Word embeddings map discrete word tokens to real-valued vector representations, where semantically similar words have similar vectors and are therefore closer in the embedding space. In general, pretraining of such word embeddings has been found to be beneficial for several natural language processing tasks, allowing for faster model convergence and task performance improvements. Therefore, we trained an unsupervised embedding model, FastText [68], with our full Twitter collection as training data and with the aforementioned preprocessing. Compared to other word representation models, FastText can produce word vectors for out-of-vocabulary words and has been proven to be a strong baseline for short text similarity, with its open-sourced implementation allowing for faster training [69]. More specifically, FastText produced 300-dimensional vector representations for each term in our vocabulary, which was used as the initialization for our model's embedding layer. We also experimented with Wikipedia-pretrained embeddings and without any pretrained embeddings: our experiments showed that the model performed better in terms of accuracy when initialized with HPV-related pretrained word embeddings.

Finally, we divided the annotated data into 60% training, 20% validation, and 20% testing, keeping the same splits across all models for a fair comparison. Deduplication of tweets with exact matches within each set left 3661 tweets in total (2142 for training, 758 for validation, and 761 for the test set). We experimented with several model architectures, including convolutional neural network (CNN) [70], bidirectional long short-term memory (BiLSTM), and traditional models, including support vector machine and Naive Bayes. We trained with cross-entropy, Adaptive Moment Estimation with a 10^–4 learning rate, 0.01 decay, and a 32 batch size for the neural models. Hyperparameter tuning was performed using the Tune library [71]. In Table 1, we report the mean and SD of the top-5 performing model variations. Our experimental evaluation showed that CNNs performed better than did the other models (see Figures 2 and 3 for respective confusion matrices and the area under the receiver operating characteristic curve comparisons between neural networks). Of the top-5 best performing models for either of the neural networks, the CNN required less training time than did the BiLSTM. The mean training time per epoch for the CNN was 11.5 ms (SD 1.09, minimum 16, maximum 16, median 12), whereas the mean training time per epoch for the BiLSTM was 51.3 ms (SD 34.07, minimum 14, maximum 88, median 81). Our best-performing CNN model had 256 convolutional filters, including –3 kernels of width (3,4, and 5) and rectified linear unit nonlinearities; a max pooling layer, a fully connected layer of 128 units with rectified linear unit activations and 0.1 dropout, and a final softmax output layer that produced the classification prediction.

To identify the risk perceptions attributed to the HPV vaccines, we first applied our classifier to a set of 291,037 tweets from which we are able to tag 124,031 as false tweets and 167,006 as true tweets. Using a dictionary of causal terms derived by Kayesh et al [72] for Twitter causality detection, we screened for tweets that contained at least 1 of these terms and kept tweets classified as false information if the classification confidence was at least 0.998, as this maintained high fidelity with our classifier. Thus, a total of 9352 tweets were used for the causal relationship mining process (Table 2). We then used a dependency parser for tweets to tag and merge multi-word expressions [73]. As tweets can have multiple utterances (ie, independent sentences or fragments), we kept the noun phrases that appeared with the causal cue regardless of whether they had a dependency related to the causal cue, which is in contrast to the work by Kayesh et al [72]. A candidate causal phrase is a set of terms pertaining to a tweet that contains a causal cue and precipitates the candidate effect phrase.

We could then compute the pointwise mutual information (PMI) for the causal set C = {c₁,...,c_m} and the effect set E = {e₁,...,e_m} where the candidate causal phrase, c_i and effect phrase, e_j, are sets that contain terms w_c∈V_c and w_e∈V_e, respectively. Here, V_c is the set of terms, noun phrases, and multi-word expressions derived from candidate causal phrases in the tweets (excluding terms with a minimum frequency of 1 and removing stopwords) and V_e is the vocabulary derived from the candidate effect phrases.

To compute the PMI for terms w_e ∈e_j and w_c ∈c_i we have,

We can apply Laplace smoothing to ensure the probability distributions are nonzero [74] and can compute the normalized pointwise mutual information (NPMI) [75] as follows:

The range of values of NPMI are from –1 to 1, where –1 means the terms never occur together, 0 means they are independent, and 1 is complete co-occurrence.

As our model was completely unsupervised and included retweets, tweet messages could become very redundant, but our method could detect many near-duplicate candidate effect phrases. To collapse these phrases, we clustered the terms using semantic similarity derived from embedding representations of the candidate effect phrases. In particular, we used the package HuggingFace [76] to acquire the sum of the last 4 layers of the bidirectional encoder representations from transformers model [77]. To compute the word embedding, we then averaged these word embeddings in the candidate effect phrase to produce the embedding vectors.

Density-based spatial clustering of applications with noise (DBSCAN) [78] was then used to cluster the candidate effect phrases. There were 2 parameters of importance: (1) reachability, which is the max distance between the 2 “points” to be considered in the same cluster; and (2) the minimum number of “points” to be considered clusters. By points here we mean an embedded real-valued vector representation of the effect sets. We set the reachability to 0.1 and the minimum number of points to 1, as we wanted to retrieve only the closest semantically similar words while maintaining meaningful clusters. Note that there are other alternatives to DBSCAN, such as ordering points to identify the clustering structure (OPTICS) [79] and hierarchical density-based spatial clustering of applications with noise (HDBSCAN) [80] for spectral clustering; however, we only needed to reduce the number of effects compared at query time, meaning that DBSCAN was sufficient. We then selected the cluster cores as representatives for each cluster to collapse the effects.

To identify the perceptions associated with different causal words, we formulated this as a retrieval problem. Given a causality-related query q, we ranked the associated effects by using the NPMI. To compute the scoring function, we used the following:

This scoring function computes the average NPMI for all pairs of terms in the query and candidate effect phrase. We can then compute the cumulative NPMI score for a category C_a of effect phrases as follows:

To validate the candidate causal mining approach, we took a lexicon pertaining to risk perceptions (ie, perceived effects) concerning HPV vaccines. The HPV-Vaccine Risk Lexicon (HPVVR) is a consumer-facing lexicon to capture how laymen describe their risk perceptions about HPV vaccines (including their perceived harms and benefits about HPV vaccines) [81]. The HPVVR was developed in 2 stages. The first stage involved adopting the risk expressions and HPV-vaccine–related consumer-facing vocabulary from the Department of Homeland Security Risk Lexicon, MedlinePlus Consumer Health Topic Vocabulary, and Consumer Health Vocabulary (Unified Medical Language System) [82,83]. The second stage was to extract layman language about the descriptions of risk perceptions based on the user-generated content (from randomly sampled user-generated content from 2013 to 2018, including from Twitter and Facebook) by 2 trained annotators (interrater reliability: Cohen's kappa coefficient (κ)=0.80). The HPVVR covers more than 200 terms or phrases across 29 categories of risk perception-related vocabulary.

This gold standard list of effects, G, was then matched with the effect set E. In particular, we defined a partial match to be present if some terms in the ground truth effect phrase were matched with some of the candidate effect phrase (ie, g∩e ≠ Ø). For example, we mined “prevent throat cancer” from the data, which is a partial match to “prevent cervical cancer” in the HPVVR. There were 2 other kinds of partial match. We defined a match to be proper if the candidate effect phrase was a more specific example of the ground truth effect phrase, g ∈ G. For example, we mined “early onset menopause” from the data, which is a proper match to “menopause” in the HPVVR. We considered a reverse match to be present if the candidate effect phrase was a more general form of the ground truth effect phrase, e ⊆ g. For example, we mined “fatigue” from the data, which is a reverse match to “extreme fatigue” in the HPVVR. Out of a total of 136 ground-truth effect phrases, we found 55 (40.4%) matches, 78 (57.4%) partial matches, 48 (35.3%) reverse matches, and 103 (75.7%) either partial or proper matches or both. Meanwhile, there were also some candidate effect phrases which were newly discovered effects.

As the causality mining method is a completely bottom-up, unsupervised method, we could automatically mine candidate effects for any set of tweets. In particular, for the predicted false tweets, one of the largest candidate effect clusters contained terms relating to the reactions of different entities, such as “Japan,” “Denmark,” and “college,” on the potential issues with the HPV vaccine, such as “recall,” “lose support,” and “banned.” Another such cluster contained terms relating to infertility misconceptions of the HPV vaccine, such as “premature ovarian failure” and “early menopause on young girls.” Another large candidate effect cluster was about the misconceptions of severe adverse events and complications, such as “sudden death,” “paralysis,” and “stroke.” Note that it is possible that some candidate effect phrases may not be directly related to health effects. Thus, to alleviate this limitation in further analysis, we limited the effects to the terms in the ground truth (ie, V_e = G).

The performance of the CNN and BiLSTM models used in this study showed the feasibility of discerning misinformation from factual information regarding HPV vaccines using the text of tweets. On average, both models predicted either class with high confidence. Although both models performed almost identically in terms of accuracy (and confidence) during testing, the CNN trained much more expediently than did the BiLSTM model, leading to its choice as the preferred model.

To examine the risk perceptions pertaining to HPV vaccines, we leveraged the false information classifier and the effect ranker. Figure 4 shows the cumulative NPMI scores after our effect ranker querying for both “HPV Vaccine” and “Gardasil” was applied. We could categorize these perceptions around the costs and benefits of HPV vaccines. In general, people discussed the benefits or low risk of harms in the true HPV vaccine tweets and various adverse events in the false HPV vaccine tweets. The main effects associated with the HPV vaccines in the true HPV vaccine tweets were about the prevention of HPV infection–related cancer and the denial of risk of increased unprotected sexual behavior of the vaccinated teens. The main effects associated with the HPV vaccines in the false HPV vaccine tweets were regarding infertility-related conditions (such as ovarian injury), child developmental disorder, death, and toxic ingredients in the HPV vaccines. Following our previous work on the patient-driven HPVVR, the findings from causality mining aided us in identifying the major concerns related to HPV vaccines, whose solutions could then be prioritized.

The results show that false HPV vaccine messages not only span a wide variety of topics in risk perceptions but also involve a more diverse vocabulary to describe these topics compared to the fewer topics and relatively limited terminology found in true messages. This phenomenon of medical-based term frequency and topic diversity within false or misleading messages has also been noted in similar work regarding anti- and provaccine literature [84]. A possible explanation for this discrepancy is that true information requires an evidentiary consensus, thereby restricting terminology and outcomes to a specific selection of topics or phrases used to describe these topics. Misinformation lacks such restrictions to terms or outcomes and tends to use narrative language or mention novel topics to gauge attention [46,85].

We also observed differences in message framing in true and false HPV vaccine messages through causality mining (see Figure 4). True information contained both gain-framed and loss-framed messages, especially those highlighting the effectiveness of the vaccine at preventing HPV-related cancers, the link between HPV infection and cancer, and negating the potential harms of vaccines such as carefree or unprotected sexual behavior (Figure 4). Conversely, false messages were largely loss-framed, focusing on negative outcomes purportedly caused by the vaccine, such as those causing HPV-related cancer or other serious adverse events (infertility, neurological disorder, or death; Figure 4). The use of the risk-indicating causative verb (eg, vaccines “prevent” versus vaccines “harm” or “cause”, etc) might be diagnostic for differentiating the true and false information. Future studies should leverage previous findings on the effectiveness of message framing to examine the impact of misinformation with different framing [86,87].

Health-related misinformation research spans a broad range of disciplines [58,59], with several studies focusing on different medical domains, such as cancer, sexually transmitted disease and infections, influenza, and more recently, COVID-19 [25,60,61]. In vaccine-related domains, several papers have examined vaccine behavior as well as geographic and demographic patterns on the dissemination of antivaccine and misinformation tweets in social media with respect to autism spectrum disorder [60], influenza (flu) vaccines [88], and cancer treatments [89]. Several research endeavors tackle key issues, such as mitigating label scarcity with additional weak social supervision signals, improving intractability with attention mechanisms, and leveraging network and group or user information [65,90-92]. In general, the distinction between vaccine hesitancy identification and vaccination behavior detection is that the former involves an attitude or stance, while the latter is concerned with detecting the action of getting vaccinated [93]. Our study is more similar to the research in vaccine hesitancy but differs in that we focused on extracting causality from tweets through examining risk perceptions; attributable causes of HPV vaccine–related health concerns or expected gain; and using natural language processing, machine learning, and unsupervised causal mining techniques.

We observed that the convolutional models with multiple filter sizes [70,94] worked better than did BiLSTM models for domains with short text, such as tweets. Intuitively, the CNN architecture captures the most common n-grams (of lengths 3, 4, and 5) and therefore is better at discovering discriminative text patterns in short text. Although we tested more sophisticated BiLSTM architectures, overall, the CNN model performed better than did the other model variations and was faster to train. These findings can be useful for social media health–related analysis, in particular with regards to the set of models that practitioners in this domain should explore for social media text classification.

With respect to causality mining, early works use hand-coded, domain-specific knowledge bases [95,96]. A challenge in identifying causal relations is the variety of ways in which it can be observed via various linguistic constructions. A previous study [97] showed that a classifier can determine whether a causal-verb expression, automatically extracted from predefined linguistic patterns of the form <noun phrase–verb–noun phrase> is a causal relationship or not. However, supervised methods require large amounts of manually annotated causes and effects and are thus resource demanding. Recent work has compared unsupervised methods for causal relationship mining including co-occurrence methods, such as point-wise mutual information, and discourse cue-based methods, which are based on information retrieval techniques, to count the number of matches in a cause-effect query [98]. Such comparisons were performed on large-scale document collections, and thus their insights are not applicable to our tasks that, in contrast, have limited amounts of data. Finally, event causality detection in tweets restricts the causal relationship mining to certain events of interest. In “Event causality detection in tweets by context word extension and neural net” [72], the authors propose an approach to encode both the candidate causal phrase and the candidate effect phrase for developing a feed-forward network classifier. Our method is not restricted to certain events. Most importantly, we focus on health-related messages pertaining to HPV vaccines, an approach which can be generalized to other health topics.

One common bottleneck when applying supervised learning methods is the requirement of large amounts of high-quality annotated data for training. Due to the complex nature of the task-at-hand, and the need for extensive manual effort, our data set size might be restrictive in providing insights that can generalize across other domains and data sources. Additionally, due to frequent linguistic variations found in informal user-generated language, closely worded instances might have evaded deduplication. In the future, we hope to address the shortage of available labeled data by incorporating weak supervision methods and denoising mechanisms. Nevertheless, we chose to continue with supervised learning for higher precision, as weak supervision may result in label noise being injected into the false information detection models and thus affecting the subsequent causality mining steps.

Another limitation stems from the misalignment of model confidence and accuracy. In other words, model confidence might not be indicative of model correctness, a problem that is well-known in the machine learning research community [99]. In our experiments, we observed that the BiLSTM model produced high confidence estimates for most false negatives (ie, it misplaced more confidence when predicting factual text), while the CNN model had an equal number of false positives and false negatives for high-confidence examples. Approximately 20% of CNN’s incorrect predictions had low confidence. Overall, the BiLSTM model seems to be overconfident in one direction and could be potentially calibrated better. Further analyses on these high-confidence inaccurate predictions are required to discover interpretable patterns that can identify misinformation subtopics and statements that share strong similarities to factual counterparts.

Finally, we should note that any use of additional metadata requires caution, especially for information that is added by the user, such as user profile characteristics, as well as reported timestamps and social network links, as recent studies show that misinformation spreaders tend to manipulate not only social network structure by forming groups to increase influence [100] but also several types of metadata [101]. In this study, we did not use these types of additional data sources, and thus we can only interpret content-based results and not along any other dimension other than the relationships found in the text.

The study has demonstrated a systematic, automatic approach to developing computational models for identifying false HPV vaccine–related information and its associated effects on social media. This approach could be generalized to other social media health information and provide insights into estimating the potential effects of a given health topic.