An Honest Conversation

Transparently Combining Machine and Human Speech Assistance in Public Spaces

Thomas Reitmaier, Simon Robinson, Jennifer Pearson, Matt Jones Computational Foundry @ Swansea University Dani Raju Studio Hasi, Mumbai

April 28^th 2020, CHI 2020, Honolulu, Hawaii

Overview

Introduce the TalkBack system and its previous iterations.
Discuss results from longitudinal deployments of TalkBack.
Contextualize results and surface wider implications for HCI/AI

Background & Context

Dharavi

is one of Asia’s largest slums (~700,000 residents)
is one of the most densely populated areas in the world
has a large informal economy
is linguistically diverse (Hindi, Marathi, Urdu, & English).
residents have varying levels of textual & technological literacies

Phase One: Exploratory Design

In Dharavi we conducted design workshops to imagine and experiment with designs for future technologies.

We rapidly prototyped resulting ideas for a public smart speaker and tested it in Dharavi using a wizard of oz approach.

Phase Two: StreetWise Deployment

Machine-Powered with instant responses (MPI)

Human-Powered with delayed responses (HPD)

Given the success of exploratory work, we built two high-fidelity prototypes, which together we call StreetWise.

The first StreetWise system was essentially a Hindi-language version of Google Assistant. To use it, people press the button and ask a question. The response from Google Assistant is then played through the speaker. Thus the system was Machine-powered and gave instant responses.

In contrast, the second system was human-powered and gave delayed responses. The initial interaction is the same.

You press the button and ask a question. Only this time the response insn’t immediate.
Instead the StreetWise system prompts the user to remember a four-digit number such as 1234 and return after about 10 minutes.
During those 10 minutes, the question is forwarded to a crowd of people, who subsequently record an answer.
Upon returning, users enter their code using the keypad and the system plays the answer.

Phase Two: StreetWise High-level Findings

People were willing to experiment with speech interaction.
They asked over 12,000 questions during a 40-day longitudinal deployment.

People generally appreciated the speed of MPI systems.
People were often not satisfied with the quality of MPI answers.

People generally appreciated the quality of HPD answers.
People were often not satisfied with the speed of the HPD system.

The TalkBack prototype

We combined both MPI & HPD systems into the TalkBack prototype by adapting the StreetWise open-source toolkit.

Deployment & Results

TalkBack Deployment

We deployed the TalkBack system:

for 25 days and
across 10 shops in Dharavi.

TalkBack Deployment

Interaction Flow & Results

People asked 4018 questions

Results (continued)

Question & Answer Examples

Q: Who is playing today?
- A: Today is IPL of Rajasthan Royals and Chennai Super Kings?

Q: What is the exit poll?
Q: When is the election?
Q: Who is the corporator of ward number 184?
Q: Documents required for PAN card?

Who is playing today? This is an example of a typical, indexical question that exposes gaps in current conversational agent abilities. A question where Google Assistant responds: I can’t help you with that.

It is also the type of question, where humans can fill in the gap … through shared experience and background knowledge. In this case the answerer could tell that the Indian Premiere League cricket match was the likely answer.

Because our deployment took place over the 2019 elections, there were also lots of political questions:

What is the exit poll? When is the election?
Who runs ward 184?
Or what documents are required for a PAN card – a type of tax identification number?

Again the crowd of human answerers were able to step in and provide appropriate answers where Google Assistant failed to do so.

Shopkeepers & Answerers

Some shopkeepers occasionally helped through intermediation.

Answerers noted it could be difficult – especially for a computer – to understand the dialect of Dharavi and the way residents phrased questions.

Answerers reported that it could be tiring to answer repetitive questions;
- they even suggested automation.
Answerers were motivated by, and found value in, directly helping residents of Dharavi.

Some shopkeepers reported occasionally helping first-time users. For instance, by demonstrating the device or by indicated how users can rephrase their questions so they were more likely to be understood and answered by Google Assistant.

Answerers similarly noted it could be difficult – especially for a computer – to understand the dialect of Dharavi and the way residents phrased questions.

They also reported that it could be tiring to answer repetitive questions and even suggested ways of automating responses in cases of repetition.

Another theme that emerged during interviews was that answerers were motivated by directly helping users.

The paper has a lot more details on user questions and system responses along with comparisons to previous studies.

In the paper, we also report on a subsequent deployment where we follow the answerers suggestion to automate responses to repetitive questions. We did this by pooling already answered questions into a corpus. Next we extended the TalkBack devices to use incoming question transcriptions to find and retrieve answers from the corpus. But we also still gave users the option of forwarding their questions to a human. Please checkout the paper for more details.

Wider Implications for HCI/AI Research

71% Discussion

Users requested a human response to their questions in 71% of cases, if they were not satisfied with the automatic machine answer.

Bloomberg

BBC

Google

Lau

Ghost Work

It takes human intelligence to push the boundary of what machines can learn
Relegating people to the shadows reinforces the myth that machines (alone) are smarter than they actually are.
People working in the growing shadow of AI face harsh realities.

We would all be better served if we knew how AI supply chains actually functioned.

Our work is also in dialogue with critical scholarship that scrutinizes subcontracted AI supply chains, such as Gray & Suri’s book titled Ghost Work.

The book argues that it takes human intelligence to push the boundary of what machines can learn.
Think about the “Who is playing today?” question … where the human answerer was able to draw on their broad knowledge of the world. It is thanks to millions of similar acts of transcribing or data labeling by subcontracted workers that machines are able to learn.
Relegating people doing this type of work to the shadows, however, reinforces the myth that machines (alone) are smarter than they actually are.
This exacerbates the harsh realities of those working in the shadows of AI, but who also form the critical infrastructure on which AI services depend.

But it doens’t need to be that way, as the GhostWork book argues and our research demonstrates.

We would all be better served – users and Ghost Workers – if we knew how AI supply chains actually functioned.

Combining Human & Machine Intelligence

Through TalkBack we combine human & machine intelligence in a way that mirrors core aspects of AI services that that rely on humans-in-the-loop.

When AI systems fall short, humans pick-up the work.
With TalkBack this handover is done transparently and at the point of interaction.

Transparency

Everyday definition of transparent:

allowing light to pass through so that objects behind can be distinctly seen.

Redefinition for AI & Computer Scientists:

functioning without the user being aware of its presence.

For users of AI-infused systems?

Lets think about that word – transparency – for a minute and lookup its definitions.

A window, for example is transparent, because you can see whats behind it.

But this is only the colloquial meaning of the word.

But in computer science terms like “information”, “communication”, or in our case “transparent” loose their colloquial sensibilities. Here the Dictionary lists a domain-specific definition … the opposite of its everyday meaning. When something functions without the user being aware of its presence. Typically it is used when speaking about a lower level of abstraction. Something the user doesn’t need to concern themselves with.

But what does or rather should transparency mean for the users of AI-infused systems, such as speech assistant services.

Conclusions

We need to have more honest conversations with users.
AI needs humans to fill in the gaps.
Users awareness and control for enlisting human help are important design considerations for AI services.

We thank Manik, Aparna, Shashank, Manjiri and Nayeem for their help with the deployments; shopkeepers and question answerers; as well as reviewers for their thoughtful comments. This work was supported by EPSRC grants EP/M00421X/1, EP/M022722/1 and EP/R511614/1.

Our work clearly resonates more with the colloquial, rather than the technical, meaning of transparency.

We also shows that AI needs humans and benefits from their creativity, general knowledge, and ability to understand.

The TalkBack system further demonstrates that it is possible, practical, and beneficial to combine machine and human intelligence in such a way that it is clear to the user when and how human help is enlisted.

While we focused on speech assistance in this work, our clear recommendation is to consider how the design ethic we demonstrate might improve other AI-infused products that rely on, but currently render invisible, the humans that are brought into the loop.

With that being said, we would like to thank our funders and especially all the people who contributed to this work.