DOES LINGUISTIC ALIGNMENT FACILITATE SUPPORT FUNCTIONS: A BIG DATA ANALYSIS ON ONLINE HEALTH COMMUNITIES

Abstract

Nowadays, an increasing number of people with serious diseases can seek and provide social support. Analyzing such on group discussions boosts our understanding of the impact of OHCs as well as the different characteristics of support-oriented interactions. This thesis describes four computational studies that analyze the relationship between linguistic alignment, a universal communication phenomenon, and social support.

Firstly, we introduce a communication phenomenon, linguistic adaptation. Linguistic adaptation in web-based communication means people tend to adjust their language use to one another both in terms of word choice and sentence structure. So far, our understanding of the relationship between linguistic alignment in social support-oriented conversations and its possible connection to member benefits is limited. We quantify linguistic alignment in an OHC at two language levels: word choices and syntactic rules. Our finding results show alignment at both lexical and syntactic level, while the speed of the decay on linguistic adaptation is much slower than previous corpus studies. These different patterns not only can be potentially revealed through alignment theories, but also help researchers understand the unique function of OHCs, i.e. exchanging social support.

As support seekers play a key role in support-oriented conversations, we focus on how community members provide support to support seekers. We construct a computational model analyzing linguistic alignment between support seekers and support providers in OHCs. Surprisingly, our finding results show that lexical alignment and syntactic alignment have distinct correlations to the support functions. This result makes potential theoretical progress, revealing the relationship between different levels of linguistic adaptation. It also motivates further research regarding potential refinement of computational linguistic theories, such as the Interactive Alignment Model, in other web-based conversations.

We further analyze how support providers adapted to each other in such interactions. Specially, we study whether latter social support providers are influenced by early support providers. From the theoretical perspective, as adaptation can occur at lexical, syntactic and pragmatic levels, relation between alignment across multiple levels is neither theoretically nor empirically understood. Thus, in this study, we develop a computational model predicting the support choice given the message from early responders using linguistic alignment measures. We find that community members align on social support type. Also, lexical adaptation, not syntactic adaptation, reliably indicates emotional support. These findings can help us understand linguistic signature of different types of social support, as well as the interactive alignment theory.

Finally, we introduce linguistic adaptation phenomenon into a causal relationship inference framework, probabilistic Kripke structure, quantifying the effectiveness of emotional support. We further analyze the framework by examining prima facie causes, and their significance. The result presents that replies with positive sentiment and high alignment score are consistently prima facie causes of resulting positive sentiment of the thread initiator, which means that linguistic adaptation is a temporal causal factor for high level communication.

Our research makes a potential theoretical progress of the mechanism of linguistic adaptation, especially in web-based conversations. From the applied perspective, this thesis also contributes to a better understanding of the impact of supportoriented interactions for benefiting members in OHCs.

 

Chapter 1       |

Introduction

With rapid expansion of Internet access, an increasing number of individuals can get online to achieve certain health purposes, such as searching for general/serious health information, sign-up for email alerts about health, exchanging social support with peers, and so on. A Pew research technical report in 2011 (Fox, 2011b) shows that 80% of U.S. Internet users have gone online for health information. In addition to searching general health related information, people also get online to exchange help to each other, called peer-to-peer online healthcare (Fox, 2011a). Almost one in three (34%) Internet users have read others’ personal experience on health, and about one in five (18%) have found others who share similar health information (Fox, 2011b). Analyzing and improving such “peer-to-peer” online healthcare is beneficial not only for people with health-related concerns but also for healthcare professionals.

A primary way of “peer-to-peer” healthcare interaction is via social network websites, such as Facebook groups, WebMed forums, Cancer Survivor Networks, and etc. Previous research studies (Eysenbach, Powell, Englesakis, Rizo, & Stern,

Figure 1.1: A fabricated and anonymized example of support interactions. It is adapted according to a real interactions in Cancer Survivor Networks. The illustration of four studies in this thesis.

2004) suggested that patients with health-related problems would like to participate in such communications in a supportive way, in a social ne twork of people with similar problems. Specifically, patients and caregivers tend to share their personal experiences, get social support and ask questions in online support-oriented groups. Such supportive online messages not only help the person who asks questions, but also help other group members in the future (Fox, 2011b). Previous research showed that participants of online support-oriented groups report increased social support, reduced levels of depression and psychological stress, increased optimism, increased ability to cope with patients’ health conditions, and improved patients’ quality of life (Beaudoin & Tao, 2008; Bouma et al., 2015; Dunkel-Schetter, 1984; Maloney-Krichmar & Preece, 2005; Rodgers & Chen, 2005). A fabricated example of such supportive interactions in online health communities is shown in Figure 1.1.

Different from traditional face-to-face support-oriented interactions, most of the web-based support-oriented interactions in Online Health Communities (OHCs) are text based interactions which happened online. Hence, the characteristics of such interactions in social support groups are different:

• Text-based. Conversations in OHCs are recorded as text, which means that other features, such as speech rate, gestures, and so on, in general communication are missing in this setting.
• Support-oriented. The purpose of a support-oriented conversation is exchanging social support. For example, a conversational thread in OHCs often starts with seeking social support from peers. The conversation participants provide social support, such as general information, personal experiences, caring, and so on, to the person who starts the conversation.
• Multi-party. Most conversations in support-oriented groups have more than two conversation participants. Normally, there is one person who starts a conversation. Then, more than one person participates in the conversation via replying to the person who posts the concern.
• Conversations in support-oriented groups often are not realtime dialogues. Under this setting, conversational participants may not be online, and engage the conversation at the same time. People always read the previous messages in the past.

• No specific task. Different from some experimental studies of task-oriented conversations, conversations in support-oriented groups are ordinary dialogues. They occur naturally without a clear objective, such as solving a problem or describing a picture.
• Have a specific topic. Conversations in support-oriented groups have a specific topic, which is described in the beginning. The rest of conversation participants are providing relevant social support within the topic.

In this thesis, we will focus on web-based support-oriented interactions in an online health community, Cancer Survivor Networks ^[1] . We will see how these

unique characteristics facilitate such support-oriented interactions.

1.1 Research Questions

While the impact of online support-oriented groups is well studied, a limited number of studies have considered communication accommodation, which was one important communication phenomenon that happened in communications, impacting support-oriented interactions. In addition, there are few studies have analyzed the relationship between social support function, the key function of online supportoriented groups, and linguistic adaptation, a text-based form of communication accommodation. To this end, this thesis describes four computational approaches that analyze users’ web-based supportive interactions in a special social media setting: online health communities.

1.2 Overview

This dissertation includes four studies, analyzing and modeling support-oriented dialogues using linguistic adaptation phenomena. These four studies model linguistic adaptation phenomenon across different communication representation levels with different types of relationships. The illustration of these four studies is shown in Figure 1.1. Firstly, We show the existence of linguistic adaptation as well as emphasize different characteristics of linguistic alignment in OHCs (Study 1). Then, we show the relationship between lower-level linguistic alignment, like lexical and syntactic alignment, and higher-level communicative representations, like social support functions in OHCs (Study 2, 3). Finally, we show the temporal causal relationship between linguistic alignment and the sentiment change of thread initiator in a conversation (Study 4).

1.2.1 Study 1: Linguistic Adaptation in Support Oriented Forums

Alignment phenomenon is a universal phenomenon which happens in most conversations. Previous studies of alignment have focused on two-party conversations under experimental settings. However, limited number of studies have explored multi-party conversations, especially support-oriented interactions. In this study, we conducted four experiments analyzing whether linguistic adaptation phenomenon of support-oriented interpersonal interactions in an online health community, Cancer Survivor Networks.

The result findings showed that support-oriented interactions have different characteristics than other experimental conversations or corpus studies. Specifically, in thread-based conversations, the decay of lexical alignment to an initial post and posts from the initial author was weaker than that to any post in the thread. Further, although, syntactic alignment decayed for all posts in a thread, there was less syntactic alignment to the initial post, and there was even anti-alignment in the beginning of the thread. This finding result may be explained by the function of social support and the properties of online health communities. The result findings also showed the existence of linguistic adaptation phenomenon in support-oriented interactions.

1.2.2 Study 2: Lexical Alignment is Associated with Emotional Support

In the first study, we showed that support-oriented interpersonal interactions in online health communities presented different properties on linguistic adaptation phenomena from general interactions in psycholinguistic experimental settings and interactions. Especially, the decay of linguistic alignment to initial post, which typically seeking social support in this corpus, is weaker than general interactions. Although we argued that this may be explained by the properties of supportoriented interactions, we didn’t show how support providers adapt to support seekers influenced by different types of social support.

In this study, we analyzed the relationship between social support providers and social support seekers with regard to the choice of social support. The modeling results showed that support providers provide social support by adapting to each other at the lexical level. Also, adaptation differed with the type of social support,

i.e. emotional support has greater lexical and less syntactic adaptation, but informational support shows opposite properties. The result findings may suggest potential refinement of Interactive Alignment Theory (Pickering & Garrod, 2004a) that lexical and syntactic alignment are not positively correlated in the setting of social support interactions.

1.2.3             Study 3: Social Support Type is Aligned in a Thread

In the previous study, we showed that support providers showed higher lexical alignment to support seekers while providing emotional support. In this study, we focused on one actor, support providers, in support-oriented interactions. Specifically, we further explored how support providers themselves adapt to each other, not only at lexical/syntactic alignment levels, but also at the level of social support choice.

As social support is important in support-oriented interactions, we examined whether the type of social support provided by later responders was influenced by the type of social support provided by earlier responders. The finding results showed that there was a positive relationship between these two. Theoretically, we considered these types of alignment as pragmatic alignment. Further, we also modeled how linguistic alignment, i.e. at both lexical and syntactic levels, influences the support choice. The result showed alignment at a social support choice level tend to coincide with lexical alignment. This result finding partially supports the interactive alignment model, which suggested that linguistic alignment builds up higher level communication representation.

1.2.4 Study 4: Is Linguistic Alignment a Temporal Causal Factor of Sentiment Change?

In previous three studies, we showed linguistic adaptation phenomena, and how social support choice was influenced by linguistic alignment in support-oriented interactions in online health communities. However, as all these studies are observational studies, none of these studies suggested any causal account. In this study, we utilized a novel temporal causality framework exploring how linguistic alignment and the sentiment of replies facilitated the sentiment change of thread initiator, which is an indicator of the effectiveness of social support.

Our analysis showed that high lexical alignment between the reply posts and their thread initiating post, together with the positive sentiment of reply posts, were the prima facie causes of a positive change in the sentiment of the thread originator. Our results also provided additional insights about temporal causal factors for sentiment change in support-oriented interactions. Furthermore, it provides an important basis for further studies and experiments regarding causal factors for other types of benefits of online health communities.

1.3 Outlines

This thesis is structured as following:

Chapter 2 provides the necessary background information and related work of each individual problem which we will focus on in this thesis. Chapter 3 starts with proposing the linguistic alignment measure, at both lexical and syntactic levels, which we will use in this thesis. Specifically, we examine the existence of linguistic adaptation in longitudinal conversations in online forums. Chapter 5 introduces pragmatic alignment, e.g. social support types, in longitudinal conversations. The correlation between pragmatic alignment and linguistic alignment is also estimated. Chapter 4 examines whether two types of linguistic alignment related to the pragmatic communication (emotional support and informational support) are the same or different. We further discuss the interactive alignment model (Pickering & Garrod, 2004a) as well. In addition to correlation analysis, Chapter 6 shows a probabilistic Kripke structure examining whether linguistic alignment is a temporal causal factor of sentiment change.

[1] https://csn.cancer.org/forum

DOES LINGUISTIC ALIGNMENT FACILITATE SUPPORT FUNCTIONS: A BIG DATA ANALYSIS ON ONLINE HEALTH COMMUNITIES

Sharing is caring!