Discourse analysis of “just chatting” streams on YouTube live: focusing on the interaction between virtual YouTubers and viewers

4.1 To what extent do streamers refer to viewers’ messages?

First, we analyzed the extent to which the streamers in our dataset of “just chatting” streams referred to messages sent by viewers. Previous research includes analyses suggesting that streamers may be expected to read the messages that their viewers send them (Licoppe and Morel 2018) as well as analyses that indicate this is not invariably the case (Recktenwald 2018). In “just chatting” streams, in which interaction between the streamer and viewers is the primary activity, it is expected that large numbers of utterances and messages that are expressly aimed at facilitating interaction will be exchanged. However, an imbalance between the number of such messages and the streamer’s ability to respond to them is also anticipated (Recktenwald 2018). In fact, in Stream A, which had the most messages among the data we observed, 2,769 unique viewers sent messages within the first 60 min of the video. In light of this, we counted the number of messages sent during the streams that the streamers referred to in a manner that allowed us to identify the specific content of the message. Table 2 presents the results of this analysis (“Number of Messages” and “Percentage of total”), along with the number of unique viewers who sent messages during the first 60 min of the video.

As Table 2 demonstrates, excluding Super Chats involving monetary contributions, the proportion of messages explicitly referred to by the streamer ranges from 8.7 % (Stream F) to 0.4 % (Stream C), indicating that over 90 % of messages were not specifically addressed by the streamer. Furthermore, 0.4 % of messages were referred to in Stream A, which had the highest total number of messages, while 7.6 % were referred to in Stream D, which had the lowest total number of messages. This indicates that the imbalance between the number of messages and the streamer’s capacity to respond influences the interaction, as found by Recktenwald (2018). However, Stream C, in which the proportion of referred messages was as low as that in Stream A, received only approximately one-third of the total number of messages sent to Stream A. Stream F, which had the highest proportion, received roughly five times the total messages sent to Stream D. This suggests that the frequency with which streamers refer to messages may also vary depending on the streamer’s personality, their relationship with viewers, and the content discussed.

Similarly, as Recktenwald (2018) observed on Twitch, when the streamer referred to messages from viewers, it was often in response to messages requiring an answer. However, there were also instances in our data that did not fit this characteristic, such as in example (1), in which the streamer picked up comments that did not require a response. In example (1), the streamer was eating a McDonald’s hamburger during the stream, and the message from Viewer 2 in line 14 was a simple, unprompted comment (“Now I want to eat McDonald’s fries”). However, the streamer read this message aloud in line 20 and then advised the viewer in line 29 to hold off on buying fries due to an upcoming price drop sale. Thus, while the streamer only refers to a fraction of viewer messages, these references do not invariably conform to a Q&A format. This is likely to be related to the inherently interactive nature of “just chatting” streams, which are aimed at bidirectional communication.

Previous quantitative studies focused on Twitch have reported that the length and diversity of individual messages, the temporal spacing between messages, and the frequency with which conversations take place among viewers change depending on the size of the audience (Nematzadeh et al. 2019; Olejniczak 2015). Given that our focus was on the streamer’s language use, detailed analysis of viewer messages is beyond this paper’s scope, though little conversation took place among the viewers in all the streams.^[3]

(1)

4.2 What is the specific length of the time lag that occurs between when a message is sent and when the streamer refers to it?

Next, we examine the specific time lag that occurs when streamers refer to messages from viewers. A common characteristic of CMC is that, due to the nature of communication via the internet and electronic devices, it takes longer for others to recognize an utterance or message after it has been sent, as compared to face-to-face conversations (Schoenenberg et al. 2014). This issue also naturally applies to nearly synchronous video-mediated and video-text communication in Japanese-speaking environments. For example, Hosoma and Muraoka (2022) have noted that even slight transmission delays in Zoom can affect the synchronization of actions and turn-taking among participants. Moreover, in live streaming, in addition to the system-induced time lags, delays may occur due to the viewers’ use of text. Similarly, delays often occur on the streamer’s side, as they often need to engage in multiple activities simultaneously, such as interacting with viewers, sharing personal stories, and playing games, depending on the progression of the discourse (Recktenwald 2018).

In YouTube Live, part of the system-induced lag (server-related delay) can be controlled, and streamers must select “normal latency”, “low latency”, or “ultra-low latency” when commencing a stream (YouTube n.d.b). All streams we examined are likely to have been set to “ultra-low latency”, which is estimated to result in an approximate 3-second server processing delay.^[4] Meanwhile, based on the analysis in Section 4.1, we extracted the timestamps of each message explicitly referred to by the streamer only when the streamer clearly referred to a single message (i.e., excluding cases where multiple identical or highly similar messages were sent within a short period). We then compared these message timestamps with the start times of the streamer’s corresponding utterances, as extracted from auto-generated subtitles. Based on these comparisons, we calculated the average time span between the sending of a message and the streamer’s response for each “just chatting” stream (Table 3).

As Table 3 suggests, it takes more than twice as long for a streamer in a “just chatting” stream to refer to a viewer’s message after it has been sent, as compared to the system-induced lag in YouTube Live. However, the length of this lag varies across the data. Comparison of Streams A, B, and I, each of which had over 10,000 messages in the first 60 min of the video, with Streams D and G, which each had fewer than 1,000 messages, reveals that more messages result in quicker references by the streamer, while fewer messages result in a longer time before the streamer refers to them.

In example (2) from Stream B, for example, the streamer discusses the difficulty of building relationships with heroines in a dating simulation game. Meanwhile, viewers continuously send messages expressing empathy or responding to the streamer’s words. The streamer refers to individual messages within 7 s of their sending time, and these references are observed to be concise – all within 5 s, except for the utterance in line 1. Streams A, B, and I, which have particularly large viewer numbers, all exhibit these characteristics, with frequent and concise references repeated throughout the video.

(2)

By contrast, in example (3) from Stream D, the streamer reflects on a lengthy Let’s Play stream from several days previously, and individual viewers mention how much of the stream they watched. The interval between messages is always more than 1 s, indicating that the streamer has relatively ample time to check each message. However, 32 s elapse between the time the message in line 1 is sent to the time it is referred to in line 13, indicating that the references are not prompt. Additionally, the utterances following line 13 are not limited to concise responses, given that the streamer uses the conjunction “Demo” (but) to extend the utterance into a reflective speech that continues for 22 s until line 16.

(3)

As this contrast reveals, the streamer adjusts the amount of time dedicated to a single topic or individual viewer based on the flow of messages.

4.3 How do streamers linguistically cope with the constraints of live streaming?

Finally, we analyze the linguistic methods that streamers use to interact smoothly with viewers while live streaming. As noted earlier, it is difficult for streamers to refer to all viewer messages and, even if they can, a certain time lag is unavoidable. Therefore, to overcome these constraints, our data identified three methods that streamers appear to use: “reading messages aloud”, “reading usernames aloud”, and “using adjacency pairs”.^[5]

Of these methods, “reading messages aloud” is similar – or identical – to the Topicalizer technique that Recktenwald (2018) reported based on observations of English-language live streams and the RAR format reported by Licoppe and Morel (2018) based on observations of French live streaming activities. For example, in lines 1 and 6 of (4), the streamer asks about the duration of the Obon holiday, a unique Japanese holiday. Meanwhile, during the streamer’s utterance, viewers sent their responses to the question. The streamer subsequently read aloud one of these responses from Viewer 2 (line 3) and declared at line 10 that the question was resolved.

(4)

Similarly, “reading usernames aloud” entails the streamer reading the usernames of those who sent messages (strictly speaking, their YouTube channel names). In (5), the streamer explains the day’s content in lines 11 and 14 and the preceding utterances. Meanwhile, viewers send messages reporting that it is their first time watching the live stream (lines 01, 02, 03, and 13) or simply greeting the streamer (line 12). Based on these messages, the streamer reads the usernames aloud in lines 17 and 18 and expresses gratitude.^[6]

(5)

Furthermore, “using adjacency pairs” involves the streamer making utterances that are designed as the second part of adjacency pairs (Schegloff and Sacks 1973), such as an “answer” to a “question”. In (6), the streamer gives her impressions of the latest song she has covered, comparing it to a previously covered song, “Tokumei M”, in lines 02 and 05. After this segment, the streamer responds in line 13 to a viewer’s question in line 01 (“So, did you also do a lot of takes for the song Tokumei M?”). The streamer’s utterance in line 13 begins with the negative response “Iya” (“No”), indicating that it is the second part of an adjacency pair, with the first part found in the preceding message.

(6)

It should be noted that these three methods are not mutually exclusive within a single utterance. For example, the utterance in line 17 of (5) (“Um, Viewer 2,” “I’m a first-time viewer.” “Good evening and thank you.”) has characteristics of both “reading usernames aloud” and “reading messages aloud”. Similarly, just as Recktenwald (2018) presents the Topicalizer as a method for referencing messages that “request a response”, utterances that fulfill the category of “reading messages aloud” are also often the second part of an adjacency pair.^[7]

However, “Reading messages aloud” is clearly the most frequently used of the three methods. Even counting only those instances in which not only a single noun or noun phrase but also entire clauses or sentences were read aloud, it was observed that in all but one stream, more than half of the references fell into the “reading messages aloud” category, as Table 4 demonstrates.

On the other hand, instances in which “reading usernames aloud” or “using adjacency pairs” occurred independently, as seen in (5) and (6), were relatively rare, with fewer than 10 examples in most streams.^[8] Therefore, “reading messages aloud” may also be acknowledged as a commonly used technique in live streaming in the Japanese environment.

Moreover, from a grammar research perspective, an interesting point is that, despite some ambiguities attributable to listening comprehension issues, more than 90 % of instances of “reading messages aloud” omitted quotative particles, such as “tte”, and reporting verbs, such as “iu” (to say). In modern Japanese, quotative particles and reporting verbs are generally difficult to omit, with few exceptions. Therefore, the omission of these elements in live streaming may be regarded as a rare example in which the characteristics of the medium and the situational context interfere with the grammar itself. The next section will discuss this peculiarity in detail.

5.1 “Reading messages aloud” as a quotative expression

As observed in 4.3, “reading messages aloud” is used in a way that is embedded within the streamer’s own speech (Licoppe and Morel 2018). Therefore, before proceeding with the discussion, we first clarified whether the viewer’s message, as expressed via this method, should be regarded as a quotation or merely as a form of “reading aloud”. In modern Japanese, the most typical approach to indicating quotations is to use the quotative particles “to” or “tte” (or their equivalents “toka” and “nante”) and a reporting verb, as shown in (7) (Fujita 2000; Kato 2010; Sunakawa 2003). In spoken language, one of these two elements may be omitted, but it is uncommon for both to be dropped.^[9]

On the other hand, when it comes to simply reading aloud, neither the quotative particle nor the reporting verb are necessarily required, as seen in (8).

Thus, the peculiarity of omitting quotative particles and reporting verbs while “reading messages aloud” could be identified only when this method was a form of quotation.

Concerning this issue, we considered “reading messages aloud” during live streaming to fulfill the characteristics of a quotation (albeit not entirely) for two reasons. The first reason was that, in the cases we observed, “reading messages aloud” did not merely involve reading messages verbatim but sometimes involved slight modifications to the message form. For example, in (9), a viewer sends a message in the polite form (-hayattemasita) in line 01. However, when the streamer reads it aloud in line 12, the sentence-ending form has been altered to the plain form (-hayatteta).

(9)

The second reason is that, as reported in 4.3, although more than 90 % of the cases observed in “just chatting” streams omitted quotative particles and reporting verbs, a few instances did involve quotative particles. For example, in (10), where the streamer jokingly recommends an unnecessary item (a clip for the nose) to the viewer, the streamer refers to the viewer’s message from line 01 in line 12. In this instance, the act of “reading messages aloud” in line 12 involves a similar modification in form (changing “ureru” (sell) to “urikireru” (sell out)) as seen in (9), as well as the use of the quotative particle “tte” at the end.

(10)

According to Fujita (2000), three conditions (A), (B), and (C) are necessary for a typical syntactic quotation, made explicit through grammatical forms in Japanese, to be established:

1. Intentionality of expression or stance in expression (i.e., something meant to “reproduce”)

2. Characteristics as a linguistic entity (sign) (i.e., used as “something like that”)

3. Characteristics as a syntactic unit (i.e., based on B), (the linguistic entity functions as a cohesive unit and serves as a constituent of the sentence).

4. (Fujita 2000: 10)

Of these, (A) points to the essence of “quotation” as an act of “reproduction”, based on identity. However, the determination of such identity depends on the speaker’s interpretation and, thus, the quotation and the quoted text may involve modifications to the extent that facts remain unchanged. Additionally, (B) indicates that the quoted speech should be recognizable as “quoted speech” and thus be distinguishable from the “narrative part”. Furthermore, (C) suggests that a “syntactic quotation” is typically embedded in the sentence by a quotative particle.^[11]

Based on Fujita (2000), the alterations in linguistic form observed in (9) and (10) suggest that the act of “reading messages aloud” in these instances reflects the speaker’s interpretation as they reproduce the content. Additionally, in (9), the rate of speech, prosody, and pauses distinguish the act of “reading messages aloud” from the streamer’s own speech (thus distinguishing between “quoted speech” and the “narrative part”). Similarly, in (10), with the use of the quotative particle “tte”, all of the conditions outlined by Fujita (2000) are met. These observations indicate that live streaming can function as a venue where “reading messages aloud” may be performed not merely as “reading aloud” but as “quotation”. Therefore, the lack of quotative particles and reporting verbs can in many cases, as in (9), indeed be considered an unusual phenomenon in the grammatical structure of Japanese.

5.2 Differences in relation to existing expressions

Incidentally, previous studies on spoken Japanese have also pointed out that, even in cases clearly recognized as quotations, there is a possibility that quotative particles and reporting verbs may not appear in certain contexts or discourse genres.^[12] For example, (11) is an utterance in a casual conversation among university students discussing an incident in which their mutual acquaintance, M, inquired about another acquaintance, A. In this utterance, the first two sentences in quotation marks use the particle “toka” and the reporting verb “iu,” but the third sentence omits these elements. Sunakawa (2003) notes that such omissions are possible when a story is being reenacted within a conversation and the speakers act as characters.

It is noteworthy that, in English conversation also, introductory clauses such as “He said” may be omitted when participants reenact reported speech (Holt 2007). Holt (2007) observes that such omissions predominantly occur in jokes and are established in three cases: when the introductory clause has already been used in preceding utterances, when prosodic or vocal quality changes accompany the quoted speech, and when the quoted speech includes dialect or offensive language, clearly indicating that it is not the speaker’s own words. In general, when indicating direct speech, English explicitly prefaces quotations with an introductory clause with a subject and a reporting verb before the speech, as shown in (12). In contrast, as shown in 5.1, Japanese explicitly indicates quotations by placing a particle and a reporting verb after the speech. Thus, although the positioning of linguistic markers for quotations differs between the two languages, Holt’s (2007) observations may still be applicable to Japanese to some extent, as illustrated by (11), in which the first two quotations do not omit linguistic markers.

Another exception in spoken Japanese is that of political speeches. For instance, in the speech in (13), the words of another person regarding Japan’s national defense are quoted without the use of quotative particles or reporting verbs.

Okubo (2013) refers to such expressions in political speeches as “Zero-gata inyo hyogen” (Zero-marked quotative) and asserts that they are used to evaluate the words of others who are not present or to draw the audience into the narrative world by vividly depicting the conversation.

The omission of quotative particles and reporting verbs in dialogues and monologues noted in previous studies shares similarities and differences with the omission of these elements in “reading messages aloud” during live streaming. One similarity, as pointed out by Holt (2007), is that transitioning from reading a message (quotation) to the VTuber’s own speech often involves changes in prosody and voice quality. For example, in the utterance from (4) discussed in Section 4.3 (excerpted in (4)′), the quoted part was spoken more slowly than the subsequent part (i.e., the streamer’s own speech).

Similarly, the streamer’s “reading messages aloud” invariably serves the purpose of expressing their own response to the message, akin to politicians’ quotation of others’ words in speeches, to express their own evaluations (Okubo 2013).

Meanwhile, differences include the frequency of occurrence and the origin of the quoted speech. In face-to-face or telephone conversations, the omission of linguistic forms indicating quotations is possible only in very limited contexts, and the majority of quotations are explicitly marked. By contrast, in VTuber streams, as mentioned above, despite some ambiguous cases due to changes in prosody or voice quality, at least 90 % of instances omit both quotative particles and reporting verbs.

Similarly, the omission of linguistic forms in previous studies occurs when the speaker is quoting the words of someone not present (including past or future selves) or when reenacting a different conversation. By contrast, in live streaming, the messages read aloud are the words of viewers currently participating in the stream, and the conversation is happening in the here and now. Thus, at least within the scope of actual spoken utterances, the omission of quotative particles and reporting verbs in “reading messages aloud” possesses a uniqueness not observed in other discourse environments.

5.3 Similarities with existing expressions

The simplest explanation for the omission of quotative particles and reporting verbs in live streaming is that, due to the high frequency of “reading messages aloud”, streamers may avoid repeatedly uttering redundant elements. However, in Japanese radio broadcasts, which also feature a few broadcasters with many listeners and lack shared spatial presence, even when emails or letters from listeners are frequently read aloud, it is observed, as in (14), that the sender is always identified at the beginning of the utterance and, in many cases, quotative particles and reporting verbs are appended at the end. Therefore, it is suggested that the peculiarity of language use in live streaming cannot be explained merely by the act of “reading messages aloud” itself or its high frequency.

Meanwhile, if we shift our perspective and consider that the omission of linguistic forms is driven by more active motivations rather than redundancy, we may obtain a more appropriate explanation. Specifically, although this may seem contrary to the discussion in Section 5.1, we propose that zero-marked quotatives in live streaming still possess an expressive effect of recreating some form of dialogue, similar to analogous utterances observed in face-to-face and telephone conversations (Holt 2007; Sunakawa 2003). Furthermore, we suggest that such an expressive effect is necessitated by both technical and situational issues.

The technical issue, as clarified in Section 4.2, is that live streaming, despite being labeled “live”, is never a truly synchronous CMC mode. This applies not only between the streamer and viewers but also among viewers themselves. The streamer and viewers do not share the same time-space, and there can be no guarantee that they are watching the same video (some viewers might have the browser completely in the background, listening only to the audio). Therefore, the streamer is constantly obliged to clarify what they are referencing.

Similarly, the situational issue is that in “just chatting” streams, the streamer is literally required to chat with viewers. As mentioned, although live streaming is not fully synchronous, participants are not entirely free to act asynchronously. In particular, the streamer, as the only one who can speak audibly, must talk continuously to sustain the viewers’ interest. In Japanese face-to-face or telephone conversations, feedback from the receiver guides the progression of the conversation (Horiguchi 1997). However, in live streaming, such feedback typically arrives with a slight delay. Therefore, without any measures, the streamer’s discourse is persistently at risk of becoming a monologue.

Given these two issues, the omission of quotative particles and reporting verbs may be recognized as a highly effective method for clarifying what is being referenced and restoring a conversational feel. Streamers can introduce viewers’ messages without converting them into their own words, such as using the phrase “to itteru” (“the viewer is saying”), and immediately follow up with responses to those words, overcoming the temporal and spatial gaps between the streamer and viewers, and effectively “recreating dialogue”.

Furthermore, the relationship between the technical and situational issues in “just chatting” streams and language use also explains why the other methods observed in Section 4 are less favorable than “reading messages aloud”. “Reading usernames aloud” and using “adjacency pairs” merely communicate the intention to the specific user who sent the message while being unfriendly to the many viewers who usually do not continuously monitor the chat section. Moreover, “reading usernames aloud”, in particular, can single out one viewer from among many, which may risk antagonizing other viewers, particularly where popular VTubers are concerned. Therefore, the method of reading messages aloud without mentioning usernames and continuing the streamer’s own speech may be regarded as a highly sophisticated practice.

In summary, the zero-marked quotative occurring in “reading messages aloud” shares expressive effects with other zero-marked quotatives observed in different contexts, in terms of “recreating dialogue”. However, the motivations for utilizing such effects clearly possess unique aspects closely tied to the characteristics of live streaming and the interpersonal communication that takes place therein. Additionally, the fact that zero-marked quotatives observed in live streaming differ from similar, previously identified elliptical expressions, and functioning as a format that indicates conversations happening here and now, suggests that new media and situational contexts can generate new linguistic forms that were not previously permissible in traditional Japanese.