NewsPod: Automatic and Interactive News Podcasts

1 INTRODUCTION

• 57% of U.S. adults obtain some news through audio-only platforms, and 62% of smart-speaker owners listen to news at least an hour weekly.The introduction presents interactive audio interfaces as a possible way to broaden in-depth news engagement.Source: §1 · p. 1
• Most news podcasts are handcrafted, while unchanged article playback is not adapted to audio as a manually recorded podcast is.A typical episode of The Daily involves a host, three producers, two editors, and an engineer.Source: §1 · pp. 1–2
• NewsPod generates an introductory summary followed by a multi-voice Q&A session, and lets listeners select stories, listen non-linearly, and ask questions for automatic answers.The system aims to bridge automated news reading and curated podcasts.Source: §1 · p. 2
• The paper evaluates the narrative style against automated baselines and a manual reference, and studies listener interaction and whether periodic breaks encourage questions.The reported contributions say the conversational format is preferred over a non-conversational baseline and that many participants ask questions when given the opportunity.Source: §1 · p. 2

Automating News Podcasts

• Unlike open-ended news dialogue that requires user input to advance, NewsPod prepares a duration-targeted podcast that supports passive listening and interruptions at any point.The authors identify Yoshino and Kawahara as the closest conceptual work and distinguish NewsPod by its use of multiple synthesized voices.Source: §2 · p. 2
• Position papers forecast personalized audio news and conversational formats that could support engagement, multiple viewpoints, and follow-up questions.These are forecasts and proposals, not evaluations of NewsPod.Source: §2 · p. 2
• NewsPod follows four recommendations for automated podcasts: audio-adapted bite-size content, key-point summaries, specialized voices, and context-adapted narration.The recommendations were proposed by Lochrie et al..Source: §2 · p. 2

Conversational Interfaces for News Recommendations

• News recommendation systems use dialogue to retrieve or suggest articles, including text-chat retrieval and an Alexa Prize agent that recommends stories.The Alexa Prize system found that less formal, more specific recommendations had significantly higher listener acceptance.Source: §2 · p. 2

Conversational Narration Style

• Prior work reports benefits of multiple distinguishable sources and conversational narration; NewsPod assigns distinct voices to roles that ask and answer questions.The cited work links multiple voices in podcast tours with social presence and enjoyment, and conversational narration with enjoyment and social presence.Source: §2 · pp. 2–3

3 BACKGROUND: NATURAL LANGUAGE PROCESSING

• The paper frames NLP component capabilities and limitations as shaping both NewsPod’s design choices and the interaction possibilities offered to users.It applies its procedure to English and notes that many components have equivalents in other languages.Source: §3 · p. 3

3.1 Question Answering

• NewsPod uses extractive RoBERTa question answering over a limited set of relevant news articles, favoring simplicity and low computation while restricting answers to article-supported questions.Conversational QA adds contextual co-reference and follow-up questions; open-domain QA additionally retrieves relevant passages from a large corpus.Source: §3.1 · p. 3

3.2 Summarization

• Automated abstractive summaries can hallucinate or misrepresent source documents, and NewsPod uses PEGASUS when no journalist-written summary is available.The paper notes that factuality is not guaranteed for generated summaries [22].Source: §3.2 · p. 3

3.3 Question Generation

• Question generation is under-constrained and can produce generic questions; reported errors include poor fluency, off-target questions, and focus on unimportant paragraph details.NewsPod fine-tunes GPT-2 for question generation, produces seven questions per paragraph, then uses a graph-based method to select relevant questions.Source: §3.3 · pp. 3–4

4.1 Selecting Podcast Segments

• Lee chooses a five-minute target duration and selects three segments about a Tesla self-driving ban, the Rohingya crisis, and an Amazon union vote.The podcast begins with a greeting and transitions into the first segment.Source: §4.1 · p. 4

4.2 Segment Reading

• The Tesla Self Driving ban segment opens with a recent headline and summary, then a Q&A speaker change asks what counts as fully autonomous; the answer describes Tesla’s current system.It identifies the current system as Level 2, requiring driver attention and hands on the wheel, and describes Level 5 as never requiring driver attention.Source: §4.2 · p. 4
• The listener can open a transcript when acronyms are unfamiliar, and a speaker change can refocus attention after a distraction.The demonstration illustrates these features during the Tesla segment.Source: §4.2 · p. 4

4.3 Interacting with the Podcast

• During playback, Lee can interrupt with questions; the system answers when it finds a response and resumes the podcast after either an answer or a failure.It answers “Who are the Rohingya?” but cannot answer why Amazon workers might oppose forming a union.Source: §4.3 · p. 4

5.1 Podcast Interface

• Listeners can navigate segments, follow a sentence-by-sentence transcript, and activate live Q&A with a microphone button rather than continuous listening.The authors cite accidental detection of podcast speech and privacy concerns as reasons for button-based activation.Source: §5.1 · pp. 4–5

5.2 Podcast Organization

• Users choose topics and podcast duration; an inverted-pyramid segment places a headline and summary before Q&A pairs, so shorter versions retain the opening information.A segment can be truncated at the word target, such as after 100 of 200 words.Source: §5.2 · p. 5

5.3 Identifying Podcast Segments

• Related articles are grouped as source articles; NewsPod selects a short headline and uses a qualifying human-written summary or the highest-likelihood PEGASUS summary.The introduction is intended to give listeners a high-level understanding before the Q&A.Source: §5.3 · p. 6

5.4 Generating a Q&A Session

• The Q&A pipeline filters article paragraphs, generates candidate questions, links questions to answer-bearing paragraphs, then greedily selects pairs while limiting repeated content.It generates seven question types per valid paragraph; the graph uses answer frequency to approximate question relevance.Source: §5.4 · p. 7

5.5 From transcript to speech

• Three distinct voices distinguish introductions and answers, questions, and quotes; SSML rules add emphasis and silences to quote paragraphs.The selected WaveNet voices use American English, matching the study’s mainly U.S. crowd-worker participants.Source: §5.5 · p. 8

5.6 Live Q&A Response

• Live Q&A uses an extractive model that answers only from segment paragraphs and cannot combine information or synthesize answers to some open-ended questions.The model takes about four seconds on average; when it cannot find a confident answer, the system asks the listener to rephrase or resumes the podcast.Source: §5.6 · p. 8

5.7 Implementation Details

• The system continuously preprocesses news into segments; a typical segment uses six articles and 50–100 paragraphs, with Q&A generation taking about 50 seconds.Question generation averages nine seconds and QA graph construction about 40 seconds.Source: §5.7 · pp. 8–9

6.1 Study A Design

• Study A compared four 200-word settings to test Q&A against article reading, automated against hand-written Q&A, and question ordering against random selection.The settings were Baseline, QA Best, QA Rand, and a journalist-authored Reference; the target corresponded to about 90 seconds at 120–150 words per minute.Source: §6.1 · p. 9

6.2 Study A Specifics

• The 61 Mechanical Turk participants listened at a computer, selected five of seven topics, and could not interrupt with their own questions.The authors chose this setting to focus participants on podcast content and left other listening contexts for future work.Source: §6.2 · p. 9, §6.3 · p. 10

6.3 Study A Results

• 56% opened the transcript at least once, 22% paused at least once, and 5% skipped a segment, with no significant differences across settings..Source: §6.3 · p. 10
• The hand-written Reference significantly outperformed every other setting on interest and coherence; QA Best ranked second, while QA Rand scored 0.8 stars lower than QA Best on coherence.QA Best was narrowly above Baseline on both ratings, without a statistically significant difference.Source: §6.3 · p. 10
• Interestingness and coherence ratings correlated at 0.5, while prior awareness correlated with interestingness at 0.24 and coherence at about 0.1.The authors suggest coherence may be a prerequisite for interest.Source: §6.3 · p. 10
• Among open-ended feedback, 30 participants expressed positive interaction; 11 mentioned problematic speech quality, and 9 found the interface easy to use.Three participants each mentioned speech that was too fast, a desire for more segment choices, or the usefulness of the transcript.Source: §6.3 · p. 10
• 80% of QA Best listeners said they would use the system, and participants significantly preferred Q&A over unaltered-article reading.Automated Q&A did not match the hand-written reference on coherence and interestingness.Source: §6.3 · p. 10
• Greedy QA Graph ordering significantly outperformed random question selection on coherence and interestingness.The randomized condition also randomized the answering paragraph.Source: §6.1 · p. 9, §6.3 · p. 11

7.1 Study B Design

• Study B replaced microphone input with a text box and offered two recommended questions per segment; breaks added a 15-second question period after each segment.Participants could also type or select a recommended question at any point during a segment.Source: §7.1 · p. 11

7.2 Study B Specifics

• Forty participants were split evenly between break settings and chose four of six newly generated segments to keep the study within 15 minutes.Source: §7.2 · p. 11

7.3 Study B Results

• Study B segments averaged 3.9 stars for interest and 4.1 for coherence, versus 3.6 and 4.2, respectively, in Study A.Source: §7.3 · p. 12
• Breaks significantly increased average recommended questions clicked from 1.5 to 3.2, while typed questions rose nonsignificantly from 0.9 to 1.1.Source: §7.3 · p. 12
• With breaks, 85% of listeners asked at least one question, versus roughly half without breaks.Table 2 reports interaction statistics for 40 participants randomly assigned to the two settings.Source: Table 2 · p. 11, §7.3 · p. 12
• Participants rated 79% of system-generated answers Good/Relevant, but rated 76% of answers to their own typed questions Irrelevant/Confusing.The paper identifies this as a limitation of applying extractive question answering to open question answering.Source: §7.3 · p. 12
• Questions were categorized across seven types: factoid (36%), synthesis (18%), encyclopedic (14%), clarification (13%), prediction (9%), rhetorical (5%), and self-relevance (5%).Source: §7.3 · p. 12

• Almost 90% would use the hand-written Reference again, compared with 63% for the unmodified-news Baseline and 80% for QA Best.The authors report that neural generation of summaries, questions, and content ordering still loses quality relative to manual curation.Source: §8 · p. 12
• With invitations and silences in the script, the proportion asking questions rose from 55% to 85%; current question-answering systems remained too limited for diverse listener questions.Source: §8 · p. 12
• The studies suggest audio interfaces could support in-depth news engagement by letting listeners choose topics, ask questions, and steer discussion dynamically.Source: §8 · p. 12

• Participants mainly used laptops at desks, so the studies do not establish how NewsPod performs while driving, walking, cooking, or doing other everyday activities.Source: §9 · p. 12
• Quotation handling and greeting or closing wording were not explicitly evaluated, so conclusions should stay limited to the main elements tested, such as the Q&A format.Source: §9 · pp. 12–13
• Mechanical Turk participants differ from the U.S. population on age, gender, and income, which could affect the results.Source: §9 · p. 13
• Around 15% noticed low speech quality, and the Q&A system failed to satisfy a majority; the studies did not measure how much these components affected evaluations.Source: §9 · p. 13
• Summaries and generated questions entered the podcast without human verification and could contain factual errors that mislead or confuse listeners.Source: §9 · p. 13
• The algorithms do not exercise journalistic judgment about balance and sources; for example, they present only one quotation without an answering quotation.Source: §9 · p. 13

• A listener’s question could change later podcast content or prompt follow-up questions, rather than having the system resume its planned segment unchanged.Source: §10 · p. 13
• Remembering prior podcasts could let future episodes incorporate updates related to topics discussed with a returning listener.Source: §10 · p. 13
• Music and sound could serve as communication tools in automated news podcasts and help soften the perceptibility of automated speech flaws.Source: §10 · p. 13

11 CONCLUSION

• 80% of participants said they would use NewsPod again to stay informed after finding its Q&A narrative more enjoyable and less monotonous.The second study found that 85% of listeners asked questions when appropriate breaks were included.Source: §11 · p. 13
• NewsPod organizes news into a simulated conversation with several automated voices, and listeners can ask questions that the system attempts to answer automatically.Source: §11 · p. 13
• A majority of participant-prompted questions tripped up current neural network-based question-answering models, revealing current NLP limitations.The questions were diverse.Source: §11 · p. 13
• The authors suggest this alternative news-consumption approach may eventually broaden the base of users who engage deeply with news.Source: §11 · p. 13

A USABILITY STUDY INSTRUCTIONS

• Participants were told the task would take no more than 15 minutes and included a tutorial followed by a seven-minute podcast.Source: A USABILITY STUDY INSTRUCTIONS · p. 15
• The task required Chrome or Firefox and enabled sound; participants could leave at any point but would not complete the HIT.Source: A USABILITY STUDY INSTRUCTIONS · p. 15
• Participants could complete the task once, agreed to research use of collected data, and were told no personally identifiable information would be collected.They were asked to accept the terms and perform the task to the best of their ability.Source: A USABILITY STUDY INSTRUCTIONS · p. 15

B EXAMPLE PODCAST SEGMENTS

• The four compared settings are Baseline, QA Rand, QA Best, and Reference; the example concerns the Large Iceberg Break-off segment.Figure 8 presents these four settings.Source: B EXAMPLE PODCAST SEGMENTS · p. 15, Figure 8 · p. 16
• For the Large Iceberg Break-off topic, QA Rand and QA Best did not extract quotations, so neither automatic segment included a quotation paragraph.Source: B EXAMPLE PODCAST SEGMENTS · p. 15
• The baseline segment describes the calving event, monitoring by the British Antarctic Survey, and the station’s winter closure and relocation.It also reports that researchers found no evidence climate change played a role.Source: B EXAMPLE PODCAST SEGMENTS · p. 16
• The QA Rand example summarizes the iceberg’s size and break-off, while its prompt asks who said there was no evidence of climate change.Source: B EXAMPLE PODCAST SEGMENTS · p. 16
• The example’s question prompts ask who described the picture as unclear and what the iceberg might do; another asks when the cracks were discovered.The supplied answer text says the cracks were first discovered a decade earlier.Source: B EXAMPLE PODCAST SEGMENTS · p. 16