KingZ23 commited on
Commit
f340ba4
·
verified ·
1 Parent(s): 1498cdf

Add transcript-only release: 3 benchmarks x 4 accents x 1000 questions

Browse files
Files changed (4) hide show
  1. 2wikimultihopqa.jsonl +0 -0
  2. README.md +177 -0
  3. hotpotqa.jsonl +0 -0
  4. musique.jsonl +0 -0
2wikimultihopqa.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
README.md ADDED
@@ -0,0 +1,177 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-sa-4.0
3
+ language:
4
+ - en
5
+ task_categories:
6
+ - question-answering
7
+ size_categories:
8
+ - 10K<n<100K
9
+ tags:
10
+ - multi-hop-qa
11
+ - retrieval-augmented-generation
12
+ - speech-recognition
13
+ - asr-robustness
14
+ - accented-speech
15
+ configs:
16
+ - config_name: hotpotqa
17
+ data_files:
18
+ - split: test
19
+ path: hotpotqa.jsonl
20
+ - config_name: 2wikimultihopqa
21
+ data_files:
22
+ - split: test
23
+ path: 2wikimultihopqa.jsonl
24
+ - config_name: musique
25
+ data_files:
26
+ - split: test
27
+ path: musique.jsonl
28
+ ---
29
+
30
+ # Spoken Multi-hop QA: ASR Transcripts Across Four English Accents
31
+
32
+ ASR transcriptions of 3,000 multi-hop QA questions, each spoken in four
33
+ English accents and transcribed with Whisper-large-v3. Released as the
34
+ data companion to [*Better Retrieval, Worse Robustness: How Multi-hop RAG
35
+ Amplifies Upstream ASR Errors*](https://arxiv.org/abs/2608.22872)
36
+ (EMNLP 2026, Main Conference).
37
+
38
+ The dataset exists to make one thing cheap to study: what happens to a
39
+ retrieval pipeline when its query arrives through ASR rather than as clean
40
+ text. Every row pairs a gold question (the clean-text oracle input) with
41
+ the transcription actually produced for one accent, so the degradation is
42
+ directly measurable without re-running any speech model.
43
+
44
+ **Code:** https://github.com/ZhenghuaBao/spoken-multihop-rag
45
+
46
+ ## Contents
47
+
48
+ | Config | Questions | Rows | Question types |
49
+ |---|---|---|---|
50
+ | `hotpotqa` | 1,000 | 4,000 | bridge, comparison |
51
+ | `2wikimultihopqa` | 1,000 | 4,000 | + compositional, inference |
52
+ | `musique` | 1,000 | 4,000 | 2-, 3-, 4-hop compositions |
53
+
54
+ One row per (question, accent). Questions are sampled uniformly at random
55
+ from each benchmark's official validation set with seed `42`.
56
+
57
+ ## Schema
58
+
59
+ | Field | Type | Description |
60
+ |---|---|---|
61
+ | `id` | string | Question id from the source benchmark |
62
+ | `benchmark` | string | `hotpotqa`, `2wikimultihopqa`, `musique` |
63
+ | `accent` | string | `us`, `in`, `ph`, `ng` |
64
+ | `voice` | string | Edge TTS voice used to synthesize this row |
65
+ | `question` | string | Gold question text, i.e. the clean-text oracle input |
66
+ | `answer` | string | Gold answer from the source benchmark |
67
+ | `transcription` | string | Whisper-large-v3 top-1 output. `null` when `tts_failed` |
68
+ | `wer` | float | Word error rate of `transcription` against `question`. `null` when `tts_failed` |
69
+ | `tts_failed` | bool | `true` for four rows whose speech synthesis produced an empty file |
70
+
71
+ `wer` uses unit-cost Levenshtein alignment over lowercased,
72
+ whitespace-tokenized words, with no punctuation stripping. This matches the
73
+ implementation used for every number in the paper
74
+ (`evaluation/core/metrics.py`).
75
+
76
+ ## Mean WER
77
+
78
+ | Config | US | IN | PH | NG |
79
+ |---|---|---|---|---|
80
+ | `hotpotqa` | 9.4% | 11.0% | 10.6% | 14.5% |
81
+ | `2wikimultihopqa` | 13.6% | 14.1% | 13.8% | 17.1% |
82
+ | `musique` | 5.2% | 5.5% | 5.7% | 8.0% |
83
+
84
+ ## Usage
85
+
86
+ ```python
87
+ from datasets import load_dataset
88
+
89
+ ds = load_dataset("KingZ23/spoken-multihop-rag", "2wikimultihopqa", split="test")
90
+
91
+ # clean-text vs ASR input for one accent
92
+ ng = ds.filter(lambda r: r["accent"] == "ng" and not r["tts_failed"])
93
+ print(ng[0]["question"]) # oracle input
94
+ print(ng[0]["transcription"]) # what the retriever actually receives
95
+ print(ng[0]["wer"])
96
+ ```
97
+
98
+ ## How it was built
99
+
100
+ **Speech synthesis.** Each question was synthesized with Microsoft Edge TTS
101
+ at 24 kHz, one voice per accent:
102
+
103
+ | Accent | Voice | Role |
104
+ |---|---|---|
105
+ | US | `en-US-JennyNeural` | High-resource baseline |
106
+ | IN | `en-IN-NeerjaNeural` | Non-native (Indian English) |
107
+ | PH | `en-PH-RosaNeural` | Non-native (Filipino English) |
108
+ | NG | `en-NG-EzinneNeural` | Low-resource (Nigerian English) |
109
+
110
+ **Transcription.** Whisper-large-v3, English forced, without timestamps.
111
+
112
+ ## Limitations
113
+
114
+ Read these before using the dataset as an ASR benchmark.
115
+
116
+ - **No audio.** Only transcripts are released. The synthesized speech is
117
+ Microsoft Edge TTS output and is not redistributed here. Audio can be
118
+ regenerated with `data/build_dataset.py` in the code repository, though
119
+ the service is not version-pinned, so regenerated audio and any
120
+ transcripts derived from it will not match these byte for byte.
121
+ - **One voice per accent.** Accent is varied, speaker is not. Results
122
+ should be read as a controlled probe of accent-induced ASR error, not as
123
+ an estimate of population-level performance for any speaker group.
124
+ - **Synthetic speech understates real difficulty.** On a public corpus of
125
+ real Nigerian-accented English, Whisper-large-v3 reaches 28.9% mean WER,
126
+ against 17.1% / 14.5% / 7.9% for the synthesized Nigerian condition on
127
+ 2WikiMultiHopQA / HotpotQA / MuSiQue. The synthesized setting is easier
128
+ by a factor of roughly 1.7x to 3.7x.
129
+ - **Four rows are unusable.** Speech synthesis returned an empty file for
130
+ the four (question, accent) pairs listed below, so no transcription
131
+ exists for them. They carry `tts_failed: true` with null transcription
132
+ and WER, which keeps them out of aggregate statistics.
133
+
134
+ | Config | Accent | id |
135
+ |---|---|---|
136
+ | `hotpotqa` | ng | `5a8603f655429960ec39b601` |
137
+ | `hotpotqa` | ph | `5ade063d5542997dc7907129` |
138
+ | `2wikimultihopqa` | ng | `69a429bb097911ebbdb0ac1f6bf848b6` |
139
+ | `musique` | ng | `3hop2__326964_36852_7713` |
140
+
141
+ - **English only.** No claim is made about other languages.
142
+
143
+ ## Licensing and attribution
144
+
145
+ This dataset is released under **CC BY-SA 4.0**, inherited from the most
146
+ restrictive upstream source. Each row's `question` and `answer` are
147
+ redistributed from one of three benchmarks:
148
+
149
+ | Source | License | Reference |
150
+ |---|---|---|
151
+ | HotpotQA | CC BY-SA 4.0 | [hotpotqa.github.io](https://hotpotqa.github.io/) |
152
+ | 2WikiMultiHopQA | Apache-2.0 | [Alab-NII/2wikimultihop](https://github.com/Alab-NII/2wikimultihop) |
153
+ | MuSiQue | CC BY 4.0 | [StonyBrookNLP/musique](https://github.com/StonyBrookNLP/musique) |
154
+
155
+ HotpotQA's ShareAlike term propagates to derivative works, which is why the
156
+ whole dataset carries CC BY-SA 4.0 rather than a per-config license. The
157
+ `transcription` field holds Whisper-large-v3 output; Whisper is
158
+ MIT-licensed. Accompanying code is Apache-2.0.
159
+
160
+ If you redistribute or build on this dataset, keep the attributions above
161
+ and license derivatives compatibly.
162
+
163
+ ## Citation
164
+
165
+ ```bibtex
166
+ @inproceedings{bao2026better,
167
+ title = {Better Retrieval, Worse Robustness: How Multi-hop {RAG} Amplifies Upstream {ASR} Errors},
168
+ author = {Bao, Zhenghua},
169
+ booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
170
+ year = {2026},
171
+ note = {To appear},
172
+ url = {https://arxiv.org/abs/2608.22872}
173
+ }
174
+ ```
175
+
176
+ Please also cite the source benchmark you use, and Whisper
177
+ (Radford et al., 2023) for the transcriptions.
hotpotqa.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
musique.jsonl ADDED
The diff for this file is too large to render. See raw diff