Voice for the Voiceless: Active Sampling to Detect Comments Supporting the Rohingyas

Voice for the Voiceless: Active Sampling to Detect Comments Supporting the Rohingyas

Shriphani Palakodety
\AndAshiqur R. KhudaBukhsh
Carnegie Mellon University
\AndJaime G. Carbonell
Carnegie Mellon University
Shriphani Palakodety and Ashiqur R. KhudaBukhsh are equal contribution first authors.

The Rohingya refugee crisis is one of the biggest humanitarian crises of modern times with more than 600,000 Rohingyas rendered homeless according to the United Nations High Commissioner for Refugees. While it has received sustained press attention globally, no comprehensive research has been performed on social media pertaining to this large evolving crisis. In this work, we construct a substantial corpus of YouTube video comments (263,482 comments from 113,250 users in 5,153 relevant videos) with an aim to analyze the possible role of AI in helping a marginalized community. Using a novel combination of multiple Active Learning strategies and a novel active sampling strategy based on nearest-neighbors in the comment-embedding space, we construct a classifier that can detect comments defending the Rohingyas among larger numbers of disparaging and neutral ones. We advocate that beyond the burgeoning field of hate-speech detection, automatic detection of help-speech can lend voice to the voiceless people and make the internet safer for marginalized communities.


Rohingya refugee crisis Active Learning AI for the marginalized

1 Introduction

On 25th August, 2017 extreme violence was allegedly perpetrated against the Rohingya community in Rakhine state, Myanmar [1]. Since then, according to Human Rights Watch [2], more than 671,000 Rohingyas have fled Rakhine state to escape persecution. The Myanmar military’s alleged large-scale campaign of ethnic cleansing [3] has led to one of the fastest-growing refugee crises in the 21st century. However, Myanmar’s military and civilian officials have repeatedly denied any claims of atrocities - which are contradicted by extensive evidence [1] and witness accounts indicating widespread genocide or ethnic cleansing.

Modern history has witnessed multiple instances of mass migration of persecuted communities. Our main goal in this paper is not to argue about highly debated issues like Rohingyas’ citizenship rights or make politically contentious claims about the Myanmar government, the alleged oppressor’s involvement in this crisis. Rather, our goal is to present the first-of-its-kind case-study of what we call a 21st century problem: migrant crisis in the era of the ubiquitous internet, where the global audience can weigh in on the matter, shape public opinion about the persecuted community through social media comments, clamor for justice for the oppressed, mobilize help to the community in distress, and perhaps side with the alleged oppressor and paint a picture of distrust, fear and threat about a persecuted minority. In online forums, persecuted communities may have little or no voice in discussions centered around them because (i) much of the discussion occurs in a global language (e.g., English) in which they may have limited proficiency, or (ii) they may have minimal access to internet, or most importantly, (iii) survival is possibly the highest priority demanding a significant chunk of their resources. Online activities disparaging refugee communities may result in serious real world consequences; prior research has even identified a close, causal link between online hate speech and offline violence targeting refugees [4].


  1. Domain: In this paper, via a substantial corpus constructed using comments on YouTube videos (5,153 videos, 263,482 comments posted by 113,250 users) relevant to the Rohingya refugee crisis, we characterize several key aspects of the discourse and show that a medium as powerful as the internet can create an asymmetric discourse where an (allegedly) oppressed minority may have little or no voice to defend themselves from (possibly misinformed) global vitriol. To the best of our knowledge, this is the first AI-focused comprehensive analysis of the Rohingya refugee crisis. In the last decade, besides the Rohingya immigrant crisis, the world has witnessed migrant crises in central America [5], Venezuela [6], Italy [7], and the long-standing Syrian refugee crisis [8] resulting in the displacement of millions of people. We believe our work in characterizing key aspects of the Rohingya migrant crisis will open the gates for similar AI research in this humanitarian domain.

  2. Voice for the voiceless: Existing discourse moderation tools on social media platforms focus on minimizing hate speech through deletion of hostile content and/or flagging belligerent members. We argue that beyond the important field of hate-speech detection, automatic identification of user-generated web content that champions the cause of a minority can be equally critical for making the internet a safer, better, and healthier place. For a balanced and nuanced discussion on issues around oppressed minorities, lending greater visibility to pro-minority voice (e.g., through pinning or highlighting content) from a large pool of hostile or ambivalent comments is critical and can be greatly facilitated through automatic methods. To this end, we construct a classifier dubbed the voice-for-the-voiceless classifier, that detects content championing the cause of an oppressed minority, in this case the Rohingyas.

  3. Machine Learning: We propose two new active sampling techniques. Our voice-for-the-voiceless classifier is constructed using a novel nearest neighbor active-sampling technique in the comment embedding space which effectively uncovers rare positives in a negatively-skewed corpus. In addition, we demonstrate that our proposed technique can be extended to another novel nearest neighbor active-sampling sampling technique in the user embedding space to identify sympathetic users and effectively uncover rare positives.

2 Concise Overview of the Crisis

In this section, we present a concise timeline to illustrate the sequence of events that led to this massive humanitarian crisis [9]. As shown in the timeline, arguably, the community experienced a long-standing systemic bias which led to this evolving crisis.

1982 In the 135 national ethnic groups listed in new citizenship law, Rohingyas are excluded, effectively rendering them stateless.
Nov 13, 2010 Aung San Suu Kyi, opposition leader and Nobel peace prize winner, is released from house arrest.
Jun, 2012 Religious violence erupts in predominantly Rohingya region Rakhine leaving more than 200 dead and close to 150,000 homeless. In the next three years, more than 112,000 flee to Malaysia by boat.
2014 In Myanmar’s first census in three decades, Rohingyas are excluded.
Nov, 2015 Suu Kyi’s party wins in first election ending military rule. Rohingyas are not allowed to vote let alone contest.
Oct 9, 2016 Rohingya insurgent group Arakan Rohingya Salvation Army (ARSA) claims an attack that killed 9 police officers according to the state media. A massive military crackdown ensues triggering an exodus of 87,000 Rohingyas to Bangladesh.
Aug 25, 2017 State media claims that 12 police officers were killed by ARSA in a coordinated attack on 20 police posts. A large number of Rohingyas flee to Bangladesh as military responds with (allegedly) burning down villages as a part of what they describe as “clearance operations”.
Oct 23, 2017 Since Aug 25, 2017, a continuous stream of Rohingya refugees arrive in Bangladesh with the refugee count reaching more than 600,000.

3 Related Work

Hate-speech detection: There is a growing body of literature on analyzing and detecting hate-speech in social media such as Facebook [10], Twitter [11, 12], Reddit [13] and YouTube [14]. While hate-speech detection and subsequent intervention (by moderating content or flagging users) are extremely helpful in maintaining a positive web environment, these tools are inadequate in this setting. In our problem, a persecuted community is largely absent in an overwhelmingly negative discussion about them, possibly due to language barriers or simply because they lack sufficient access to the web or social media. Detecting comments that advocate their cause is crucial in representing their views.

Active Learning: We drew inspiration from several existing lines of Active Learning research for constructing our voice-for-the-voiceless classifier [15, 16, 17]. Since sequentially labeling and retraining models may not be practically feasible, following [18], we adopted a batch Active Learning setting to expand our pool of labeled samples. As we shall demonstrate, a large majority of comments are unfavorable to the Rohingyas, making this classification task one with significant class imbalance. Active Learning with class imbalance is a widely studied research problem (see, e.g., [19, 20, 21, 22]). Our proposed sampling strategy leverages recent advances in language modeling to obtain comment embeddings [23, 24] and then mines nearest neighbors in the comment embedding space to alleviate the class imbalance issue. Our proposed solution can thus can be considered a skew-specialized Active Learning approach. However, unlike [25], instead of constructing a synthetic sample, our method yields samples from the actual pool of unlabeled data. In addition to nearest-neighbor sampling and the more traditional Active Learning sampling strategy, uncertainty sampling, we considered minority-class certainty sampling since it was found to be useful in reducing class imbalance in short document classification tasks  [26, 27, 28].

Research on migrant crisis: Extensive research on migrant crises including the Rohingya refugee crisis has been performed from a social science perspective [29, 30, 31]. In what follows, we focus on relevant literature with an AI emphasis. Large-scale social media analysis of the Syrian refugee crisis to explore social and communicative networks in Twitter has been performed in [32]. Using a small set of curated Twitter accounts, community detection has been analyzed in [33]. In terms of domain, our work is most similar to [34] where a classifier to label comments favorable to the Rohingyas’ resettlement in Bangladesh was constructed. Our work is different in terms of scale, focus and analysis in the following ways. First, we consider a substantially larger corpus of 263,482 comments on videos retrieved using high-frequency search queries from 19 different countries (listed in Table 1), whereas [34] focuses on 5,000 Bengali tweets generated by Bangladeshis. Presence of multiple countries adds to our linguistic challenges as expression of intent may become more diverse. Second, we provide a comprehensive analysis of the corpus employing topic modeling, user-level analysis to demonstrate under-representation of Rohingya sympathizers in the global discussion, and overall sentiment analysis of the corpus using domain-specific sentiment lexicons. Third, our voice-for-the-voiceless classifier is more nuanced than merely the sentiment towards settlement in one particular country. Finally, we address a critical challenge of mining positive examples in a rare-class learning problem with a novel approach of nearest-neighbor sampling.

Figure 1: Temporal distribution of videos in .

4 Data Collection

Our data collection process consists of the following steps:

  1. We construct a query set, (116 unique queries), by including related queries from Google Trends111https://trends.google.com/trends/?geo=US for the query [Rohingya] from countries listed in Table 1.

  2. For each query in , we obtained the top 200 YouTube video search results. In addition to these, we performed a targeted crawl focusing on the three months time-duration (July 1, 2017 through September 30, 2017) when the crisis reached its peak. For a given month, for each query in , we obtained the top 50 YouTube video search results. Our final consolidated video data set, , consists of 5,153 unique videos.

  3. We used the publicly available YouTube API to crawl the comments for the videos obtaining 263,482 comments posted by 113,250 unique users.

  4. Since contains queries from multiple countries where English is not the native tongue, we expected the comment corpus to be a mixed bag of different languages with English being the predominant one. Hence, we required an automated method to identify the English comments. We used an unsupervised language filtering approach [35]. The size of our English-filtered corpus, denoted as , is 138,978 comments (i.e., more than 50% of the corpus was written in English). We identified German, Hindi, Bengali, Malay, Urdu, French, and Arabic in our non-English corpus indicating a global presence in the discussion.

Countries considered Reason for inclusion
Myanmar Origin country of crisis
Malaysia, Indonesia and Bangladesh Countries that offered most help
Bangladesh, China, Laos, India and Thailand Border sharing countries

Germany, Australia, Austria, Canada, Sweden, Norway, Switzerland, Finland, Ireland and Spain
Granted asylum to Syrian refugees
USA Received 10% of all asylum applications in the OECD countries in 1998-2007 [36].

Table 1: List of countries.

Video and comment statistics: We present views, comments and like-related statistics of our videos in Table 2. The total number of views of all videos exceeded 15 million. As shown in Figure 1, a substantial chunk of the obtained videos come from the month of September, 2017, when the crisis reached its peak. Hence, we believe that our video corpus captured most of the relevant coverage on this issue.

Mean Standard deviation
Views 29,724 132,869
Likes 452 2100
Dislikes 36 145
Comments 51 238

Table 2: Statistics related to videos in .

General perception: We first analyze phrases in the comments that match a set of high-frequency text templates. We focus on: [Rohingya are] (or [Rohingyas are]) and the negated variants ([Rohingya are not] and [Rohingyas are not]). The tokens that follow these templates are visualized in Figure 2 demonstrating that the prevalent perception of Rohingyas was largely negative. The templates were chosen by observing the top phrases that occurred in the corpus. A substantial majority often equated them with terrorists and only a tiny fraction of comments stated they were innocent. When the negated versions were used (Figure 2(b)) however, we noticed that innocent became a high-frequency term. Hence, their innocence is questioned by general commenters. Additionally, Figure 2(b) indicates a debate around the ethnicity of the Rohingyas - several commenters stated that they are neither Burmese nor Bengali, which leads to our next analysis.

(a) Rohingyas are
(b) Rohingyas are not
Figure 2: A word cloud visualization of Rohingyas’ perception. Among 992,841 unique bigrams and 2,465,453 unique trigrams, in terms of frequency, [Rohingyas are] and [Rohingyas are not] rank 188 and 358, respectively.
Figure 3: Resettlement debate. Among 2,465,453 unique trigrams, in terms of frequency, [Send them to] ranks 72 respectively.
Topic 1 (77.5%) Topic 2 (16.3%) Topic 3 (3.2%)
people allah child
muslims quran rapist
rohingya god aisha
myanmar muhammad puberty
india islam mentally
bangladesh prophet muhammedans
Table 3: Most relevant tokens for three major topics discovered in the Rohingya corpus using [37].

The resettlement debate: We focused on the template [send them to] to analyze public perception of where they should be resettled and which community or country is responsible for providing assistance and protection. Figure 3 shows that the issue of origin and resettlement is a highly discussed issue in the corpus. Apart from rich Muslim countries, and obvious choices like neighboring countries India and Bangladesh, we were alarmed to notice that [send them to hell] was also a frequently occurring 4-gram in the corpus; among 3,215,489 unique 4-grams, its percentile rank is 99.75.

Aspects of the discussion: We next employ topic modeling to obtain a fuller picture of the discussion in the corpus. For our topic model analysis, we first filtered comments with 15 tokens or more and then ran the LDA algorithm [37] obtaining best results for topic count of 4 (the 4th topic contained code-mixed incoherent tokens). We discover three main themes: (i) a geopolitical discussion centered around India, Bangladesh, and Myanmar - the three geographically significant countries in the area, (ii) a religion-centric discussion with an appeal for help, and (iii) an anti-Islamic cluster primarily consisting of religion-themed slurs. Table 3 contains a list of the most relevant terms per topic. The relevance score is from [38].

User level analysis: We were curious to examine: is it possible to estimate how many Rohingyas engage with videos where global users post highly negative comments about them? It is not possible to unambiguously identify if a YouTube user is Rohingya or not. However, we identified several Rohingya-focused YouTube channels many of which use the Rohingya language (this language has 1.8 million native speakers as compared to 380 million native speakers of English). Users who predominantly commented on videos hosted by such channels could possibly be Rohingya or Rohingya-sympathizers. Consequently, we divided our set of videos into two mutually exclusive sets: videos that are hosted by Rohingya-focused channels (e.g., Voice of Rohingya) denoted as (1,727 videos from 123 channels), the other, denoted as is the complement of . Videos belonging to (3,426 videos from 1,244 channels) are primarily hosted by News channels (e.g.. BBC, Al Jazeera, CNN) and a few individual contributors.

Of the 113,250 users, 11,326 and 104,973 users commented on and , respectively. The overlap between the two sets was 3,049 users (Jaccard similarity index 0.0269). Due to disparate size of the two sets, we admit that instead of looking at Jaccard similarity, a more interesting follow up research question could be who posts more negative comments? Is it the users who are frequent visitors of but occasionally visit ? Or the other way around? We focus on the 3,049 users who have commented on at least one video belonging to and one video belonging to and define two mutually exclusive user sets: (users with more than 80% of comments posted on videos in ) and (users with more than 80% of comments posted on videos in ).

Positive Negative
peace, great, peaceful, accept, kind, thank, love, care, humanity, innocent terrorist, genocide, war, hate, bad, violence, rape, illegal, evil, shame
Table 4: Manually selected seed words for SENTPROP.

We next obtained English comments made by these two user sets (denoted by ) and (denoted by ). Lexicon-based sentiment analysis is an established tool for computing sentiment scores [39]. In this scheme, tokens are assigned scores and individual documents’ (comments in our case) scores are obtained by combining the constituent token scores. For effective sentiment analysis, obtaining a domain-specific lexicon is crucial [40]. We considered two existing lexicons induced on popular sub-reddits [41] (politics and India sub-reddits) and a new custom lexicon induced on our corpus using 100-dimensional FastText embeddings [23] and a lexicon inducing algorithm (SENTPROP)  [41]. SENTPROP requires a set of positive and negative seed words (listed in Table 4). Our test for positive or negative adds the individual token scores and if the cumulative comment score is greater than 3 (or less than -3), the comment is considered positive (or negative).

As shown in Table 5, across all three lexicons, we found that posted substantially fewer negative comments than positive comments in comparison to where the ratio of positive to negative comments was reversed. Human evaluation on a random sample of 200 comments revealed that a larger share of negative comments posted by were disparaging to Rohingyas, and the small fraction of negative 200 comments posted by were mostly against the Myanmar government’s (alleged) atrocities.

Politics India Induced on
sub-reddit sub-reddit
pos = 30.40% pos = 25.20% pos = 48.57%
neg = 2.38% neg = 1.85% neg = 2.88%
pos = 10.01%, pos = 12.45% pos = 18.55%
neg = 32.92% neg = 32.76% neg = 35.30%

Table 5: Percentage of positive and negative comments using lexicons presented in  [41] and a lexicon induced on .

5 Voice for the Voiceless Classifier

We start with pointing out a subtle but important distinction: Voice-for-the-voiceless speech is not absence of hate speech. The goals of a hate-speech classifier and our voice-for-the-voiceless classifier are different and complement each other. Identifying hateful content for possible moderation certainly has a positive role in making the internet a safer place for a vulnerable community. However, surfacing comments marked as not hate speech does not necessarily lend a voice to the voiceless. For instance, say a user from India respectfully states that India is an over-populated country and does not have enough resources for Rohingyas. This is clearly not hate-speech against the Rohingyas, but it is also not voicing the concerns of the voiceless (the Rohingyas).

We next present a definition of voice-for-the-voiceless speech and provide examples picked from the corpus or written by us (italicized) to illustrate the point. Understandably, the italicized comments succinctly express a given condition in correct English while examples from the corpus might contain grammatical errors.

Definition 1: A comment is marked as voice-for-the-voiceless speech, if the comment

  1. actively seeks to help one or more persons belonging to the (allegedly) oppressed minority (e.g., [how can we help the Rohingyas])

  2. urges other people or organizations (such as the UN) to help the (allegedly) oppressed minorities (e.g., UN should help Rohingyas)

  3. urges other people to come forward and assist or take a humane stance (e.g., [value the humanity thy r migrating for lives not for luxury])

  4. advocates for the (allegedly) oppressed community’s rights (e.g., Rohingyas should get Myanmar citizenship)

  5. condemns the atrocities against the (allegedly) oppressed (e.g., Myanmar government shame on you)

  6. sympathizes with the (allegedly) oppressed community’s plight (e.g., [This just breaks my heart. I wish I could help. All these people commenting about muslims and hindus should be ashamed. The bottom line is these are humans being killed, children being killed. It doesnt matter who started it. It needs to stop!])

and a comment is not voice-for-the-voiceless if it

  1. expresses violent intent to a specific entity (including the alleged oppressors) or broad bias against any religious community (e.g., [Pakistan please nuke Myanmar bhudists])

  2. calls for aggressive action against the oppressed community (e.g., deport them all)

  3. demonstrates proverbial whataboutism (e.g., [what about Yazidis])

  4. paints a general picture that the community is a threat (e.g.,[Rohingyas are terrorists])

  5. shows solidarity with the (alleged) oppressors (e.g., [well done Myanmar])

Active learning with class imbalance: Typically, in Active Learning, a seed set of samples is used to construct a classifier which then samples from the unlabeled pool and seeks labels [42]. For better generalizability of the classifier in the wild, it is often critical that the training set is (i) balanced, i.e., contains sufficient number of examples from both classes (ii) diverse, i.e. captures a wide variety of data points we may encounter in the wild.

We faced the following two research challenges:

  • How to obtain a sufficient number of positive comments in a corpus largely disparaging to the Rohingyas?

  • How to cover a wide range of aspects of positive (and also negative) comments in our training set so that the classifier performs well in the wild?

Active Learning meets document embeddings: Note that, key phrases (e.g., send them to, deport them all, breaks my heart) may express user intent. However, in a corpus largely filled with negative comments and with a high variance in English proficiency among the contributors, simple mining techniques using exact phrase-level match may not yield sufficient number of positives. In the extreme case, the phrase we are looking for, may not even yield a single exact match. Moreover, the matched comments run the risk of being highly similar to each other and hence may not capture the entire space of varied expressions. Using semantic embeddings to find comments similar to an example positive phrase (or negative phrase) may be effective; however, with a smaller corpus, semantic embeddings may be more prone to inaccuracy (alleviated with a human-in-the-loop in the Active Learning setting). In this work, we meld recent advances in sentence embeddings with Active Learning and propose a novel Active Sampling technique to augment our seed set. The model described in [43] is used to obtain a real-valued vector for each comment in the corpus and used to retrieve a comment’s nearest neighbors in this embedding space. In conjunction with random sampling, this technique helps us discover a broader, more diverse set of positive examples and helps us combat extreme class imbalance.

5.1 Our Active Learning Approach

As illustrated in Figure 4, our approach consists of the following steps:

  1. Construct a seed set of positive and negative comments.

  2. Expand the seed set by randomly sampling comments from the unseen corpus.

  3. Obtain real valued embeddings for the comments, find the nearest neighbors of the seed set and include them in the corpus (new technique presented in this paper).

  4. Further expand using minority-class certainty sampling.

  5. Perform final expansion using uncertainty sampling.

Seed set: Our seed set (6 positives, 5 negatives) consists of the same set of examples presented in Definition 1.

Random sampling: In order to have better coverage, we randomly sampled 300 comments and labeled them. We obtained 32 positives and 268 negatives, i.e., 10.66% positives. For evaluating our sampling strategies this acts as the baseline. All rounds of manual labeling were performed by two annotators proficient in English. The annotators were presented with the definition and example seed set. They were first asked to label independently, and then allowed to discuss and resolve the disagreed labels. We obtained strong agreement in every round (lowest Cohen’s coefficient across all rounds was 0.8766 indicating strong inter-rater agreement).

Nearest-neighbor sampling (NN sampling): We use a well-known document embedding model from [43] to obtain a real-valued embedding for each comment in our corpus. Starting from the seed set, we obtained the seed embeddings and then obtained the comments from the unlabeled corpus whose embeddings were closest to the seed embeddings (i.e. the nearest neighbors). following  [44, 43] cosine distance was used as the distance metric.

can someone tell me where i can help charity to them
all the countries should take a stand for these people and force mayanmar government to accept them
No country is too small to take on refugees and camp them for period of time until the problem is solved by the world leaders making every problem a political issue is just creating dangerous matters for the poor public in some countries animals are cared for more than the humans
sanction myanmar till they understand international law and give up ethnic cleansing
Table 6: Random sample of positive comments obtained using the nearest-neighbor sampling.
Figure 4: System diagram.

Advantages of our technique: First, it allows flexibility while specifying an example comment. Without sufficient knowledge of the corpus, it may be difficult to uncover a rare positive satisfying a particular aspect of the target concept. Some of the examples in our seed set (italicized) did not have an exact match in the actual corpus yet the semantic similarity technique uncovered similar rare positives. Second, our embedding method [43] employs sub-word information and thus is robust to spelling variations or outright erroneous spellings (e.g., Buddhists was incorrectly spelled as bhudists). In addition, it can handle both short and long comments. In Table 6, we list a few positives we obtained through our technique to highlight its effectiveness. Out of the 300 nearest neighbors obtained from 6 positive seed comments, we obtained 101 unique positives (33.67%) from 292 unique comments. The large number of unique comments indicates that our technique found a diverse set of samples. We obtained more than 3x the number of positives than discovered by random sampling (10.66%). Our method also uncovered a diverse set of negative comments about the Rohingyas; a representative sample is listed in Table 7 to emphasize why we believed the community required protection from online attacks.

We next trained a Support Vector Machine (SVM) classifier on our consolidated labeled data set with 164 unique positives and 662 unique negatives (we also included the randomly sampled instances) with token unigram, bigram and trigram features.

rohingyas are very strong in breeding
rohingya muslims are terrorists they have been killing buddhist from 1947 onwards they deserve whatever they are getting
kick them all out fuc ing like swines and changing our demographics
just kill them all soon because they are terrorists bastards
Table 7: Random sample of negative comments obtained using the nearest-neighbor sampling.
Performance Seed set + random sampling + + Certainty + Uncertainty
measure NN in the embedding space sampling sampling
Precision 67.17 9.90% 71.27 5.23% 73.65 3.45%
Recall 32.35 7.65% 72.52 4.23% 79.39 3.72%
Accuracy 82.04 2.34% 75.95 3.10% 75.38 2.76%
F1 score 43.02 7.90% 71.75 4.32% 76.34 2.77%
AUC 83.61 2.88% 83.64 2.84% 83.67 2.61%

Table 8: Voice-for-the-voiceless classifier performance.
Performance SVM (n gram) SVM (n gram + embedding)
Precision 73.65 3.45% 76.49 3.51
Recall 79.39 3.72% 80.30 3.73
Accuracy 75.38 2.76% 77.71 2.56
F1 score 76.34 2.77% 78.28 2.71
AUC 83.67 2.61% 85.91 2.32

Table 9: Model improvement.

Certainty sampling. While our NN sampling technique effectively uncovered a considerable number of rare positives, the class imbalance was still present with positives merely constituting 19.85% of the labeled data set. We bridged this gap through employing certainty sampling, a sampling technique first proposed in [27, 26]. In batch certainty sampling, we pick (set to 1000) unlabeled samples with highest predicted probability for the minority class. In this step, we closed the gap between the number of positives and negatives as we obtained 611 unique positives and 389 unique negatives. We re-train our classifier with our consolidated data set of 775 positives and 1,051 negatives.

Uncertainty sampling. Finally, we used uncertainty sampling to add 1000 more samples where the predicted class probability was close to 0.5. Our final data set consists of 2790 comments with roughly equal numbers of positives and negatives (1,391 positives and 1,399 negatives). Hence, via (i) active learning, (ii) combining multiple existing sampling techniques and (iii) our proposed nearest-neighbor sampling, we succeeded in addressing class imbalance.

keep helping these poor innocent people the myanmar government is really kind of like animals not like human beings so thats why they genocide innocent people in myanmar
i am from nepal where buddha was born i have seen buddhist who is so kindful n helpful but i never seen buddhist who murder poor n innocent people i really fell so shameful that they are killing innocent poor people children and old age people they are torturing kid and womens for god shake please stop this violence
thank you so for news today and vi want full human rights in arakan myanmar and stop nvc card and ples vi want myanmar army goverment to the icc kireminal courd justice and vi want full setizenthip in arakan myanmar vi no bangali vi setizenthip in arakan myanmar and myanmar army reped womens rohingya and etnik kilingsing of rohingya and genocide of rohingya and ples vi want hlep from un konsiel and from human rights wohc ples hlep stop genocide of rohingya and humanty in myanmar and thank you so lot god bles you all
its genocide ethnic cleansing brutality reach the level of where words cannot describe its inhuman government of myanmar monk are killing babies and womens into pieces sushi should be punished by court of law
Table 10: Performance in the wild.

Classifier performance. We used a 90/10 train/test split; we trained an SVM classifier [45] with token n-grams as features (with up to 3) and evaluated the performance on the test set. Since performance can be sensitive to individual test/train splits, we repeated the experiment 100 times on 100 randomly chosen test-train splits. Our intermediate classifiers and final classifier’s performance is summarized in Table 8. In a class-imbalanced problem, simply predicting the majority class can yield a high accuracy, F1 score is the more meaningful measure. After each round of labeled data acquisition, we noticed a steady rise in the F1 score with a final performance of 76.50 2.85%. We obtain further improvement by adding comment embeddings as features as shown in Table 9.

Figure 5: Breakdown of positive comments found in the wild. A single comment can satisfy multiple criteria.

Performance in the wild: Our goal is to identify comments supporting a persecuted minority in the wild. We ran our classifier on the unlabeled corpus (i.e., on comments neither belonging to the train or test set) and conducted a human evaluation of the top 100 comments predicted as voice-for-the-voiceless ranked by confidence. Of the 100 comments, 88 were annotated as positives which indicates the classifier would substantially reduce manual effort (random sampling only found 10.66 positives) to find supportive comments for Rohingyas in the wild. In Figure 5, we present the breakdown of the positives into six broad categories as presented in our definition (active help, appeal to organizations, humanitarian call, vocalize rights, condemn oppressor, express sympathy). We found that our classifier found comments belonging to all broad categories from the wild. In Table 10, we highlight few randomly sampled comments to illustrate two points. First, we draw attention to the comment highlighted in blue. We suspect that this comment is written by a Rohingya YouTube user. Broken sentences, grammatical disfluency and a large number of spelling errors indicate how the language barrier may make it difficult for a marginalized community to voice their opinion. Our classifier correctly labeled this comment with high confidence, indicating our approach holds promise in surfacing minority voices. However, since the minority is experiencing (alleged) persecution, it is reasonable to observe measured expression of rational negativity while condemning the (alleged) oppressor. The comment marked in red opens up an interesting philosophical question: where should we draw the line? For instance, one particular comment supporting the Rohingyas found in the wild used a gendered insult to refer to the Prime Minister of Myanmar which our annotators marked as negative. Hence, we conclude this analysis by saying that our classifier holds promise to substantially lessen the burden of moderators to automatically find content supporting a minority, however it may require some further supervision and human judgement to ensure fairness. This research question merits deeper exploration.

5.2 Voice-for-the-voiceless community

We conclude our paper with a small exploratory study on the possibility of finding rare positives through user embeddings. For a given user, we constructed the corresponding user embedding using the embeddings of all comments posted by the user, normalizing them and finally averaging these normalized embeddings. Our user-focused nearest-neighbor sampling consists of the following steps:
(1) Obtain top positive comments (ranked by predicted class probability) predicted by the voice-for-the-voiceless classifier. (2) Next, identify the set of unique users, , who posted these comments. (3) Next, for each user in , obtain nearest neighbors in the user embedding space. (4) Finally, sample comments from the nearest neighbors.

We set both and to 10. We obtained 9 unique users who posted the top 10 comments. Of the 90 nearest neighbors, 88 were unique indicating our user-focused nearest neighbor sampling was able to uncover a diverse set of users. We next randomly sampled 300 comments posted by the nearest neighbors. Our hypothesis was if our user embedding-based sampling indeed identifies a set of positive users, the sampled comments will have more positives than the baseline (random sampling fetched 10.66% positives). Our annotators identified 105 positives (35%). Hence, our user-focused sampling performed 3x better than the baseline. Hence, both Active Sampling strategies proposed in this paper substantially outperformed the baseline in finding rare positives supporting a persecuted minority. We conclude with the hope that this work will motivate further AI research in this important humanitarian domain.


  • [1] Nathan Thompson. Myanmar: Un fact-finding mission releases its full account of massive violations by military in rakhine, kachin and shan statesa, 2018. [Online; accessed 12-May-2019].
  • [2] Human Rights Watch. Rohingya crisis. [Online; accessed 12-May-2019].
  • [3] Chris Beyrer and Adeeba Kamarulzaman. Ethnic cleansing in myanmar: the rohingya crisis and human rights. The Lancet, 390(10102):1570–1573, 2017.
  • [4] Karsten Müller and Carlo Schwarz. Fanning the flames of hate: Social media and hate crime. Available at SSRN 3082972, 2018.
  • [5] BBC News. Us-mexico border official says migrant crisis ‘at breaking point’, 2019. [Online; accessed 12-May-2019].
  • [6] Mark Leon Goldberg. Venezuela is a refugee crisis, 2019. [Online; accessed 12-May-2019].
  • [7] BBC News. Italy migrant crisis: Government passes tough bill, 2019. [Online; accessed 12-May-2019].
  • [8] Office of the United Nations High Commissioner for Refugees. Seven years on: Timeline of the syria crisis, 2018. [Online; accessed 12-May-2019].
  • [9] Katie Hunt. Rohingya crisis: How we got here, 2017. [Online; accessed 12-May-2019].
  • [10] Fabio Del Vigna12, Andrea Cimino23, Felice Dell’Orletta, Marinella Petrocchi, and Maurizio Tesconi. Hate me, hate me not: Hate speech detection on facebook. In Proceedings of the First Italian Conference on Cybersecurity, 2017.
  • [11] Thomas Davidson, Dana Warmsley, Michael Macy, and Ingmar Weber. Automated hate speech detection and the problem of offensive language. In Eleventh International AAAI Conference on Web and Social Media, 2017.
  • [12] Pinkesh Badjatiya, Shashank Gupta, Manish Gupta, and Vasudeva Varma. Deep learning for hate speech detection in tweets. In Proceedings of the 26th International Conference on World Wide Web Companion, pages 759–760. International World Wide Web Conferences Steering Committee, 2017.
  • [13] Eshwar Chandrasekharan, Umashanthi Pavalanathan, Anirudh Srinivasan, Adam Glynn, Jacob Eisenstein, and Eric Gilbert. You can’t stay here: The efficacy of reddit’s 2015 ban examined through hate speech. Proceedings of the ACM on Human-Computer Interaction, 1(CSCW):31, 2017.
  • [14] Karthik Dinakar, Birago Jones, Catherine Havasi, Henry Lieberman, and Rosalind Picard. Common sense reasoning for detection, prevention, and mitigation of cyberbullying. ACM Transactions on Interactive Intelligent Systems (TiiS), 2(3):18, 2012.
  • [15] Nicholas Roy and Andrew McCallum. Toward optimal active learning through monte carlo estimation of error reduction. ICML, Williamstown, pages 441–448, 2001.
  • [16] Yoram Baram, Ran El Yaniv, and Kobi Luz. Online choice of active learning algorithms. Journal of Machine Learning Research, 5(Mar):255–291, 2004.
  • [17] Pinar Donmez, Jaime G Carbonell, and Paul N Bennett. Dual strategy active learning. In European Conference on Machine Learning, pages 116–127. Springer, 2007.
  • [18] Liu Yang and Jaime Carbonell. Buy-in-bulk active learning. In Advances in neural information processing systems, pages 2229–2237, 2013.
  • [19] Burr Settles and Mark Craven. An analysis of active learning strategies for sequence labeling tasks. In Proceedings of EMNLP, pages 1070–1079, 2008.
  • [20] Hieu T Nguyen and Arnold Smeulders. Active learning using pre-clustering. In Proceedings of the twenty-first ICML, page 79. ACM, 2004.
  • [21] Pinar Donmez and Jaime G Carbonell. Paired-sampling in density-sensitive active learning. In Proceedings of the 10th International Symposium on Artificial Intelligence and Mathematics, 2008.
  • [22] Katrin Tomanek and Udo Hahn. Reducing class imbalance during active learning for named entity annotation. In Proceedings of the fifth international conference on Knowledge capture, pages 105–112. ACM, 2009.
  • [23] Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. In Proceedings of the 15th EACL: Volume 2, Short Papers, pages 427–431, 2017.
  • [24] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. Transactions of the Association for Computational Linguistics, 5:135–146, 2017.
  • [25] Seyda Ertekin. Learning in extreme conditions: Online and active learning with massive, imbalanced and noisy data. PhD thesis, The Pennsylvania State University, 2009.
  • [26] Josh Attenberg, Prem Melville, and Foster Provost. A unified approach to active dual supervision for labeling features and examples. In ECML/PKDD, pages 40–55. Springer, 2010.
  • [27] Vikas Sindhwani, Prem Melville, and Richard D Lawrence. Uncertainty sampling and transductive experimental design for active dual supervision. In Proceedings of the 26th ICML, pages 953–960. ACM, 2009.
  • [28] Ashiqur R KhudaBukhsh, Paul N Bennett, and Ryen W White. Building effective query classifiers: a case study in self-harm intent detection. In Proceedings of the 24th ACM CIKM conference, pages 1735–1738. ACM, 2015.
  • [29] XChange.org. The rohingya survey 2017, 2017. [Online; accessed 12-May-2019].
  • [30] Abhishek Bhatia, Ayesha Mahmud, Arlan Fuller, Rebecca Shin, Azad Rahman, Tanvir Shatil, Mahmuda Sultana, KA M Morshed, Jennifer Leaning, and Satchit Balsari. The rohingya in cox’s bazar: When the stateless seek refuge. Health and human rights, 20(2):105, 2018.
  • [31] Abul Milton, Mijanur Rahman, Sumaira Hussain, Charulata Jindal, Sushmita Choudhury, Shahnaz Akter, Shahana Ferdousi, Tafzila Mouly, John Hall, and Jimmy Efird. Trapped in statelessness: Rohingya refugees in bangladesh. International journal of environmental research and public health, 14(8):942, 2017.
  • [32] Marc Lynch, Deen Freelon, and Sean Aday. Syria’s socially mediated civil war. Universitäts-und Landesbibliothek Sachsen-Anhalt, 2014.
  • [33] Derek O’Callaghan, Nico Prucha, Derek Greene, Maura Conway, Joe Carthy, and Pádraig Cunningham. Online social media in the syria conflict: Encompassing the extremes and the in-betweens. In Proceedings of the 2014 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining, pages 409–416. IEEE Press, 2014.
  • [34] Hemayet Ahmed Chowdhury, Tanvir Alam Nibir, Md Islam, et al. Sentiment analysis of comments on rohingya movement with support vector machine. arXiv preprint arXiv:1803.08790, 2018.
  • [35] Shriphani Palakodety, Ashiqur R. KhudaBukhsh, and Jaime G. Carbonell. Kashmir: A computational analysis of the voice of peace. CoRR, abs/1909.12940, 2019.
  • [36] Chad C Haddal. Refugee and asylum-seeker inflows in the united states and other oecd member states. Congressional Research Service, Library of Congress, 2009.
  • [37] David M Blei, Andrew Y Ng, and Michael I Jordan. Latent dirichlet allocation. JMLR, 3(Jan):993–1022, 2003.
  • [38] Carson Sievert and Kenneth Shirley. Ldavis: A method for visualizing and interpreting topics. In Proceedings of the workshop on interactive language learning, visualization, and interfaces, pages 63–70, 2014.
  • [39] Brendan O’Connor, Ramnath Balasubramanyan, Bryan R Routledge, and Noah A Smith. From tweets to polls: Linking text sentiment to public opinion time series. In Fourth International AAAI Conference on Weblogs and Social Media, 2010.
  • [40] Leonid Velikovich, Sasha Blair-Goldensohn, Kerry Hannan, and Ryan McDonald. The viability of web-derived polarity lexicons. In NAACL, pages 777–785. Association for Computational Linguistics, 2010.
  • [41] William L Hamilton, Kevin Clark, Jure Leskovec, and Dan Jurafsky. Inducing domain-specific sentiment lexicons from unlabeled corpora. In EMNLP, volume 2016, page 595. NIH Public Access, 2016.
  • [42] Burr Settles. Active learning literature survey. Technical report, University of Wisconsin-Madison Department of Computer Sciences, 2009.
  • [43] Matteo Pagliardini, Prakhar Gupta, and Martin Jaggi. Unsupervised Learning of Sentence Embeddings using Compositional n-Gram Features. In NAACL 2018, 2018.
  • [44] Dorottya Demszky, Nikhil Garg, Rob Voigt, James Zou, Matthew Gentzkow, Jesse Shapiro, and Dan Jurafsky. Analyzing polarization in social media: Method and application to tweets on 21 mass shootings. In Proceedings of the 17th Annual NAACL), 2019.
  • [45] Simon Tong and Daphne Koller. Support vector machine active learning with applications to text classification. JMLR, 2(Nov):45–66, 2001.
Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
Add comment
Loading ...
This is a comment super asjknd jkasnjk adsnkj
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test description