that pretrained embeddings should be kept fixed, when little labeled data are available to retrain them.

In order to ensure the accuracy of professional-consumer concept pairs, manual review is inevitable. Table 3 showed the top 10 candidates for word “diarrhea” (“腹泻”) provided by our method for reviewers. Most words illustrated here are symptoms or clinical findings related to “diarrhea,” such as “vomiting” (“呕吐”), “abdominal pain” (“腹痛”), and “dyspepsia” (“消化不良”). We see two synonyms in the table: “having loose bowels” (“拉肚子”, ranked third) and “diarrhea” (“腹泄”, ranked seventh). The former is a consumer health term that is rarely used by professionals, and the latter is a typo of “diarrhea” (“腹泻”). In the manual review, researchers reviewed a sample of the candidate terms suggested by the system to assess whether these terms should be added into the consumer health vocabularies.