| { |
| "File Number": "1053", |
| "Title": "Evaluating the Robustness of Adverse Drug Event Classification Models using Templates", |
| "7 Limitations": "While the approach of generating new inputs by templates undoubtedly has benefits, it also introduces some limitations. For instance, the combination of all entity fill-ins with all templates can produce some unnatural phrases. An example of this is the Temporal Order template \"After taking {drug}, I had {ade}.\". The ADE entity \"weight gain\" creates the unnatural sounding test case \"After taking cymbalta, I had weight gain.\" instead of \"After taking cymbalta, I gained weight.\" The un-\nnatural use of language may introduce a bias. This should be kept in mind when using the templates. However, not all entity fill-ins will introduce such a bias and the model’s performance on the test cases cannot be fully attributed to the effect of unnatural language use.\nA second potential bias when using templates is that it may not be able to depict a large variety of language when only few templates were used. An example of this are the templates for the positive class Beneficial Effect test where each test case includes the word \"problem\". A model could use this as a proxy for correctly classifying the test cases.\nLastly, as described in Section 4, not all features of social media tests were used when creating templates. No anonymized usernames, hashtags, non-standard punctuation, and colloquialisms other than contractions were applied in the templates. This may introduce a bias as there is a slight difference in language variety between the templates and the training data. A researcher should keep in mind that slight changes in the model performance may also be attributed to this shift in language variety.", |
| "abstractText": "An adverse drug effect (ADE) is any harmful event resulting from medical drug treatment. Despite their importance, ADEs are often under-reported in official channels. Some research has therefore turned to detecting discussions of ADEs in social media. Impressive results have been achieved in various attempts to detect ADEs. In a high-stakes domain such as medicine, however, an in-depth evaluation of a model’s abilities is crucial. We address the issue of thorough performance evaluation in English-language ADE detection with handcrafted templates for four capabilities: Temporal order, negation, sentiment, and beneficial effect. We find that models with similar performance on held-out test sets have varying results on these capabilities.", |
| "1 Introduction": "When a trained model is applied to real-world data, it may be confronted with phenomena that are under-represented or non-existent in the training data (Belinkov and Bisk, 2019; Moradi and Samwald, 2022). This raises the question of how to evaluate a model’s performance and generalization abilities. Reporting summary statistics and heldout test set performance is a common practice in model evaluation. While this can provide an indication of the model’s performance and ability to generalize, there are some issues with this practice. Firstly, held-out test sets often arise from the same distribution as the training data and will, therefore, exhibit the same patterns and biases to a high degree. Real-world data, however, may have different feature distribution or exhibit noise. Held-out testing, therefore, often provides an unsatisfactory estimation of a model’s performance and generalization abilities (Belinkov and Bisk, 2019; McCoy et al., 2019; Ribeiro et al., 2018).\nSecondly, a high model score does not necessarily reveal what the model has learned during train-\ning. Research has shown that a model may not learn relevant patterns but instead base its decisions on shallow heuristics or proxies (McCoy et al., 2019). Benchmark challenges have attempted to address this issue by testing models on a wide range of aspects of language (Wang et al., 2019). However, not all aspects can be tested in a benchmark, and the benchmark itself may exhibit unintended biases (Kiela et al., 2021), so the question of what a model has learned remains.\nInspired by the behavioral testing suite CheckList (Ribeiro et al., 2020), we propose the use of template-based test cases to test different capabilities of adverse drug effect (ADE) classification models. ADEs are any harmful consequence to a patient due to medical drug intake. Due to the potential detrimental outcomes of ADEs, the detection of ADEs is an important goal in health-related NLP and has been a subject of research for a considerable time. We test models in understanding of temporal order, positive sentiment, beneficial effects and negation (see Table 1).\nIn high-stakes domains such as medicine, an indepth evaluation of a model’s abilities is crucial. Related work (Section 2), however, suggests that shortcomings towards selected linguistic phenomena and reliance on proxies for model decisions may exist in models in the biomedical domain.\nIn this work1, two transformer-based models for the detection of ADEs in user reports on social media were fine-tuned and tested by conventional held-out testing as well as additional templatebased tests. The results of held-out testing and the template-based tests were compared in order to better understand (i) the models’ shortcomings and (ii) the potential gaps in knowledge that can occur when a model’s abilities are only evaluated via test set performance. We find that models underper-\n1The templates and code can be found at https://github. com/dfki-nlp/ade_templates\nform on some capabilities and show differences in some capabilities despite highly similar F1-scores on the held-out test set. We therefore provide the following contributions:\n• A curated test bench of 99 templates with 1505 variations to investigate the robustness of ADE classification models across four capabilities.\n• A comparison of two popular transformerbased models on long-tail linguistic phenomena in the classification of ADEs.", |
| "2 Related Work": "Studies on the detection of ADEs in user-generated texts have been conducted since approximately 2010, when Leaman et al. published the first English dataset within this domain. The usual downstream tasks are those common in information extraction: Document classification, to find relevant documents containing mentions of adverse effects;\nnamed entity recognition, to identify medication and disease-related mentions; and relation classification, to establish associations between the entity mentions. Approaches for all of these tasks range from rule- and lexicon-based systems (Leaman et al., 2010; Nikfarjam and Gonzalez, 2011) to traditional machine learning pipelines (Gurulingappa et al., 2012; Ginn et al., 2014; Segura-Bedmar et al., 2014) and, recently, deep neural networks (Huynh et al., 2016), specifically transformer-based setups (Weissenbacher et al., 2019; Miftahutdinov et al., 2020; Gusev et al., 2020; Magge et al., 2021b).\nHowever, even advanced models struggle with the supposedly simple task of classifying a document into either “contains an ADE” (henceforth ADE) or “does not contain an ADE” (no ADE), a standard binary classification that is still necessary to find relevant documents for further information extraction. This is often due to a strong class imbalance (in most cases, the documents containing ADEs are in the minority), the usual noise\nin social media data, ambiguities in health-related statements of patients, and general weaknesses of language models in coping with certain linguistic phenomena not only with respect to ADEs.\nFor example, Scaboro et al. (2021) have studied the extraction of ADEs from tweets using BERT, SpanBERT (Joshi et al., 2020), and PubMedBERT (Gu et al., 2021). They tested all three models’ ability to handle negation and detect shortcomings in all three models. Moradi and Samwald (2022) investigated the robustness of four transformer models specialized in the biomedical and clinical domain over a variety of tasks such as sentence classification, inference, and question answering. The models’ robustness is tested by adding minor meaning-preserving changes to the input with the goal of fooling the model. Their findings highlight the vulnerability of state-of-the-art transformer-based models to adversarial input.\nFinally, there is CheckList (Ribeiro et al., 2020), a model-agnostic framework aimed at testing a trained model’s behavior and gaining an in-depth understanding of its potential shortcomings. CheckList guides the creation of test cases based on natural language capabilities, which are used as new inputs to the trained model and subsequently evaluated. The idea is to determine which capabilities (e.g., negation handling, robustness) are necessary for the task the model is intended to perform. Ribeiro et al. (2020) identify three possible test types which can be used for testing the capabilities: the Minimum Functionality Test (MFT), which targets a specific behavior similar to a unit test; the Invariance Test (INV), where the model’s robustness to irrelevant perturbations is tested; and the Directional Expectation Test (DIR), which consists of adding perturbations that are expected to lead to a specific outcome. Ribeiro et al. (2020) observe that the CheckList-based evaluation approach could not only uncover bugs in previously tested models but also that CheckList can make the search for bugs more systematic. Recently, updates to CheckList, AdaTest (Ribeiro and Lundberg, 2022) and AdaTest++ (Rastogi et al., 2023), were proposed which assist the user in finding bugs by suggesting topics and test cases in a semi-automated process. While these are valuable additions, we decided to use the template-based approach for this project because we had pre-selected capabilities that we wanted to test with full control over the template design.\nCheckList applications include the evaluation of general capabilities of models (Xie et al., 2021) as\nwell as evaluating models in specialized tasks such as offensive speech detection (Bhatt et al., 2021; Manerba and Tonelli, 2021) and automatic text simplification (Cumbicus-Pineda et al., 2021). For the specialized tasks, the authors use CheckList to guide their testing approach by defining new capabilities specific to the task at hand.\nIn the biomedical and clinical domain, Ahsan et al. (2021) use CheckList to test four linguistic capabilities (negation, temporal order, misspellings, and attributions) on their transformer-based model with a dataset of clinical discharge notes. One of their findings is that the model struggles to correctly distinguish between past and present mentions of substance use in the discharge notes. The detection of ADEs, however, is not part of the research.\nThe exposure of potential weaknesses in transformer-based models in the biomedical domain motivates an in-depth analysis of models used for ADE detection. To our knowledge, a systematic template-based approach to test model capabilities has not yet been applied to ADE detection.", |
| "3 Methods": "We use templates to test a selection of linguistic capabilities of binary ADE classification models. To this end, we first manually create templates (see Section 3.1) and then sets of test cases, by using entities to fill placeholders in the templates (see Section 3.3). We then evaluate two fine-tuned classification models on these test cases and compare their predictions with each other and with the models’ performance on the held-out test set.\nExample 1: Template for Temporal Order (ADE)\nI started taking {drug} before I experienced {ade}.\nWe test four capabilities: Temporal Order, Positive Sentiment, Beneficial Effect, and Negation (see Section 3.2). 99 base templates are created with 1505 variations (for details see Table 5 in Appendix A). Each template is also assigned a label (ADE/no ADE) in accordance with published guidelines for the annotation of ADEs (see Section 4.1.1). The template in Example 1 provides a test case for the capability Temporal Order and has a positive label (ADE). Filled-in template examples for every capability we test are listed in Table 1. The filledin templates (test cases) serve as the input to the fine-tuned model for inference. In the following, we present more details about the template creation\nand the investigated capabilities.", |
| "3.1 Template Creation": "Template-based evaluation is most effective with a large number of test cases that cover a diverse range of potential inputs. These test cases are based on templates, which include placeholders. For every placeholder, there is a list of potential entity fill-ins as in Example 1, {drug} and {ade}, which could be filled with, e.g., Effexor and nausea. The abstraction of test cases to templates allows to systematically capture important linguistic scenarios while creating a large number of different test cases. The process is visualized in Figure 1.\nIn the interest of linguistic diversity, variations of base templates were introduced for all capabilities except Beneficial Effect. For Temporal Order and Negation templates, the vocabulary of the base template was modified to increase diversity. Positive Sentiment templates underwent syntax variations by exchanging or removing the conjunction between the two phrases.\nThe templates have a mean token count of 10.6 and 13.4 for the no ADE and ADE class respectively2. After filling in the entities for the placeholders, the average test case length in the experiments is 14.7 for the no ADE class and 16.6 for the ADE class.", |
| "3.2 Capabilities": "The choice of capabilities for this work is inspired by considerations on abilities a robust ADE classification model should possess and shortcomings of biomedical models as reported in Section 2. We based the phrasing of the templates on linguistic properties of social media posts: First-person\n2Tokens were split at whitespace.\nusage, mostly single short sentences, and colloquial language. Contractions were used occasionally. However, usernames, misspellings, and nonstandard grammar and punctuation were not applied in the templates as they manifest a separate capability. All templates created can be viewed as templates for a CheckList Minimum Functionality Test (Ribeiro et al., 2020).\nTo verify the existence of the described phenomena in the dataset, we randomly sampled 1,000 documents and let two annotators check each tweet for the occurrence of these phenomena. The annotations showed that eight of the sampled tweets contain expressions of temporal order, one positive sentiment, one beneficial effect, and one negation. This sample showed that, as expected, the phenomena are rather rare but still exist in the long tail of the data distribution. Nevertheless, an expert would expect a good classification model to have these capabilities.\nTemporal Order The templates for testing Temporal Order adapt the temporal structure test of Ribeiro et al. (2020) and investigate the model’s ability to correctly process information on past, present, and future as expressed in text. In the context of ADE detection, it is important for the model to “understand” temporal order since an effect cannot be an ADE if it occurred before the drug intake. According to the annotation guidelines based on which the data we use for fine-tuning was annotated, an effect occurring after a drug intake was labeled as ADE if the patient draws a connection between the effect and drug intake. Therefore, the templates assume an ADE when a harmful effect occurs after the drug intake.\nPositive Sentiment ADEs are often reported using negative sentiment (Alhuzali and Ananiadou, 2019). If many ADE reports contain negative sentiment, an ADE detection model might perform well by using negative sentiment as a proxy. Nevertheless, a report might also be expressed favorably. This could be the case when a patient experiences relief from the original symptoms alongside a mild ADE. Therefore, an ADE detection model should recognize ADEs even when expressed in a positive framing so as not to miss out on less severe ADEs.\nBeneficial Effects The third capability is the correct distinction between ADEs and beneficial effects. The latter are secondary effects of a drug that are not related to the reason for using the med-\nication and which have, nevertheless, a positive outcome for the patient. Note that an effect may be regarded as positive or negative depending on the patient, their general health, and the context. Weight loss, for instance, may be considered a negative secondary drug effect or a beneficial effect depending on the patient. The tests in this work assume that a positive secondary effect is a beneficial effect, not an ADE. The Beneficial Effect test that expects a negative class label (no ADE) expresses the occurrence of a beneficial effect. The positive class (ADE) test consists of test cases that express an ADE that could be classified as a beneficial effect, but the context states that the user views the effect as negative.\nNegation Negation templates test the model’s ability to process negation in text. Negation detection is a general challenge in NLP and a common phenomenon in language (Hossain et al., 2022; Truong et al., 2022). Thus, it is also an important capability for ADE detection. The Negation test that expects a negative class label (no ADE) contains a negated ADE. The positive class (ADE) test cases include an ADE mention as well as a negation without negating the ADE.", |
| "3.3 Entity Placeholders": "All templates have entity placeholders for a drug name. Templates for Temporal Order, Positive Sentiment, and Negation also have a placeholder for an ADE entity. Templates for Beneficial Effect contain an effect that may be considered an ADE or a beneficial effect depending on the context. A list of the effects used in the Beneficial Effects tests is provided in Appendix A.2. Template variations of the Temporal Order capability that use time entities have placeholders for time expressions. The placeholders are filled with the respective time expressions from a self-created list of entities.", |
| "4 Experiments": "We frame ADE detection as a binary classification task. We first describe the experiments on the custom dataset and then the experiments on our template-based test cases.", |
| "4.1 Fine-Tuning Experiments": "The following describes data, training and evaluation on the custom dataset.", |
| "4.1.1 Data": "The custom dataset for our experiments consists of three social media corpora: The SMM4H-2021 Shared Task 1a training data (Magge et al., 2021a) (61% of the custom dataset), the SMM4H-2017 Shared Task dataset (Sarker et al., 2018) (38%), and artificially negated tweets from the NADE dataset (Scaboro et al., 2021) (1%), resulting in 28,468 tweets. The data flow and their origin are shown in Figure 2. Dataset statistics are covered in Table 2. In the user-reported texts, each sample either describes an ADE (ADE) or does not contain an ADE mention (no ADE).\nThe SMM4H-2021 Shared Task 1a training data (Magge et al., 2021a) itself consists of posts from Twitter and DailyStrength3 collected using a list of 81 drugs widespread on the US market (Nikfarjam et al., 2015). The data was annotated by two expert annotators. The annotators did not include beneficial effects in the ADE definition. It further includes some data previously used in the SMM4H-2017 Shared Task (Sarker et al., 2018).\nThe SMM4H-2017 Shared Task data was collected from Twitter using generic drug names with a total of 250 keywords and subsequently annotated by two annotators. Again, the annotators excluded\n3www.dailystrength.org\nbeneficial effects from the ADE definition. Overlapping texts between the SMM4H-2021 data and the SMM4H-2017 data used for our merged custom dataset were removed.\nThe last part of our custom dataset are artificially negated tweets from the NADE dataset (Scaboro et al., 2021). This dataset consists of tweets originating from the SMM4H-2019 Shared Task (Weissenbacher et al., 2019) and manually negated by annotators. Each negated tweet contains a statement that negates the presence of an ADE. The three components are shown again in Table 2.\nWe use this merged version of multiple datasets to give the fine-tuning models the best chance to learn different capabilities from varied data. The texts in the custom dataset are between 1 and 34 tokens long.4 Negative (no ADE) samples are slightly shorter on average (16.2 tokens) than positive (ADE) samples (18.4 tokens). These are slightly longer than our test cases with an average length of 14.7 tokens and 16.6 tokens. Data splits for training, validation, and testing were created with a 70-10-20 ratio and stratified sampling by class label.", |
| "4.1.2 Model Fine-Tuning": "For the task of ADE classification, we fine-tune BioRedditBERT (Basaldella et al., 2020) and XLMRoBERTa (Conneau et al., 2020) on the custom dataset described in Section 4.1.1. BioRedditBERT is a BERT-base uncased model related to BioBERT (Lee et al., 2019), a model pre-trained on the original BERT training corpus (English Wikipedia + BookCorpus) as well as on medical texts sourced from PubMed and PMC. It was then further finetuned on a corpus of health-related Reddit posts. XLM-RoBERTa is a popular multilingual model with no specific medical pre-training data. We chose these models to gain insights on robustness of a language model with medical knowledge compared with an general domain language model that has no specific medical knowledge.\nThe inputs were sampled with replacement weighted by class ratio due to the class imbalance. This sampling strategy resulted in a better F1-score on the validation dataset.", |
| "4.1.3 Held-Out Test Set Evaluation": "We evaluate the fine-tuned models on the test set using precision, recall, and F1-score for each class. The main metric we focused on is F1 of the positive class due to the large class imbalance. This\n4Tokens were split at white spaces.\nmetric was also used for hyperparameter tuning on the validation set. We compare per-class recall to the models’ performances on each capability of the test cases. The goal of this comparison is to determine whether the template-based evaluation approach contradicts the overall impression of the model performance measured by held-out test set performance.", |
| "4.2 Test Case Experiments": "We use all templates for each test and randomly select only one template variation per base template for the capabilities Temporal Order, Positive Sentiment, and Negation to have a manageable number of test cases. We created a total of 11,265 test cases, of which 4,620 test cases belong to the negative class (no ADE) and 6,645 belong to the positive class (ADE). Table 3 shows the number of test cases run per test.\nA random sample of 15 ADEs, 15 mild ADEs, 5 drug names, 7 single time entities, and 7 relational time entities was taken. A list of sampled ADEs, mild ADEs, and drug names can be viewed in Appendix B.", |
| "4.2.1 Drug and ADE Template Fill-Ins": "We need expressions of ADEs and medical drugs to fill in the placeholders in the templates. These are automatically extracted from the PsyTAR dataset (Zolnoori et al., 2019) of patient reports on psychiatric medications. The dataset consists of 891 Ask-a-Patient5 patient forum posts on the topic of four psychiatric medications: Zoloft, Lexapro,\n5www.askapatient.com\nCymbalta, and Effexor XR. The corpus was annotated for ADE mentions by four annotators with a health-related background. A mention was considered an ADE “if there is an explicit report of any sign/symptom that the patient explicitly associated them with the drug consumption” (Zolnoori et al., 2019). All four drug names of PsyTAR were extracted as well as two spelling variations of “Effexor XR” and lowercase versions of all drug names. Statistics on the occurrences of the drug names in the custom training dataset can be found in Table 7 in Appendix B. Extracting ADEs and drug names from the same domain ensures a high likelihood of compatibility between ADEs and medications.\nThe ADE entities extracted from PsyTAR are user-generated descriptions of ADEs that are often multi-word expressions and which use nonstandardized terms. We did not correct grammar and spelling errors in the extracted ADEs.\nWe created the templates in a way that most short noun phrases6 fit as ADE entities, therefore, short noun phrases were filtered from the ADE mentions in PsyTAR. A total of 1,227 unique ADEs were extracted, amounting to 36.50% of unique ADE entities in PsyTAR.7\nFor the Positive Sentiment test, the extracted ADEs were manually filtered to collect 60 less severe ADEs. This was a necessary step to avoid creating unrealistic test cases such as “I always have severe pain in my hands when I’m on Cymbalta, but I am happy my symptoms have reduced”.\nThe time entities for the variations in Temporal Order tests were not extracted but generated. Numbers between 1 and 25 inclusive were combined with a noun (either “days”, “weeks”, or “months”). A random selection of these combinations was used as time entities.", |
| "5 Results": "The following presents the results of both the baselines and the template-based test cases.", |
| "5.1 Model Baseline": "The results of the baseline models can be found in Table 4. All models were evaluated on the same test split of the fine-tuning corpus.\nThe F1-score of BioRedditBERT on the positive class (ADE) is 0.698, whereas XLM-RoBERTa\n6The longest extracted ADE has a length of 7 tokens. 7More details on the extraction can be found in Ap-\npendix A.1.\nachieves a score of 0.700, which indicates very similar general performance. Due to the large class imbalance, the models reached a higher performance on the majority class (no ADE) with F1-scores of 0.972. The high overlap in data allows for comparison of this model’s performance to the best performing models proposed in the latest SMM4H Shared Task on ADE classification (Weissenbacher et al., 2022).", |
| "5.2 Template-Based Test Results": "We compare model performance on the custom dataset to each template-based capability test performance separately. Due to the variations in model performance over the two classes, we use per-class recall as a measurement of comparison between the model performance on the custom dataset and the template-based test cases as shown in Figure 3. For both models, all tests with no ADE labels fall short of the baseline performances. The highest level of performance is observed in the Negation tests where BioRedditBERT and XLM-RoBERTa pass 92% and 94% of the test cases, respectively. On the other hand, the Beneficial Effect tests perform strikingly worse than the baselines with BioRedditBERT and XLM-RoBERTa passing only 7.5% and 5.8% of the test cases, respectively. All three versions of the negative class Temporal Order tests lie below the baselines but to a varying degree with a range of 54%-78% for BioRedditBERT and a range of 62%-74% for XLM-RoBERTa.\nFor ADE, the models perform below the baseline (recall of 68% for both models) on the standard Temporal Order and double time entities Temporal Order test (25%-48%), while the baseline is exceeded on the single time entity Temporal Order test with 90% for BioRedditBERT and 80% for XLM-RoBERTa. Based on the varying model performance on different types of Temporal Order tests for both the negative and the positive class,\none can conclude that the model is not robust to changes in expression of temporal structure: The use of single time entities affects the model performance positively compared to the use of prepositions (standard Temporal Order) and double relational time entities. Furthermore, BioRedditBERT (48%) performs much better on standard Temporal Order tests than XLM-RoBERTa (25%).\nMild ADEs expressed in positive sentiment as in the Positive Sentiment test do not pose a problem to the model. The performance on the Positive Sentiment test cases (72% for BioRedditBERT and 68% for XLM-RoBERTa) lies above the baseline of the positive class for both models. Also, the models’ performance on the positive class negation test lies below the baseline, with BioRedditBERT (60%) again performing much better than XLMRoBERTa (41%).\nUnlike for the negative class test, almost all test cases in the Beneficial Effect test on the positive class are correctly classified as ADE. The poor performance on the negative Beneficial Effect test and the outstanding performance on the positive class Beneficial Effect test leads to the conclusion that the model has not learned to distinguish between ADEs and beneficial effects. Both models classify 96% of Beneficial Effects test cases as ADE, even though half of the tests have a no ADE mention. Possible explanations for this behavior are that the number of beneficial effect samples in the custom dataset is low and/or that the model does not take\nthe context into account that distinguishes an ADE from a beneficial effect.\nEach of the five selected drug name variants was used in every template allowing for an analysis of the impact of drug names in the test cases. Performance variations on test cases with different drug names indicate reduced robustness of the model. We find slight variations in the model performance over different drug names as shown in figure Figure 4 for XLM-Roberta. A potential explanation of these variations may be deviations in the occurrence of the respective drug names in the custom\nfine-tuning dataset, see Table 7 in Appendix B.", |
| "6 Conclusion and Future Work": "In this work, we present a template-based approach for evaluating capabilities of models on the task of ADE detection in social media texts. Four capabilities, Temporal Order, Positive Sentiment, Beneficial Effect, and Negation, were identified and corresponding tests were created. Two high-performing models for the task of ADE detection were evaluated using the adapted approach.\nResults show that the models’ performances vary across capabilities. While both models perform well on the Positive Sentiment tests, BioRedditBERT outperforms XLM-RoBERTa on Negation. The models are not able to distinguish between ADEs and beneficial effects and are not robust to changes in the expression of temporal structure in text. In summary, the template-based approach adapted to ADE classification has provided a better understanding of the shortcomings of highperforming models and can highlight previously undetected differences between models that perform almost identically on a held-out test set. We publish the templates to enable researchers to evaluate their own ADE classification models.\nFurther research may expand on this work by adding tests for more capabilities and evaluating other models using this approach. For example, in the phenomena annotation described in Section 3.2, we found 1.6% questions and 1.1% speculative content in the tweets. The linguistic variety of the templates could be improved by using a large language model to generate templates or test cases. A different direction of research may focus on improving the model’s faults detected during evaluation. One method of improvement is to include a subset of the test cases as new training data (McCoy et al., 2019).", |
| "Acknowledgments": "We would like to thank the anonymous reviewers for their feedback on the paper. We further thank our annotators Alon Drobickij, Selin Yeginer, and Emiliano Valdes Menchaca.\nThis work was partially supported by the German Federal Ministry of Education and Research as part of the project TRAILS. We are furthermore grateful for the support of the German Research Foundation (DFG) under the trilateral ANRDFG-JST AI Research project KEEPHA (DFG442445488), and the German Federal Ministry of Education and Research under the grant BIFOLD24B.", |
| "A Templates": "The number of templates with linguistic variations for each capability can be seen in Table 5. Example templates without filled-in entities are in Table 6.\nA.1 Extraction of ADEs from PsyTAR Sets of Parts of Speech combinations (tagsets) were created to define which sets of POS tags constitute a short noun phrase. An English POS tagger (spaCy) was then used to tag every token in the PsyTAR ADEs and filter out the chosen noun phrases.\nExamples of PsyTAR ADEs that were retrieved using this method are “listlessness”, “recurrence of ocular migraines”, and “bad pain in my right arm”. The goal of this process was to retrieve as many and diverse ADE descriptions as possible, yet the tagsets are not extensive and not all relevant ADEs were retrieved. Reasons for not passing the tagset filters were not being a noun phrase (“gained 18 pound”), incorrect POS tag assigned tagger (“heartburn”), incorrect POS tags assigned due to typos or extra whitespace, long noun phrases (“stomach cramping the first couple of days”), and punctuation marks/symbols (“increase in alcohol abuse/dependence”).\nA.2 List of Beneficial Effects List of (potential) beneficial effects used for the Beneficial Effect tests: weight loss/weight gain, sleepiness/decreased need for sleep, loss of appetite/increased appetite.", |
| "B Experiment Details": "List of entities used as fill-ins for ADE, milder ADE for the Positive Sentiment test, and drug names used in the experiments for this project.\n• drug names: zoloft, effexor, cymbalta, Effexor XR, effexorxr\n• ADEs: Incredible sweet tooth, big appetite, many dreams, Difficulty Orgasming, excellerated heart rate, Insomnia, blackouts, bad pain in my right arm, a little more lethargy, VERY vivid dreams, stiff shoulders, EXTREME DRY MOUTH, Dialated pupils, increase in Libido, acid reflux\n• milder ADEs: sugar craving, carbohydrate cravings, bouts of sleeplessness, gum pain, secretion under my toungue, weird dreams, stiff muscles, mild constipation, arm tingling, increased heat sensitivity, strange dreams, poorer concentration, cravings for sweets, hard time falling asleep, neck pain\nThe counts of the occurrences of the drug names can be found in Table 7.", |
| "C Model Details": "BioRedditBERT (Basaldella et al., 2020) is a BERT-base uncased model related to BioBERT (Lee et al., 2019), a model pre-trained on the original BERT training corpus (English Wikipedia +\nBookCorpus) as well as on medical texts sourced from PubMed and PMC. BioRedditBERT, in turn, was initialized from BioBERT and continued to pre-train on a corpus of health-related Reddit posts. The Reddit dataset contains 800.000 posts from 68 health-related subreddits collected between 2015 and 2018. The specific set of training data of BioRedditBERT was the pivotal argument for choosing this model for the task of ADE classification on the Twitter dataset.\nXLM-RoBERTa (Conneau et al., 2020) XLMRoBERTa is a popular multilingual classification model without a focus on the biomedical domain.\nWe conducted hyperparameter search for both models and tried batch sizes of 8, 16 and 32 and learning rates of 3 · 10−6, 10−5 and 3 · 10−5. Both models achieved the best performance on the development set at 16, 10−5 and trained with the AdamW (Loshchilov and Hutter, 2017) optimizer. No truncation of inputs was applied and the model was evaluated on the validation set after every epoch. The inputs were sampled (batch sampling) with replacement weighted by class ratio due to the class imbalance (see Section 4.1.1).", |
| "D Per-Template Performance": "The performance of the models on the templatebased tests also varies within each test. For all tests except the Beneficial Effect tests, the models’ performance varies for each template, see Figures 5 and 6. The dependence of the model performance on the template demonstrates that the wording of a template influences the models’ ability to handle a capability. In turn, this stresses the importance of creating a wide range of variations in templates when using template-based evaluation." |
| } |
