Identifying Depressive Symptoms from Tweets: Figurative Language Enabled Multitask Learning Framework
WARNING: This paper contains examples which are depressive in nature.

The recent survey conducted by WHO shows that a total $322$ million people in the world are living with depression. At its most severe, depression can lead to suicide and is responsible for $850,000$ deaths every year [WHO and others, 2017]. Early detection and appropriate treatment can encourage remission and prevent relapse [Halfin, 2007]. However, the stigma coupled with the depression makes patients reluctant to seek support or provide truthful answers to physicians [Haselton et al., 2015]. Additionally, clinical diagnosis is dependent on the self-reports of the patient’s behavior, which requires them to reflect and recall from the past, that may have obscured over time. In contrast, social media offers unique platform for people to share their experiences in the moment, express emotions and stress in their raw intensity, and seek social and emotional support for resilience. As such, the depression studies based on social media offer unique advantages over scheduled surveys or interviews [Coppersmith et al., 2014, De Choudhury and De, 2014, Manikonda and De Choudhury, 2017, De Choudhury et al., 2016]. Social media self-narratives contain large amounts of implicit and reliable information expressed in real-time, that are essential for practitioners to glean and understand user’s behavior outside of the controlled clinical environment. Majority of these existing studies have formulated the social media depression detection task as a binary classification problem (i.e., depressive/non-depressive) and therefore are limited to only identifying the depressive users.
To assist healthcare professionals (HPs) intervene in a timely manner such as with an automatic triaging, it is necessary to develop an intelligent decision support system that provides HPs fine-grained depression related symptoms. The triage process is a critical step in giving care to the patients because, by prioritizing patients at different triage levels based on the severity of their clinical condition, one can enhance the utilization of healthcare facilities and the efficacy of healthcare interventions. There have been a few efforts to create datasets for capturing depression severity, however they are limited to (1) only clinical interviews [Valstar et al., 2013, Ringeval et al., 2019, Gratch et al., 2014] and questionnaires [De Choudhury et al., 2013], and (2) individuals who voluntarily participate in the study [De Choudhury et al., 2014].
In this work, we exploit the Twitter data to identify the indications (specifically, PHQ-9 guided symptoms) of depression. We developed a high quality dataset consisting of total $12,000$ tweets, with $3738$ tweets posted by $205$ self-reported depressed users over $2$ weeks time, which were manually annotated using PHQ-9 questionnaire [Kroenke and Spitzer, 2002] based symptoms categories. In Table-1, we provide sample tweets associated with the nine item PHQ-9 depression symptoms. Our research hypothesis is that depressed individuals discuss their symptoms on Twitter that can be tracked reliably.

Nonetheless, user social-media post offer unique challenges as discussed below:

• •

  Usage of the figurative language: First, the depressive users often tend to use figurative language (‘FL’) elements such as sarcasm and metaphor, to describe their symptoms. For example, one user wrote metaphorically, “My skin is paper, razor is the pen”, while another user wrote “I want to cut myself”. While both of these utterances refer to one specific medical concept “Self-Harm”, the first sentence utilizes paper and pen metaphorically to convey self-harm. Furthermore, previous studies utilizing social media data reported prediction errors when drug or symptom names were utilized in a figurative sense [Iyer et al., 2019].

• •

  Usage of implicit sense: The creative expressions used by depressive users also possess implicit sense not evident from a literal reading. For example, one user expresses their desperation as, “What if life comes after death, grab my knife and find out myself.” implicitly referring to “Self-Harm”, while another user gives a compliment with “You have a killer look.”, or captures anger through “If looks could kill, I would be dead by now.”. Other challenges include recognizing misspelled words, slangs, acronyms and unconventional contractions.

• •

  Usage of highly polysemous words: The vocabulary of social media language offers polysemous words that require understanding of the context to determine the semantic labels. For example, “woke up and nose started bleeding” and “I wish I had the nerve to press the blade deeper into my skin so I don’t stop bleeding this time.”, use “bleeding” in different contexts and senses.

To account for this creative linguistic device widely observed in utterances of depressive users, we propose a Figurative Language enabled Multi-Task Learning framework (FiLaMTL) that works on the concept of task sharing mechanism [Ruder, 2017, Yadav et al., 2018b, Yadav et al., 2019]. In this work, we improve the performance and robustness of the FiLaMTL for the primary task of ‘symptom identification’ combined with the supervisory task ‘figurative usage detection’ in a multi-task learning setting. We introduce a mechanism named ‘co-task aware attention’ which enables the layer-specific soft sharing of the parameters for the tasks of interest. The proposed attention mechanism is parameterized with the task-specific scaling factor for BERT [Devlin et al., 2019] layers. BERT enables even the low-resource tasks to benefit from deep bi-directional architectures and the unsupervised training framework to obtain the context-aware encoded representation. The virtue of this model is its ability to learn the task-specific representation of the input tweet by coordinating among the layers and between the tasks.
Contributions:

1. 1.

   We propose a robust multi-task learning framework for identifying fine-grained PHQ-9 defined symptoms from depressive tweets that takes into consideration the figurative language wired in the communication of depressive users. To the best of our knowledge, this is the first study in the depression domain that accounts for figurative language in the users social-media post.

2. 2.

   We introduce an effective way to fine-tune the BERT model for multi-task learning using ‘co-task aware attention’ to better encode the feature across the different BERT layers and tasks. This mechanism allows the model to learn the layer and task-specific parameters, to control the information flow from each BERT layer, in end-to-end model training.

3. 3.

   To evaluate our study, we created a corpus of $12,155$ tweets, with $3738$ tweets annotated with 9 PHQ-9 symptom classes (validated by the collaborator psychiatrists). Additionally, these tweets were also labeled with the figurative classes: metaphor and sarcasm.

Depending upon the data modalities and depressive markers, we categorize the existing literature as follows:

1. 1.

   Linguistic Marker: Language often reflects how people think and is a well known tool used by psychiatrists to assess the mental health condition of the people [Fine, 2006]. Numerous research [Coppersmith et al., 2014, De Choudhury et al., 2013, De Choudhury et al., 2014, Yadav et al., 2020b] has shown that modeling of word-use and social language combined with network analysis has been effective in recognizing depression. A widely adopted resource for understanding the linguistic patterns in mental health is the well-known Linguistic Inquiry Word Count (LIWC) [Pennebaker et al., 2007]. Other researchers exploited sentiment analysis [Xue et al., 2014, Huang et al., 2014, Yadav et al., 2018a], topic modeling [Resnik et al., 2015] and emotion features [Chen et al., 2018, Aragón et al., 2019] to detect depression. Furthermore, substantial progress has been made with the introduction of a shared task [Coppersmith et al., 2015, Milne et al., 2016]. Recently, most of the existing studies [Yates et al., 2017, Benton et al., 2017] have drifted from the traditional linguistic indicators to automated feature generation using the neural network based technique to predict or assess at-risk depressive users.

2. 2.

   Visual Marker: Visual information such as head pose, body movement, facial expressions, gestures and eye blinks provide important cues in analyzing depression. ?) examined if there exists a relationship between non-verbal cues and depression severity using Facial Action Coding System [Ekman and Friesen, 1978]. In another prominent study utilizing FACS, ?) identified that a more downward gaze angle, dull smiles, shorter average lengths of smile, longer self-touches may predict depression. Several studies [de Melo et al., 2019, Cummins et al., 2013] have also investigated the Space-Time Interest Point (STIP) features that capture spatio-temporal changes such as facial motion and the movement in the hand, foot, and head.

3. 3.

   Speech Marker: Recent studies have shown the potential for exploiting speech for depression detection and monitoring [Cummins et al., 2015a, Cummins et al., 2015b, Scherer et al., 2014]. Numerous research [Mundt et al., 2012, Hönig et al., 2014] have revealed the strength of prosodic markers, specifically the speech-rate to analyze the level of depression. ?) proposed a depression classification system based on a wide range of acoustic feature like prosodic, spectral, voice quality, and glottal feature. Other prominent studies [Mundt et al., 2012, Cummins et al., 2011] have explored spectral features like prosodic timing measures, mel-frequency cepstral coefficients (MFCC) and glottal features to accurately classify depressed and control groups.

4. 4.

   Multi-modal Marker: In recent times, there is visible surge in investigating multi-modal indicators to diagnose depression, particularly due to publicly available datasets made available through research challenges like Audio/Visual Emotion Recognition (AVEC) Workshop Depression Sub-challenge (2013-2019) [Schuller et al., 2011, Valstar et al., 2013, Ringeval et al., 2019] and popular Distress Analysis Interview Corpus (DAIC) [Gratch et al., 2014]. Several computational models [Tzirakis et al., 2017, Ringeval et al., 2017, Ringeval et al., 2018] based on machine learning and sophisticated deep learning techniques have been proposed to address the challenges posed by AVEC each year. The best system at AVEC 2019 [Ray et al., 2019] proposed an attention based fusion technique to judiciously select the feature representation obtained from multimodal source.

In this section, we describe how we crawled our dataset Depression to (2) Symptoms (named as D2S) using the Twitter streaming application programming interface, filtered out irrelevant profiles, annotated the tweets of depressive users and verified the annotations by a psychiatrist to prepare the gold standard dataset.

1. 1.

   Dataset Crawling: We utilized the lexicon developed by [Yazdavar et al., 2017] in collaboration with a psychologist. The lexicon contains around $1000$ depression-related terms categorized into nine categories of symptoms from PHQ-9. A subset of highly informative depression indicative terms from the lexicon, that are likely to be used by depressive individuals, was used as seed terms to crawl the public profiles of twitter users with at least one of those filtered terms in their profile description. Through this process, we collected $5,000$ users and their tweets.

2. 2.

   Filtering and Identifying Depressed Users: As users on social media often use sarcasm and metaphor to implicitly express their feelings, contemporary approaches that do not capture context well, miss sub-population of depressive users. To improve upon these approaches, we proceed as follows. After filtering out the retweets, we removed the profiles with less than $100$ tweets and obtained $1567$ users. To emulate the PHQ-9 using social media, we chunked the tweets of each profile into two week buckets. To ensure the high quality data and identify potential depressive profiles with severity level mild to severe based on PHQ-9 scoring, we filtered the profiles based on their frequency of post. After filtering out the profiles who had not tweeted at least $5$ days in the most recent bucket, we obtained $575$ profiles. Although these profiles had depression-related terms in the description, due to the lack of context-sensitivity in the profile identification process, a subset of those were false positives, i.e., non-depressive. A few of these non-depressive profiles were meant to share motivational quotes for depressive users. We strictly examined the visual (i.e., profile image and shared images) and linguistic markers (i.e., profile name, description and tweets) of each of those profiles and removed the users having no depressive tweets. Finally, we obtained $205$ depressive users and selected the bucket of the most recent tweets over two weeks for annotation, which sums to $3738$ tweets.

3. 3.

   Anonymization, Annotation and Verification: Prior to annotation, we anonymized the user profiles with random numbers and replaced the mentions and URLs in tweets with strings ‘@USER’ and ‘URL’ respectively. Four native English speakers from multiple disciplines were assigned to independently annotate the tweets into $9$ categories of PHQ-9. The annotators were also asked to identify the tweets having usage of FL such as sarcasm and metaphor. The annotators were provided with the definitions and samples of annotated tweets from each of those $9$ categories of PHQ-9 and as well as FL. The average inter-annotator reliability scores for the symptoms, depressive vs. non-depressive, and figurative classes were K=$0.83$, K=$0.87$, and K=$0.79$, respectively, based on Cohen’s Kappa statistics. We resolved the conflicting annotations with the majority voting strategy and resolved the ones voted evenly by a psychiatrist. After preparing the final gold standard data, we randomly selected $100$ annotated tweets from each of the symptom categories, including the non-depressive ones and verified by a psychiatrist.

   Depressive														Non-depressive
   Symptoms											Figurative Language
   	S1	S2	S3	S4	S5	S6	S7	S8	S9	Total	Sarc	Meta	Total	Total
   #Tweets	494	657	261	301	473	1054	136	122	970	3738	668	1106	1485	8417
   Avg.Len	13.02	14.13	13.03	13.08	16.08	12.83	13.31	15.63	12.33	12.98	13.17	13.41	13.24	10.96

4. 4.

   Data Analysis: The final data¹¹1Limitation:(i) Only a sub-population (i.e., those with self-reported diagnosis) is identified by this method , (ii) Twitter users cannot be reflective of the entire population, and (iii) it cannot be verified if self-reported depressed users are being truthful comprises 3738 depressive tweets, each tagged with the number of symptoms (single or multiple), and the presence of figurative expressions (either or both) it exhibits, and 8417 non-depressive tweets. Out of these depressive tweets, 1485 ($\sim 40\%$) tweets use FL. = We performed topic analysis to examine the usage of utterances associated with each symptom. Table 3 illustrates the topic distribution of each symptom. We observe from the table, to express their feelings, the depressive individuals use metaphoric phrases such as ‘body is begging’, and ‘feel like trash’; sarcastic expressions such as ‘am eating ?’; implicit utterances such as ‘up all night’, and ‘feel myself falling’.

5. 5.

   Ethical Concerns: Psychiatric research using social media data poses several ethical concerns regarding user privacy, which should necessarily be taken into consideration [Hovy and Spruit, 2016, Valdez and Keim-Malpass, 2019]. Following the ethical practices, as adopted by the previous research on Twitter data [Coppersmith et al., 2015], we constructed our dataset using only public twitter profiles. We anonymized the profiles before presenting it to the annotators who pledged not to make attempts to contact or deanonymize any of the users or share the data with others. The dataset will be shared with researchers who agree to follow the similar ethical guidelines.

Our proposed approach to identify the depressive symptoms, is assisted by the Bidirectional Representation from Transformers (BERT) and multi-task learning [Yadav et al., 2020a] with soft-parameter sharing. This section describes the proposed methodology for identifying the depression symptoms from user tweets.

We will first provide detail about baseline models and then present our results on SI and FUD task. Later, we will assess the performance of our approach on depression detection task.

In this research, we explored a new dimension of social media in Twitter to identify depressive symptoms. Towards this end, we created a new benchmark dataset (D2S) for identifying fine-grained PHQ-9 emulated depressive symptoms that contains figurative language. We also introduce a robust BERT based MTL framework that jointly learns to automatically discover complementary features required to identify the symptoms with the help of the auxiliary task of figurative usage detection. Our experimental results convincingly show the effectiveness of introducing figurative usage detection for depressive symptoms identification. In future, we aim to enhance the dataset with the other modalities like image and memes to assist the model in better understanding of figurative sense in symptom identification.