Large-scale Analysis of Counseling Conversations: An Application of Natural Language Processing to Mental Health Tim Althoff∗, Kevin Clark∗, Jure Leskovec Stanford University {althoff, kevclark, jure}@cs.stanford.edu Abstract Mental illness is one of the most pressing pub- lic health issues of our time. While counsel- ing and psychotherapy can be effective treat- ments, our knowledge about how to conduct successful counseling conversations has been limited due to lack of large-scale data with la- beled outcomes of the conversations. In this paper, we present a large-scale, quantitative study on the discourse of text-message-based counseling conversations. We develop a set of novel computational discourse analysis meth- ods to measure how various linguistic aspects of conversations are correlated with conver- sation outcomes. Applying techniques such as sequence-based conversation models, lan- guage model comparisons, message cluster- ing, and psycholinguistics-inspired word fre- quency analyses, we discover actionable con- versation strategies that are associated with better conversation outcomes. 1 Introduction Mental illness is a major global health issue. In the U.S. alone, 43.6 million adults (18.1%) experience mental illness in a given year (National Institute of Mental Health, 2015). In addition to the person di- rectly experiencing a mental illness, family, friends, and communities are also affected (Insel, 2008). In many cases, mental health conditions can be treated effectively through psychotherapy and coun- seling (World Health Organization, 2015). However, it is far from obvious how to best conduct counsel- ing conversations. Such conversations are free-form without strict rules, and involve many choices that *Both authors contributed equally to the paper. could make a difference in someone’s life. Thus far, quantitative evidence for effective conversation strategies has been scarce, since most studies on counseling have been limited to very small sample sizes and qualitative observations (e.g., Labov and Fanshel, (1977); Haberstroh et al., (2007)). How- ever, recent advances in technology-mediated coun- seling conducted online or through texting (Haber- stroh et al., 2007) have allowed counseling ser- vices to scale with increasing demands and to col- lect large-scale data on counseling conversations and their outcomes. Here we present the largest study on counseling conversation strategies published to date. We use data from an SMS texting-based counseling service where people in crisis (depression, self-harm, sui- cidal thoughts, anxiety, etc.), engage in therapeutic conversations with counselors. The data contains millions of messages from eighty thousand counsel- ing conversations conducted by hundreds of coun- selors over the course of one year. We develop a set of computational methods suited for large-scale dis- course analysis to study how various linguistic as- pects of conversations are correlated with conversa- tion outcomes (collected via a follow-up survey). We focus our analyses on counselors instead of individual conversations because we are interested in general conversation strategies rather than prop- erties of specific issues. We find that there are sig- nificant, quantifiable differences between more suc- cessful and less successful counselors in how they conduct conversations. Our findings suggest actionable strategies that are associated with successful counseling: 463 Transactions of the Association for Computational Linguistics, vol. 4, pp. 463–476, 2016. Action Editor: Lillian Lee. Submission batch: 12/2015; Revision batch: 3/2016; 7/2016 Published 8/2016. c©2016 Association for Computational Linguistics. Distributed under a CC-BY 4.0 license. 1. Adaptability (Section 5): Measuring the dis- tance between vector representations of the lan- guage used in conversations going well and go- ing badly, we find that successful counselors are more sensitive to the current trajectory of the conversation and react accordingly. 2. Dealing with Ambiguity (Section 6): We de- velop a clustering-based method to measure differences in how counselors respond to very similar ambiguous situations. We learn that successful counselors clarify situations by writ- ing more, reflect back to check understanding, and make their conversation partner feel more comfortable through affirmation. 3. Creativity (Section 6.3): We quantify the di- versity in counselor language by measuring cluster density in the space of counselor re- sponses and find that successful counselors re- spond in a more creative way, not copying the person in distress exactly and not using too generic or “templated” responses. 4. Making Progress (Section 7): We develop a novel sequence-based unsupervised conversa- tion model able to discover ordered conversa- tion stages common to all conversations. Ana- lyzing the progression of stages, we determine that successful counselors are quicker to get to know the core issue and faster to move on to collaboratively solving the problem. 5. Change in Perspective (Section 8): We de- velop novel measures of perspective change us- ing psycholinguistics-inspired word frequency analysis. We find that people in distress are more likely to be more positive, think about the future, and consider others, when the coun- selors bring up these concepts. We further show that this perspective change is associated with better conversation outcomes consistent with psychological theories of depression. Further, we demonstrate that counseling success on the level of individual conversations is predictable using features based on our discovered conversation strategies (Section 9). Such predictive tools could be used to help counselors better progress through the conversation and could result in better counseling practices. The dataset used in this work has been re- leased publicly and more information on dataset ac- cess can be found at http://snap.stanford. edu/counseling. Although we focus on crisis counseling in this work, our proposed methods more generally apply to other conversational settings and can be used to study how language in conversations relates to con- versation outcomes. 2 Related Work Our work relates to two lines of research: Therapeutic Discourse Analysis & Psycholinguis- tics. The field of conversation analysis was born in the 1960s out of a suicide prevention center (Sacks and Jefferson, 1995; Van Dijk, 1997). Since then conversation analysis has been applied to various clinical settings including psychotherapy (Labov and Fanshel, 1977). Work in psycholinguistics has demonstrated that the words people use can re- veal important aspects of their social and psycho- logical worlds (Pennebaker et al., 2003). Previous work also found that there are linguistic cues as- sociated with depression (Ramirez-Esparza et al., 2008; Campbell and Pennebaker, 2003) as well as with suicude (Pestian et al., 2012). These find- ings are consistent with Beck’s cognitive model of depression (1967; cognitive symptoms of depres- sion precede the affective and mood symptoms) and with Pyszczynski and Greenberg’s self-focus model of depression (1987; depressed persons engage in higher levels of self-focus than non-depressed per- sons). In this work, we propose an operationalized psy- cholinguistic model of perspective change and fur- ther provide empirical evidence for these theoretical models of depression. Large-scale Computational Linguistics Applied to Conversations. Large-scale studies have re- vealed subtle dynamics in conversations such as co- ordination or style matching effects (Niederhoffer and Pennebaker, 2002; Danescu-Niculescu-Mizil, 2012) as well as expressions of social power and status (Bramsen et al., 2011; Danescu-Niculescu- Mizil et al., 2012). Other studies have connected writing to measures of success in the context of re- quests (Althoff et al., 2014), user retention (Althoff and Leskovec, 2015), novels (Ashok et al., 2013), and scientific abstracts (Guerini et al., 2012). Prior 464 http://snap.stanford.edu/counseling http://snap.stanford.edu/counseling work has modeled dialogue acts in conversational speech based on linguistic cues and discourse coher- ence (Stolcke et al., 2000). Unsupervised machine learning models have also been used to model con- versations and segment them into speech acts, top- ical clusters, or stages. Most approaches employ Hidden Markov Model-like models (Barzilay and Lee, 2004; Ritter et al., 2010; Paul, 2012; Yang et al., 2014) which are also used in this work to model progression through conversation stages. Very recently, technology-mediated counseling has allowed the collection of large datasets on coun- seling. Howes et al. (2014) find that symptom sever- ity can be predicted from transcript data with com- parable accuracy to face-to-face data but suggest that insights into style and dialogue structure are needed to predict measures of patient progress. Counseling datasets have also been used to predict the conversa- tion outcome (Huang, 2015) but without modeling the within-conversation dynamics that are studied in this work. Other work has explored how novel inter- faces based on topic models can support counselors during conversations (Dinakar et al., 2014a; 2014b; 2015; Chen, 2014). Our work joins these two lines of research by de- veloping computational discourse analysis methods applicable to large datasets that are grounded in ther- apeutic discourse analysis and psycholinguistics. 3 Dataset Description In this work, we study anonymized counseling con- versations from a not-for-profit organization provid- ing free crisis intervention via SMS messages. Text- based counseling conversations are particularly well suited for conversation analysis because all interac- tions between the two dialogue partners are fully ob- served (i.e., there are no non-textual or non-verbal cues). Moreover, the conversations are important, constrained to dialogue between two people, and outcomes can be clearly defined (i.e., we follow up with the conversation partner as to whether they feel better afterwards), which enables the study of how conversation features are associated with actual out- comes. Counseling Process. Any person in distress can text the organization’s public number. Incoming re- quests are put into a queue and an available coun- Dataset statistics Conversations 80,885 Conversations with survey response 15,555 (19.2%) Messages 3.2 million Messages with survey response 663,026 (20.6%) Counselors 408 Messages per conversation* 42.6 Words per message* 19.2 Table 1: Basic dataset statistics. Rows marked with * are computed over conversations with survey responses. selor picks the request from the queue and engages with the incoming conversation. We refer to the cri- sis counselor as the counselor and the person in dis- tress as the texter. After the conversation ends, the texter receives a follow-up question (“How are you feeling now? Better, same, or worse?”) which we use as our conversation quality ground-truth (we use binary labels: good versus same/worse, since we care about improving the situation). In contrast to previous work that has used human judges to rate a caller’s crisis state (Kalafat et al., 2007), we di- rectly obtain this feedback from the texter. Further- more, the counselor fills out a post-conversation re- port (e.g., suicide risk, main issue such as depres- sion, relationship, self-harm, suicide, etc.). All crisis counselors receive extensive training and commit to weekly shifts for a full year. Dataset Statistics. Our dataset contains 408 coun- selors and 3.2 million messages in 80,885 conversa- tions between November 2013 and November 2014 (see Table 1). All system messages (e.g., instruc- tions), as well as texts that contain survey responses (revealing the ground-truth label for the conversa- tion) were filtered out. Out of these conversations, we use the 15,555, or 19.2%, that contain a ground- truth label (whether the texter feels better or the same/worse after the conversation) for the follow- ing analyses. Conversations span a variety of issues of different difficulties (see rows one and two of Ta- ble 2). Approval to analyze the dataset was obtained from the Stanford IRB. 4 Defining Counseling Quality The primary goal of this paper is to study strategies that lead to conversations with positive outcomes. Thus, we require a ground-truth notion of conver- sation quality. In principle, we could study individ- 465 NA Depressed Relationship Self harm Family Suicide Stress Anxiety Other Success rate 0.556 0.612 0.659 0.672 0.711 0.573 0.696 0.671 0.537 Frequency 0.200 0.200 0.089 0.074 0.071 0.063 0.041 0.039 0.035 Frequency with more successful counselors 0.203 0.199 0.089 0.067 0.072 0.061 0.048 0.042 0.030 Frequency with less successful counselors 0.223 0.208 0.087 0.070 0.067 0.056 0.030 0.032 0.028 Table 2: Frequencies and success rates for the nine most common conversation issues (NA: Not available). On average, more and less successful counselors face the same distribution of issues. ual conversations and aim to understand what fac- tors make the conversation partner (texter) feel bet- ter. However, it is advantageous to focus on the conversation actor (counselor) instead of individual conversations. There are several benefits of focusing analy- ses on counselors (rather than individual conversa- tions): First, we are interested in general conversa- tion strategies rather than properties of main issues (e.g., depression vs. suicide). While each conver- sation is different and will revolve around its main issue, we assume that counselors have a particular style and strategy that is invariant across conversa- tions. Second, we assume that conversation qual- ity is noisy. Even a very good counselor will face some hard conversations in which they do every- thing right but are still unable to make their conver- sation partner feel better. Over time, however, the “true” quality of the counselor will become appar- ent. Third, our goal is to understand successful con- versation strategies and to make use of these insights in counselor training. Focusing on the counselor is helpful in understanding, monitoring, and improv- ing counselors’ conversation strategies. More vs. Less Successful Counselors. We split the counselors into two groups and then compare their behavior. Out of the 113 counselors with more than 15 labeled conversations of at least 30 messages each, we use the most successful 40 counselors as “more successful” counselors and the bottom 40 as “less successful” counselors. Their average success rates are 66.3-85.5% and 42.1-58.6%, respectively. While the counselor-level analysis is of primary con- cern, we will also differentiate between counselor behavior in “positive” versus “negative” conversa- tions (i.e., those that will eventually make the texter feel better vs. not). Thus, in the remainder of the paper we differentiate between more vs. less suc- cessful counselors and positive vs. negative conver- 0-20 20-40 40-60 60-80 80-100 Portion of conversation (% of messages) 12 14 16 18 20 22 24 A ve ra ge m es sa ge le ng th More successful counselors, positive conversations More successful counselors, negative conversations Less successful counselors, positive conversations Less successful counselors, negative conversations Figure 1: Differences in counselor message length (in #tokens) over the course of the conversation are larger between more and less successful counselors (blue cir- cle/red square) than between positive and negative con- versations (solid/dashed). Error bars in all plots corre- spond to bootstrapped 95% confidence intervals using the member bootstrapping technique from Ren et al. (2010). sations. Studying the cross product of counselors and conversations allows us to gain insights on how both groups behave in positive and negative conver- sations. For example, Figure 1 illustrates why differ- entiating between counselors and as well as conver- sations is necessary: differences in counselor mes- sage length over the course of the conversation are bigger between more and less successful counselors than between positive and negative conversations. Initial Analysis. Before focusing on detailed anal- yses of counseling strategies we address two impor- tant questions: Do counselors specialize in certain issues? And, do successful counselors appear suc- cessful only because they handle “easier” cases? To gain insights into the “specialization hypoth- esis” we make use the counselor annotation of the main issue (depression, self-harm, etc.). We com- pare success rates of counselors across different issues and find that successful counselors have a higher fraction of positive conversations across all issues and that less successful counselors typically do not excel at a particular issue. Thus, we conclude 466 that counseling quality is a general trait or skill and supporting that the split into more and less success- ful counselors is meaningful. Another simple explanation of the differences be- tween more and less successful counselors could be that successful counselors simply pick “easy” issues. However, we find that this is not the case. In par- ticular, we find that both counselor groups are very similar in how they select conversations from the queue (picking the top-most in 60.1% vs. 60.3%, respectively), work similar shifts, and handle a sim- ilar number of conversations simultaneously (1.98 vs. 1.83). Further, we find that both groups face sim- ilar distributions of issues over time (see Table 2). We attribute the largest difference, “NA” (main issue not reported), to the more successful counselors be- ing more diligent in filling out the post-conversation report and having fewer conversations that end be- fore the main issue is introduced. 5 Counselor Adaptability In the remainder of the paper we focus on factors that mediate the outcome of a conversation. First, we examine whether successful counselors are more aware that their current conversation is going well or badly and study how the counselor adapts to the situation. We investigate this question by looking for language differences between positive and neg- ative conversations. In particular, we compute a distance measure between the language counselors use in positive conversations and the language coun- selors use in negative conversations and observe how this distance changes over time. We capture the time dimension by breaking up each conversation into five even chunks of messages. Then, for each set of counselors (more successful or less successful), conversation outcome (positive or negative), and chunk (first 20%, second 20%, etc.), we build a TF-IDF vector of word occurrences to represent the language of counselors within this subset. We use the global inverse document (i.e., conversation) frequencies instead of the ones from each subset to make the vectors directly comparable and control for different counselors having differ- ent numbers of conversations by weighting conver- sations so all counselors have equal contributions. We then measure the difference between the “posi- 0-20 20-40 40-60 60-80 80-100 3ortLon of conversDtLon (% of PessDges) 0.012 0.013 0.014 0.015 0.016 0.017 0.018 0.019 D Ls tD nc e be tw ee n po sL tLv e D nd n eg Dt Lv e co nv er sD tLo ns 0ore successful counselors Less successful counselors Figure 2: More successful counselors are more varied in their language across positive/negative conversations, suggesting they adapt more. All differences between more successful and less successful counselors except for the 0-20 bucket were found to be statistically significant (p < 0.05; bootstrap resampling test). tive” and “negative” vector representations by taking the cosine distance in the induced vector space. We also explored using Jensen-Shannon divergence be- tween traditional probabilistic language models and found these methods gave similar results. Results. We find more successful counselors are more sensitive to whether the conversation is going well or badly and vary their language accordingly (Figure 2). At the beginning of the conversation, the language between positive and negative conver- sations is quite similar, but then the distance in lan- guage increases over time. This increase in distance is much larger for more successful counselors than less successful ones, suggesting they are more aware of when conversations are going poorly and adapt their counseling more in an attempt to remedy the situation. 6 Reacting to Ambiguity Observing that successful counselors are better at adapting to the conversation, we next examine how counselors differ and what factors determine the dif- ferences. In particular, domain experts have sug- gested that more successful counselors are better at handling ambiguity in the conversation (Levitt and Jacques, 2005). Here, we use ambiguity to refer to the uncertainty of the situation and the texter’s ac- tual core issue resulting from insufficiently short or uncertain descriptions. Does initial ambiguity of the situation negatively affect the conversation? How do more successful counselors deal with ambiguous sit- uations? 467 [4, 9] (9, 16] (16, 24] (24, 33] (33, 632] Length of situation setter (#tokens) 0.4 0.5 0.6 0.7 0.8 F ra ct io n o f p o si tiv e c o n v. More successful Less successful Figure 3: More ambiguous situations (length of situation setter) are less likely to result in positive conversations. (4, 8] (8, 15] (15, 23] (23, 32] (32, 476] Length of situation setter (#tokens) 0.0 0.5 1.0 1.5 2.0 2.5 |R e sp o n se | / |S itu a tio n s e tt e r| Counselor quality More successful Less successful Figure 4: All counselors react to short, ambiguous mes- sages by writing more (relative to the texter message) but more successful counselors do it more than less success- ful counselors. Ambiguity. Throughout this section we measure ambiguity in the conversation as the shortness of the texter’s responses in number of words. While ambi- guity could also be measured through concreteness ratings of the words in each message (e.g., using concreteness ratings from Brysbaert et al. (2014)), we find that results are very similar and that length and concreteness are strongly related and hard to distinguish. 6.1 Initial Ambiguity and Situation Setter It is challenging to measure ambiguity and reactions to ambiguity at arbitrary points throughout the con- versation since it strongly depends on the context of the entire conversation (i.e., all earlier messages and questions). However, we can study nearly identi- cal beginnings of conversations where we can di- rectly compare how more successful and less suc- cessful counselors react given nearly identical situa- tions (the texter first sharing their reason for texting in). We identify the situation setter within each con- versation as the first long message by the texter (typ- ically a response to a “Can you tell me more about what is going on?” question by the counselor). Results. We find that ambiguity plays an important role in counseling conversations. Figure 3 shows that more ambiguous situations (shorter length of situation setter) are less likely to result in success- ful conversations (we obtain similar results when measuring concreteness (Brysbaert et al., 2014) di- rectly). Further, we find that counselors generally react to short and ambiguous situation setters by writing significantly more than the texters (Figure 4; if counselors wrote exactly as much as the texter, we would expect a horizontal line y = 1). How- ever, more successful counselors react more strongly to ambiguous situations than less successful coun- selors. 6.2 How to Respond to Ambiguity Having observed that ambiguity plays an important role in counseling conversations, we now examine in greater detail how counselors respond to nearly identical situations. We match situation setters by representing them through TF-IDF vectors on bigrams and find similar situation setters as nearest neighbors within a certain cosine distance in the induced space.1 We only con- sider situation setters that are part of a dense cluster with at least 10 neighbors, allowing us to compare follow-up responses by the counselors (4829/12770 situation setters were part of one of 589 such clus- ters). We also used distributed word embeddings (e.g., (Mikolov et al., 2013)) instead of TF-IDF vec- tors but found the latter to produce better clusters. Based on counselor training materials we hypoth- esize that more successful counselors • address ambiguity by writing more themselves, • use more check questions (statements that tell the conversation partner that you understand 1 Threshold manually set after qualitative analysis of matches from randomly chosen clusters. Results were not overly sensitive to threshold choice, choice of representation (e.g., word vectors), and distance measure (e.g., Euclidean). 468 More S. Less S. Test % conversations successful 70.7 51.7 *** #messages in conversation 57.0 46.7 *** Situation setter length (#tokens) 12.1 10.7 *** C response length (#tokens) 15.8 11.8 *** T response length (#tokens) 20.4 18.8 *** % Cosine sim. C resp. to context 11.9 14.8 *** % Cosine sim. T resp. to context 7.6 7.3 – % C resp. w check question 12.6 4.1 *** % C resp. w suicide check 13.5 10.3 *** % C resp. w thanks 6.3 2.4 *** % C resp. w hedges 41.4 36.8 *** % C resp. w surprise 3.3 2.8 – Table 3: Differences between more and less success- ful counselors (C; More S. and Less S.) in responses to nearly identical situation setters (Sec. 6.1) by the texter (T). Last column contains significance levels of Wilcoxon Signed Rank Tests (*** p < 0.001, – p > 0.05). them while avoiding the introduction of any opinion or advice (Labov and Fanshel, 1977); e.g.“that sounds like...”), • check for suicidal thoughts early (e.g., “want to die”), • thank the texter for showing the courage to talk to them (e.g., “appreciate”), • use more hedges (mitigating words used to lessen the impact of an utterance; e.g., “maybe”, “fairly”), • and that they are less likely to respond with sur- prise (e.g., “oh, this sounds really awful”). A set of regular expressions is used to detect each class of responses (similar to the examples above). Results. We find several statistically significant dif- ferences in how counselors respond to nearly iden- tical situation setters (see Table 3). While situation setters tend to be slightly longer for more success- ful counselors (suggesting that conversations are not perfectly randomly assigned), counselor responses are significantly longer and also spur longer texter responses. Further, the more successful counselors respond in a way that is less similar to the original situation setter (measured by cosine similarity in TF- IDF space) compared to less successful counselors (but the texter’s response does not seem affected). We do find that more successful counselors use more check questions, check for suicide ideation more of- ten, show the texter more appreciation, and use more hedges, but we did not find a significant difference with respect to responding with surprise. More successful Less successful Counselor quality 8 10 12 14 16 18 20 22 24 # S im ila r co u n se lo r re a ct io n s Conversation quality Positive Negative Figure 5: More successful counselors use less com- mon/templated responses (after the texter first explains the situation). This suggests that they respond in a more creative way. There is no significant difference between positive and negative conversations. 6.3 Response Templates and Creativity In Section 6.2, we observed that more successful counselors make use of certain templates (including check questions, checks for suicidal thoughts, affir- mation, and using hedges). While this could suggest that counselors should stick to such predefined tem- plates, we find that, in fact, more successful coun- selors do respond in more creative ways. We define a measure of how “templated” the counselors responses are by counting the number of similar responses in TF-IDF space for the counselor reaction (c.f., Section 6.2; again using a manually defined and validated threshold on cosine distance). Figure 5 shows that more successful counselors use less common/templated questions. This sug- gests that while more successful counselors ques- tions follow certain patterns, they are more creative in their response to each situation. This tailoring of responses requires more effort from the counselor, which is consistent with the results in Figure 1 that showed that more successful counselors put in more effort in composing longer messages as well. 7 Ensuring Conversation Progress After demonstrating content-level differences be- tween counselors, we now explore temporal differ- ences in how counselors progress through conversa- tions. Using an unsupervised conversation model, we are able to discover distinct conversation stages and find differences between counselors in how they 469 move through these stages. We further provide ev- idence that these differences could be related to power and authority by measuring linguistic coor- dination between the counselor and texter. 7.1 Unsupervised Conversation Model Counseling conversations follow a common struc- ture due to the nature of conversation as well as counselor training. Typically, counselors first intro- duce themselves, get to know the texter and their situation, and then engage in constructive prob- lem solving. We employ unsupervised conversation modeling techniques to capture this stage-like struc- ture within conversations. Our conversation model is a message-level Hid- den Markov Model (HMM). Figure 6 illustrates the basic model where hidden states of the HMM rep- resent conversation stages. Unlike in prior work on conversation modeling, we impose a fixed ordering on the stages and only allow transitions from the cur- rent stage to the next one (Figure 7). This causes it to learn a fixed dialogue structure common to all of the counseling sessions as opposed to conversa- tion topics. Furthermore, we separately model coun- selor and texter messages by treating their turns in the conversation as distinct states. We train the con- versation model with expectation maximization, us- ing the forward-backward algorithm to produce the distributions during each expectation step. We ini- tialized the model with each stage producing mes- sages according to a unigram distribution estimated from all messages in the dataset and uniform transi- tion probabilities. The unigram language models are defined over all words occurring more than 20 times (over 98% of words in the dataset), with other words replaced by an unknown token. Results. We explored training the model with vari- ous numbers of stages and found five stages to pro- duce a distinct and easily interpretable representa- tion of a conversation’s progress. Table 4 shows the words most unique to each stage. The first and last stages consist of the basic introductions and wrap- ups common to all conversations. In stage 2, the texter introduces the main issue, while the counselor asks for clarifications and expresses empathy for the situation. In stage 3, the counselor and texter dis- cuss the problem, particularly in relation to the other s1 Ck Ws0 w0,i s0 s2 w1,i w2,i Ws1 Ws2 Figure 6: Our conversation model generates a particular conversation Ck by first generating a sequence of hid- den states s0, s1, ... according to a Markov model. Each state si then generates a message as a bag of words wi,0, wi,1, ... according a unigram language model Wsi . Counselor turn Texter turn Stage 1 c1 Stage 2 Stage k c2 ck t1 t2 tk Figure 7: Allowed state transitions for the conversation model. Counselor and texter messages are produced by distinct states and conversations must progress through the stages in increasing order. people involved. In stage 4, the counselor and tex- ter discuss actionable strategies that could help the texter. This is a well-known part of crisis counselor training called “collaborative problem solving.” 7.2 Analyzing Counselor Progression Do counselors differ in how much time they spend at each stage? In order to explore how counselors progress through the stages, we use the Viterbi al- gorithm to assign each conversation the most likely sequence of stages according to our conversation model. We then compute the average duration in messages of each stage for both more and less suc- cessful counselors. We control for the different distributions of positive and negative conversations among more successful and less successful coun- selors by giving the two classes of conversations equal weight and control for different conversation lengths by only including conversations between 40 and 60 messages long. Results. We find that more successful counselors are quicker to move past the earlier stages, partic- 470 Stage Interpretation Top words for texter Top words for counselor 1 Introductions hi, hello, name, listen, hey hi, name, hello, hey, brings 2 Problem introduction dating, moved, date, liked, ended gosh, terrible, hurtful, painful, ago 3 Problem exploration knows, worry, burden, teacher, group react, cares, considered, supportive, wants 4 Problem solving write, writing, music, reading, play hobbies, writing, activities, distract, music 5 Wrap up goodnight, bye, thank, thanks, appreciate goodnight, 247, anytime, luck, 24 Table 4: The top 5 words for counselors and texters with greatest increase in likelihood of appearing in each stage. The model successfully identifies interpretable stages consistent with counseling guidelines (qualitative interpretation based on stage assignment and model parameters; only words occurring more than five hundred times are shown). 1 2 3 4 5 Stage 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 S ta ge d ur at io n (p er ce nt o f c on ve rs at io n) More successful counselors Less successful counselors Figure 8: More successful counselors are quicker to get to know texter and issue (stage 2) and use more of their time in the “problem solving” phase (stage 4). ularly stage 2, and spend more time in later stages, particularly stage 4 (Figure 8). This suggests they are able to more quickly get to know the texter and then spend more time in the problem solving phase of the conversation, which could be one of the rea- sons they are more successful. 7.3 Coordination and Power Differences One possible explanation for the more successful counselors’ ability to quickly move through the early stages is that they have more “power” in the conversation and can thus exert more control over the progression of the conversation. We explore this idea by analyzing linguistic coordination, which measures how much the conversation partners adapt to each other’s conversational styles. Research has shown that conversation participants who have a greater position of power coordinate less (i.e., they do not adapt their linguistic style to mimic the other conversational participant as strongly) (Danescu- Niculescu-Mizil et al., 2012). In our analysis, we use the “Aggregated 2” coordi- nation measure C(B, A) from Danescu-Niculescu- Mizil (2012), which measures how much group B coordinates to group A (a higher number means more coordination). The measure is computed by counting how often specific markers (e.g., auxiliary verbs) are exhibited in conversations. If someone tends to use a particular marker right after their con- versation partner uses that marker, it suggests they are coordinating to their partner. Formally, let set S be a set of exchanges, each involving an initial utterance u1 by a ∈ A and a reply u2 by b ∈ B. Then the coordination of b to A according to a linguistic marker m is: Cm(b, A) = P(Emu2→u1|E m u1 )−P(Emu2→u1) where Emu1 is the event that utterance u1 exhibits m (i.e., contains a word from category m) and Emu2→u1 is the event that reply u2 to u1 exhibits m. The prob- abilities are estimated across all exchanges in S. To aggregate across different markers, we average the coordination values of Cm(b, A) over all markers m to get a macro-average C(b, A). The coordination between groups B and A is then defined as the mean of the coordinations of all members of group B to- wards the group A. We use eight markers from Danescu-Niculescu- Mizil (2012), which are considered to be processed by humans in a generally non-conscious fashion: ar- ticles, auxiliary verbs, conjunctions, high-frequency adverbs, indefinite pronouns, personal pronouns, prepositions, and quantifiers. Results. Texters coordinate less than coun- selors, with texters having a coordination value of C(texter, counselor)=0.019 compared to the counselor’s C(counselor, texter)=0.030, suggest- ing that the texters hold more “power” in the conversation. However, more successful counselors coordinate less than less successful ones (C(more succ. counselors, texter)=0.029 vs. C(less succ. counselors, texter)=0.032). All differ- ences are statistically significant (p < 0.01; Mann- Whitney U test). This suggests that more successful counselors act with more control over the conversa- 471 tion, which could explain why they are quicker to make it through the initial conversation stages. 8 Facilitating Perspective Change Thus far, we have studied conversation dynamics and their relation to conversation success from the counselor perspective. In this section, we show that perspective change in the texter over time is asso- ciated with a higher likelihood of conversation suc- cess. Prior work has shown that day-to-day changes in writing style are associated with positive health outcomes (Campbell and Pennebaker, 2003), and existing theories link depression to a negative view of the future (Pyszczynski et al., 1987) and a self- focusing style (Pyszczynski and Greenberg, 1987). Here, we propose a novel measure to quantify three orthogonal aspects of perspective change within a single conversation: time, self, and sentiment. Fur- ther, we show that the counselor might be able to actively induce perspective change. Time. Texters start explaining their issue largely in terms of the past and present but over time talk more about the future (see Figure 9A; each plot shows the relative amount of words in the LIWC past, present, and future categories (Tausczik and Pennebaker, 2010)). We find that texters writing more about the future are more likely to feel better after the conversation. This suggests that changing the perspective from issues in the past towards the future is associated with a higher likelihood of suc- cessfully working through the crisis. Self. Another important aspect of behavior change is to what degree the texter is able to change their perspective from talking about themselves to con- sidering others and potentially the effect of their sit- uation on others (Pyszczynski and Greenberg, 1987; Campbell and Pennebaker, 2003). We measure how much the texter is focused on themselves by the rela- tive amount of first person singular pronouns (I, me, mine) versus third person singular/plural pronouns (she, her, him / they, their), again using LIWC. Fig- ure 9B shows that a smaller amount of self-focus is associated with more successful conversations (pro- viding support for the self-focus model of depres- sion (Pyszczynski and Greenberg, 1987)). We hy- pothesize that the lack of difference at the end of the conversation is due to conversation norms such as thanking the counselor (“I really appreciate it.”) even if the texter does not actually feel better. Sentiment. Lastly, we investigate how much a change in sentiment of the texter throughout the con- versation is associated with conversation success. We measure sentiment as the relative fraction of pos- itive words using the LIWC PosEmo and NegEmo sentiment lexicons. The results in Figure 9C show that texters always start out more negative (value be- low 0.5), but that the sentiment becomes more posi- tive over time for both positive and negative conver- sations. However, we find that the separation be- tween both groups grows larger over time, which suggests that a positive perspective change through- out the conversation is related to higher likelihood of conversation success. We find that both curves increase significantly at the very end of the con- versation. Again, we attribute this to conversation norms such as thanking the counselor for listening even when the texter does not actually feel better. Together with the result on talking about the fu- ture, these findings are consistent with the theory of Pyszczynski et al. (1987) that depression is related to a negative view of the future. Role of a Counselor. Given that positive conver- sations often exhibit perspective change, a natural question is how counselors can encourage perspec- tive change in the texter. We investigate this by ex- ploring the hypothesis that the texter will tend to talk more about something (e.g., the future), if the coun- selor first talks about it. We measure this tendency using the same coordination measures as Section 7.3 except that instead of using stylistic LIWC markers (e.g., auxiliary verbs, quantifiers), we use the LIWC markers relevant to the particular aspect of perspec- tive change (e.g., Future, HeShe, PosEmo). In all cases we find a statistically significant (p < 0.01; Mann-Whitney U-test) increase in the likelihood of the texter using a LIWC marker if the counselor used it in the previous message (~4-5% change). This link between perspective change and how the counselor conducts the conversation suggests that the coun- selor might be able to actively induce measurable perspective change in the texter. 472 0-20 20-40 40-60 60-80 80-100 3ortion of Fonversation (% of Pessages) 0.06 0.07 0.08 0.09 0.10 0.11 0.12 0.13 0.14 0.15 ) ut ur e re la tiv e fr eq ue nF y 3ositive Fonversations 1egative Fonversations 0-20 20-40 40-60 60-80 80-100 3ortion of conversation (% of Pessages) 0.12 0.14 0.16 0.18 0.20 0.22 0.24 0.26 0.28 0.30 3 as t r el at iv e fr eq ue nc y 3ositive conversations 1egative conversations 0-20 20-40 40-60 60-80 80-100 Portion of conversation (% of Pessages) 0.60 0.62 0.64 0.66 0.68 0.70 0.72 0.74 0.76 P re se nt r el at iv e fr eq ue nc y Positive conversations 1egative conversations 0-20 20-40 40-60 60-80 80-100 Portion of conversation (% of Pessages) 0.70 0.75 0.80 0.85 6 el f r el at iv e fr eq ue nc y Positive conversations 1egative conversations 0-20 20-40 40-60 60-80 80-100 Portion of conversation (% of Pessages) 0.4 0.5 0.6 0.7 0.8 P os eP o re la tiv e fr eq ue nc y Positive conversations 1egative conversations A1: Past B: Self C: Sentiment A2: Present A3: Future Figure 9: A: Throughout the conversation there is a shift from talking about the past to future, where in positive conversations this shift is greater; B: Texters that talk more about others more often feel better after the conversation; C: More positive sentiment by the texter throughout the conversation is associated with successful conversations. 9 Predicting Counseling Success In this section, we combine our quantitative insights into a prediction task. We show that the linguistic as- pects of crisis counseling explored in previous sec- tions have predictive power at the level of individ- ual conversations by evaluating their effectiveness as features in classifying the outcome of conversations. Specifically, we create a balanced dataset of positive and negative conversations more than 30 messages long and train a logistic regression model to predict the outcome given the first x% of messages in the conversation. There are 3619 such negative conver- sations and and we randomly subsample the larger set of positive conversations. We train the model with batch gradient descent and use L1 regulariza- tion when n-gram features are present and L2 reg- ularization otherwise. We evaluate our model with 10-fold cross-validation and compare models using the area under the ROC curve (AUC). Features. We include three aspects of counselor messages discussed in Section 6: hedges, check questions, and the similarity between the counselor’s message and previous texter message. We add a measure of how much progress the counselor has made (Section 7) by computing the Viterbi path of stages for the conversation (only for the first x%) with the HMM conversation model and then adding the duration of each stage (in #messages) as a fea- ture. Additionally, we add average message length 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Percent of conversation seen by the model 0.56 0.58 0.60 0.62 0.64 0.66 0.68 0.70 0.72 R O C a re a un de r th e cu rv e sc or e Full model N-gram features only Figure 10: Prediction accuracies vs. percent of the con- versation seen by the model (without texter features). and average sentiment per message using VADER sentiment (Hutto and Gilbert, 2014). Further, we add temporal dynamics to the model by adding fea- ture conjunctions with the stages HMM model. Af- ter running the stages model over the x% of the con- versation available to the classifier, we add each fea- ture’s average value over each stage as additional features. Lastly, we explore the benefits of adding surface-level text features to the model by adding unigram and bigram features. Because the focus of this work is on counseling strategies, we primar- ily experiment with models using only features from counselor messages. For completeness, we also re- port results for a model including texter features. Prediction Results. The model’s accuracy increases with x, and we show that the model is able to dis- 473 Features ROC AUC Counselor unigrams only 0.630 Counselor unigrams and bigrams only 0.638 None 0.5 + hedges 0.514 (+0.014) + check questions 0.546 (+0.032) + similarity to last message 0.553 (+0.007) + duration of each stage 0.561 (+0.008) + sentiment 0.590 (+0.029) + message length 0.596 (+0.006) + stages feature conjunction 0.606 (+0.010) + counselor unigrams and bigrams 0.652 (+0.046) + texter unigrams and bigrams 0.708 (+0.056) Table 5: Performance of nested models predicting con- versation outcome given the first 80% of the conversa- tion. In bold: full models with only counselor features and with additional texter features. tinguish positive and negative conversations after only seeing the first 20% of the conversation (see Figure 10). We attribute the significant increase in performance for x = 100 (Accuracy=0.687, AUC=0.716) to strong linguistic cues that appear as a conversation wraps up (e.g., “I’m glad you feel better.”). To avoid this issue, our detailed feature analysis is performed at x = 80. Feature Analysis. The model performance as fea- tures are incrementally added to the model is shown in Table 5. All features improve model accuracy sig- nificantly (p < 0.001; paired bootstrap resampling test). Adding n-gram features produces the largest boost in AUC and significantly improves over a model just using n-gram features (0.638 vs. 0.652 AUC). Note that most features in the full model are based on word frequency counts that can be derived from n-grams which explains why a simple n-gram model already performs quite well. However, our model performs well with only a small set of lin- guistic features, demonstrating they provide a sub- stantial amount of the predictive power. The effec- tiveness of these features shows that, in addition to exhibiting group-level differences reported earlier in this paper, they provide useful signal for predicting the outcome of individual conversations. 10 Conclusion & Future Work Knowledge about how to conduct a successful coun- seling conversation has been limited by the fact that studies have remained largely qualitative and small- scale. In this work, we presented a large-scale quan- titative study on the discourse of counseling con- versations. We developed a set of novel computa- tional discourse analysis methods suited for large- scale datasets and used them to discover actionable conversation strategies that are associated with bet- ter conversation outcomes. We hope that this work will inspire future generations of tools available to people in crisis as well as their counselors. For ex- ample, our insights could help improve counselor training and give rise to real-time counseling qual- ity monitoring and answer suggestion support tools. Acknowledgements. We thank Bob Filbin for fa- cilitating the research, Cristian Danescu-Niculescu- Mizil for many helpful discussions, and Dan Ju- rafsky, Chris Manning, Justin Cheng, Peter Clark, David Hallac, Caroline Suen, Yilun Wang and the anonymous reviewers for their valuable feedback on the manuscript. This research has been sup- ported in part by NSF CNS-1010921, IIS-1149837, NIH BD2K, ARO MURI, DARPA XDATA, DARPA SIMPLEX, Stanford Data Science Initiative, Boe- ing, Lightspeed, SAP, and Volkswagen. References Tim Althoff and Jure Leskovec. 2015. Donor retention in online crowdfunding communities: A case study of DonorsChoose.org. In WWW. Tim Althoff, Cristian Danescu-Niculescu-Mizil, and Dan Jurafsky. 2014. How to ask for a favor: A case study on the success of altruistic requests. In ICWSM. Vikas Ganjigunte Ashok, Song Feng, and Yejin Choi. 2013. Success with style: Using writing style to pre- dict the success of novels. In EMNLP. Regina Barzilay and Lillian Lee. 2004. Catching the drift: Probabilistic content models, with applications to generation and summarization. In HLT-NAACL. Aaron T. Beck. 1967. Depression: Clinical, experimen- tal, and theoretical aspects. University of Pennsylva- nia Press. Philip Bramsen, Martha Escobar-Molano, Ami Patel, and Rafael Alonso. 2011. Extracting social power rela- tionships from natural language. In HLT-NAACL. Marc Brysbaert, Amy Beth Warriner, and Victor Kuper- man. 2014. Concreteness ratings for 40 thousand generally known English word lemmas. Behavior Re- search Methods, 46(3). R. Sherlock Campbell and James W. Pennebaker. 2003. The secret life of pronouns: Flexibility in writing style and physical health. Psychological Science, 14(1). 474 Ge Chen. 2014. Visualizations for mental health topic models. Master’s thesis, MIT. Cristian Danescu-Niculescu-Mizil, Lillian Lee, Bo Pang, and Jon Kleinberg. 2012. Echoes of power: Language effects and power differences in social interaction. In WWW. Cristian Danescu-Niculescu-Mizil. 2012. A computa- tional approach to linguistic style coordination. Ph.D. thesis, Cornell University. Karthik Dinakar, Allison J.B. Chaney, Henry Lieberman, and David M. Blei. 2014a. Real-time topic models for crisis counseling. In KDD DSSG Workshop. Karthik Dinakar, Emily Weinstein, Henry Lieberman, and Robert Selman. 2014b. Stacked generalization learning to analyze teenage distress. In ICWSM. Karthik Dinakar, Jackie Chen, Henry Lieberman, Ros- alind Picard, and Robert Filbin. 2015. Mixed- initiative real-time topic modeling & visualization for crisis counseling. In ACM ICIUI. Marco Guerini, Alberto Pepe, and Bruno Lepri. 2012. Do linguistic style and readability of scientific ab- stracts affect their virality? In ICWSM. Shane Haberstroh, Thelma Duffey, Marcheta Evans, Robert Gee, and Heather Trepal. 2007. The experi- ence of online counseling. Journal of Mental Health Counseling, 29(3). Christine Howes, Matthew Purver, and Rose McCabe. 2014. Linguistic indicators of severity and progress in online text-based therapy for depression. CLPsych Workshop at ACL 2014. Rongyao Huang. 2015. Language use in teenage crisis intervention and the immediate outcome: A machine automated analysis of large scale text data. Master’s thesis, Columbia University. C.J. Hutto and Eric Gilbert. 2014. Vader: A parsimo- nious rule-based model for sentiment analysis of social media text. In ICWSM. Thomas R. Insel. 2008. Assessing the economic costs of serious mental illness. The American Journal of Psychiatry, 165(6). John Kalafat, Madelyn Gould, Jimmie Lou Harris Mun- fakh, and Marjorie Kleinman. 2007. An evaluation of crisis hotline outcomes part 1: Nonsuicidal crisis callers. Suicide and Life-threatening Behavior, 37(3). William Labov and David Fanshel. 1977. Therapeutic discourse: Psychotherapy as conversation. Dana Heller Levitt and Jodi D. Jacques. 2005. Promot- ing tolerance for ambiguity in counselor training pro- grams. The Journal of Humanistic Counseling, Edu- cation and Development, 44(1). Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Cor- rado, and Jeff Dean. 2013. Distributed representa- tions of words and phrases and their compositionality. In NIPS. National Institute of Mental Health. 2015. Any mental illness (AMI) among U.S. adults. http://www.nimh.nih.gov/health/ statistics/prevalence/any-mental- illness-ami-among-us-adults.shtml. Retrieved June 3, 2016. Kate G. Niederhoffer and James W. Pennebaker. 2002. Linguistic style matching in social interaction. Jour- nal of Language and Social Psychology, 21(4). Michael J. Paul. 2012. Mixed membership Markov models for unsupervised conversation modeling. In EMNLP-CoNLL. James W. Pennebaker, Matthias R. Mehl, and Kate G. Niederhoffer. 2003. Psychological aspects of natural language use: Our words, our selves. Annual Review of Psychology, 54(1). John P. Pestian, Pawel Matykiewicz, Michelle Linn-Gust, Brett South, Ozlem Uzuner, Jan Wiebe, Kevin B. Co- hen, John Hurdle, and Christopher Brew. 2012. Senti- ment analysis of suicide notes: A shared task. Biomed- ical informatics insights, 5(Suppl. 1). Tom Pyszczynski and Jeff Greenberg. 1987. Self- regulatory perseveration and the depressive self- focusing style: A self-awareness theory of reactive de- pression. Psychological Bulletin, 102(1). Tom Pyszczynski, Kathleen Holt, and Jeff Greenberg. 1987. Depression, self-focused attention, and ex- pectancies for positive and negative future life events for self and others. Journal of Personality and Social Psychology, 52(5). Nairan Ramirez-Esparza, Cindy Chung, Ewa Kacewicz, and James W. Pennebaker. 2008. The psychology of word use in depression forums in English and in Span- ish: Testing two text analytic approaches. In ICWSM. Shiquan Ren, Hong Lai, Wenjing Tong, Mostafa Amin- zadeh, Xuezhang Hou, and Shenghan Lai. 2010. Non- parametric bootstrapping for hierarchical data. Jour- nal of Applied Statistics, 37(9). Alan Ritter, Colin Cherry, and Bill Dolan. 2010. Unsu- pervised modeling of Twitter conversations. In HLT- NAACL. Harvey Sacks and Gail Jefferson. 1995. Lectures on con- versation. Wiley-Blackwell. Andreas Stolcke, Klaus Ries, Noah Coccaro, Eliza- beth Shriberg, Rebecca Bates, Daniel Jurafsky, Paul Taylor, Rachel Martin, Carol Van Ess-Dykema, and Marie Meteer. 2000. Dialogue act modeling for automatic tagging and recognition of conversational speech. Computational Linguistics, 26(3). Yla R. Tausczik and James W. Pennebaker. 2010. The psychological meaning of words: LIWC and comput- erized text analysis methods. Journal of Language and Social Psychology, 29(1). 475 http://www.nimh.nih.gov/health/statistics/prevalence/any-mental-illness-ami-among-us-adults.shtml http://www.nimh.nih.gov/health/statistics/prevalence/any-mental-illness-ami-among-us-adults.shtml http://www.nimh.nih.gov/health/statistics/prevalence/any-mental-illness-ami-among-us-adults.shtml Teun Van Dijk. 1997. Discourse studies: A multidisci- plinary approach. SAGE. World Health Organization. 2015. Depression: Fact sheet no 369. http://www.who.int/ mediacentre/factsheets/fs369/en/. Re- trieved November 2, 2015. Jaewon Yang, Julian McAuley, Jure Leskovec, Paea LePendu, and Nigam Shah. 2014. Finding pro- gression stages in time-evolving event sequences. In WWW. 476 http://www.who.int/mediacentre/factsheets/fs369/en/ http://www.who.int/mediacentre/factsheets/fs369/en/