--- tags: - espnet - audio - automatic-speech-recognition language: en datasets: - lrs3 license: cc-by-4.0 --- ## ESPnet2 AVSR model ### `espnet/msk_lrs3_train_avsr_avhubert_large_extracted_en_bpe1000` This model was trained by ms-dot-k using lrs3 recipe in [espnet](https://github.com/espnet/espnet/). ### Demo: How to use in ESPnet2 Follow the [ESPnet installation instructions](https://espnet.github.io/espnet/installation.html) if you haven't done that already. ```bash cd espnet pip install -e . cd egs2/lrs3/avsr1 ./run.sh --skip_data_prep false --skip_train true --download_model espnet/msk_lrs3_train_avsr_avhubert_large_extracted_en_bpe1000 ``` # RESULTS ## Environments - date: `Thu Sep 28 23:59:06 KST 2023` - python version: `3.8.18 (default, Sep 11 2023, 13:40:15) [GCC 11.2.0]` - espnet version: `espnet 202308` - pytorch version: `pytorch 1.12.0` - Git hash: `5d0758e2a7063b82d1f10a8ac2de98eb6cf8a352` - Commit date: `Wed Aug 30 18:03:42 2023 -0400` ## exp/asr_train_avsr_avhubert_large_extracted_en_bpe1000 ### WER |dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err| |---|---|---|---|---|---|---|---|---| |inference_asr_model_valid.acc.ave/test|1321|9890|98.5|1.1|0.4|0.2|1.7|8.8| ### CER |dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err| |---|---|---|---|---|---|---|---|---| |inference_asr_model_valid.acc.ave/test|1321|49750|99.4|0.2|0.4|0.2|0.8|8.8| ### TER |dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err| |---|---|---|---|---|---|---|---|---| |inference_asr_model_valid.acc.ave/test|1321|14940|98.8|0.8|0.4|0.3|1.5|8.8| ## ASR config
expand ``` config: conf/train_avsr_avhubert_large.yaml print_config: false log_level: INFO drop_last_iter: false dry_run: false iterator_type: sequence valid_iterator_type: null output_dir: exp/asr_train_avsr_avhubert_large_extracted_en_bpe1000 ngpu: 1 seed: 0 num_workers: 1 num_att_plot: 3 dist_backend: nccl dist_init_method: env:// dist_world_size: 4 dist_rank: 0 local_rank: 0 dist_master_addr: localhost dist_master_port: 54927 dist_launcher: null multiprocessing_distributed: true unused_parameters: true sharded_ddp: false cudnn_enabled: true cudnn_benchmark: false cudnn_deterministic: true collect_stats: false write_collected_feats: false max_epoch: 20 patience: null val_scheduler_criterion: - valid - loss early_stopping_criterion: - valid - loss - min best_model_criterion: - - valid - acc - max keep_nbest_models: 10 nbest_averaging_interval: 0 grad_clip: 5.0 grad_clip_type: 2.0 grad_noise: false accum_grad: 1 no_forward_run: false resume: true train_dtype: float32 use_amp: false log_interval: null use_matplotlib: true use_tensorboard: true create_graph_in_tensorboard: false use_wandb: false wandb_project: null wandb_id: null wandb_entity: null wandb_name: null wandb_model_log_interval: -1 detect_anomaly: false pretrain_path: null init_param: [] ignore_init_mismatch: false freeze_param: [] num_iters_per_epoch: null batch_size: 16 valid_batch_size: null batch_bins: 1000000 valid_batch_bins: null train_shape_file: - exp/asr_stats_extracted_en_bpe1000/train/speech_shape - exp/asr_stats_extracted_en_bpe1000/train/text_shape.bpe valid_shape_file: - exp/asr_stats_extracted_en_bpe1000/valid/speech_shape - exp/asr_stats_extracted_en_bpe1000/valid/text_shape.bpe batch_type: folded valid_batch_type: null fold_length: - 800 - 150 sort_in_batch: descending shuffle_within_batch: false sort_batch: descending multiple_iterator: false chunk_length: 500 chunk_shift_ratio: 0.5 num_cache_chunks: 1024 chunk_excluded_key_prefixes: [] train_data_path_and_name_and_type: - - dump/extracted/train/feats.scp - speech - kaldi_ark - - dump/extracted/train/text - text - text valid_data_path_and_name_and_type: - - dump/extracted/val/feats.scp - speech - kaldi_ark - - dump/extracted/val/text - text - text allow_variable_data_keys: false max_cache_size: 0.0 max_cache_fd: 32 valid_max_cache_size: null exclude_weight_decay: false exclude_weight_decay_conf: {} optim: adam optim_conf: lr: 0.0003 scheduler: warmuplr scheduler_conf: warmup_steps: 8000 token_list: - - - S - ▁THE - ▁TO - ▁A - ▁AND - T - ▁I - '''' - ▁OF - ▁THAT - ▁IN - ING - D - ▁YOU - ▁WE - E - ▁IT - N - ED - ▁IS - R - M - P - Y - ▁FOR - ER - ▁THIS - ▁WAS - RE - C - G - ▁SO - A - ▁BE - ▁THEY - ▁HAVE - ▁ARE - O - ▁ - ▁ON - ▁WITH - LY - ▁WHAT - U - IN - AL - ▁MY - I - ▁S - ▁DO - B - ▁RE - L - ▁ME - ▁CAN - ▁BUT - LE - ▁ABOUT - OR - ▁NOT - VE - F - AR - RA - ▁ALL - ▁OUR - ▁PEOPLE - ▁AT - ▁C - ▁AS - IC - ▁OR - ▁LIKE - W - LL - K - ▁AN - ▁THERE - ENT - ▁ONE - ES - ▁HE - RI - 'ON' - ▁P - ▁IF - ▁FROM - ▁JUST - ▁WHEN - TH - ▁YOUR - ▁US - CE - ▁DE - ION - IT - ▁KNOW - ▁HOW - ▁T - ▁BECAUSE - CH - V - ▁OUT - ▁B - ▁UP - ▁E - ▁F - TE - ▁HAD - ▁CO - LI - ▁TIME - ▁THEIR - ▁MORE - UR - ▁WHO - ▁GO - EN - ▁G - ATION - AN - CK - TER - ▁SEE - ▁WOULD - ▁THESE - ▁NO - ▁THEM - ▁BY - ▁THINK - ▁WERE - IL - ATE - ▁GET - ▁SE - ▁VERY - ▁GOING - ▁EX - ▁REALLY - ITY - ▁WAY - ▁CON - H - RO - ▁DON - ▁NOW - ▁W - X - NE - GE - ▁WILL - ▁MAKE - ▁WANT - ▁OTHER - ▁SOME - LA - ▁WORLD - ▁ST - ▁COULD - TION - ▁WORK - MENT - ▁SHE - ▁NEED - ▁PA - LO - OL - ▁SAY - ▁MO - ▁BA - IST - ▁FA - IR - ▁MA - ERS - ▁HAS - VER - ▁PO - IVE - ▁PRO - ▁LIFE - ▁INTO - ▁WHICH - ▁THINGS - ▁WHERE - ND - ▁LA - MP - ▁BEEN - ▁SOMETHING - MA - ▁THOSE - US - ▁NEW - ▁CH - ▁RA - ▁ACTUALLY - ▁YEARS - ▁EVEN - ▁TAKE - ▁LOOK - UL - ▁RIGHT - ▁SAID - TIC - ▁UN - Z - AS - ▁DAY - ▁HER - IDE - ▁BO - ▁THAN - ▁HERE - ▁OVER - ▁BACK - ▁LO - ▁FIRST - ▁DI - ▁MOST - ▁COME - ▁ALSO - VI - KE - ▁WELL - IES - ABLE - UT - ▁THEN - ▁CHANGE - AGE - ▁MUCH - '0' - ▁MEAN - OM - ▁CA - CO - AT - ▁ANY - ▁HAPPEN - ▁ONLY - ▁PART - ▁SU - ▁HIS - ▁SP - ▁DIS - ANCE - ID - ▁MANY - ▁RO - '}' - ▁{ - OW - ▁O - IGHT - ▁GOOD - UM - ▁LIVE - ▁LOT - ▁D - ▁TWO - ▁LI - ▁THING - ▁GOT - ▁TELL - AC - ▁EVERY - EL - CI - ▁WHY - TA - FUL - ▁BEING - ANT - EST - ▁LEARN - ▁COMP - ▁DID - URE - PE - ▁FEEL - ▁DIFFERENT - ▁PRE - MO - TI - ▁HO - ▁K - ▁LITTLE - IV - ▁THROUGH - ▁1 - INE - ▁KIND - ME - RY - ▁LET - ▁HELP - UN - ICAL - ▁VI - ▁SAME - ECT - ▁HUMAN - ▁GIVE - HE - ▁TALK - ▁FE - ▁HA - ▁OWN - ▁AROUND - ▁USE - IS - ALLY - ▁IDEA - RESS - ▁PROBLEM - ▁PERSON - ▁TE - ▁FI - ▁FIND - ▁SA - ▁START - OS - TED - ▁BU - LG - NCE - ATED - ▁YEAR - ▁DIDN - ▁LOVE - HO - '5' - ▁DOWN - ▁SCHOOL - ▁TODAY - ▁QUESTION - ▁HEAR - DI - ▁MAN - ▁CAR - MI - ▁GREAT - ▁CR - ▁DOING - IG - ▁FACT - ▁LE - ▁LONG - OUS - ▁RU - ▁PUT - ▁AFTER - ▁EN - ▁M - ▁GA - ▁SHOW - OP - ▁SI - ▁SHOULD - ▁NE - ▁STA - ▁NEVER - ▁BIG - NS - ▁THOUGHT - ISH - ▁MIGHT - ▁ACT - ▁PLACE - LU - END - IZE - ▁REAL - ▁BETTER - ATIVE - IA - ▁UNDERSTAND - ▁POWER - ▁IMPORTANT - IAN - ▁BRAIN - ▁SYSTEM - UAL - NESS - ▁END - ▁ABLE - ▁BEFORE - ▁STORY - ▁OFF - TOR - FF - ▁STARTED - ▁DR - ▁MADE - ▁ASK - NA - ▁HU - ▁CREATE - ATING - ▁BI - ARY - ▁HIGH - ▁HIM - BO - ITION - ▁THREE - ▁PER - ▁AM - ▁CALLED - ▁APP - ▁CAME - ▁WOMEN - ▁OLD - TY - ▁PLAY - '4' - PP - ▁PH - AG - ▁BELIEVE - ▁HOME - ARD - ▁FRIEND - ▁RI - ▁FOUND - HA - ▁HAND - ▁DA - ▁STILL - ▁NA - ▁WORD - ▁TRANS - ▁HEALTH - OUND - ▁BUILD - ▁CARE - ▁WI - ▁NEXT - ▁THANK - ▁TURN - ▁TOGETHER - ▁TA - ▁BECOME - ▁EXPERIENCE - VING - ▁EM - ▁MEN - ISE - ▁MAR - ▁EACH - ▁WENT - ▁TRI - ▁POINT - ▁LAST - ▁MAYBE - ▁EVER - ▁CALL - WARD - ▁CHILDREN - ▁DOES - CA - ▁BIT - UC - LIC - UGH - ▁EXAMPLE - ▁FEW - ITIES - ▁ANOTHER - SH - ▁TH - ▁ALWAYS - ▁H - ▁READ - ▁INTEREST - FORM - ▁STATE - ▁MOVE - IOUS - ▁MIND - 'NO' - AM - ▁TEACH - ▁2 - ▁HARD - ▁WANTED - ▁20 - ▁GROW - ▁JOB - DA - ▁TOO - ▁VA - OME - ▁MAY - '8' - ▁SOCIAL - ▁HI - ▁FOOD - BI - ▁JO - ▁COURSE - ▁FR - BA - ▁MOMENT - ▁AGAIN - ▁DOESN - ▁SHARE - ▁AWAY - ▁BETWEEN - ▁LESS - ▁SHA - ▁MONEY - ▁UNDER - BER - ▁DEVELOP - ▁SECOND - ▁NUMBER - ▁ART - QUE - ▁FAMILY - '1' - '7' - ▁SH - '6' - ▁EVERYTHING - ▁FAR - ▁WORKING - ▁KIDS - ▁PLAN - ▁CHA - ▁AGO - ▁PI - ▁ENOUGH - ISM - ▁AMERICA - ▁THINKING - ▁USED - ▁REASON - ▁TRY - ▁SOMEONE - ▁GENE - ▁CU - ▁STUDENT - ▁TOLD - ▁GU - ▁TRYING - ▁LEAD - ▁MYSELF - ▁BEST - ▁FUTURE - ▁MILLION - ▁SMALL - ▁TECHNOLOGY - LESS - ▁PASS - ▁DONE - ▁YOUNG - '9' - ▁SPACE - ▁WATER - ▁MATTER - ▁OPEN - ▁COUNTRY - ▁REMEMBER - ▁TALKING - ▁REALIZE - LAND - ▁RESEARCH - Q - IAL - ▁WAR - ▁GROUP - ▁BOOK - ▁KEEP - ▁DEF - ▁STOP - ▁HOPE - ▁CONNECT - ▁SENSE - ▁ANSWER - ▁WALK - ▁DESIGN - ▁WEEK - ▁LANGUAGE - ▁DATA - ▁LOOKING - ▁PERCENT - ADE - ▁CLASS - ▁WHOLE - ▁BODY - ▁FOUR - ▁OFTEN - ▁ELSE - ▁WITHOUT - ▁PROCESS - ▁FREE - ▁MAKING - IBLE - ▁BRING - ▁CHILD - ▁GETTING - ▁PROBABLY - ▁ALLOW - ▁SPEAK - ▁COMMUNITY - ▁HAVING - ▁TOOK - ▁OP - ▁JU - ▁MU - ▁FACE - ▁INFORMATION - ABILITY - ▁NAME - ▁NI - '2' - ▁GIRL - ▁CELL - ▁ANYTHING - ▁SCIENCE - ▁STAND - ▁WHILE - ▁SUCH - '000' - ▁CASE - J - ANG - ▁FIVE - ▁GUY - ▁FUN - ▁BUSINESS - ▁ROOM - ▁SELF - ▁LIVING - ▁SURE - ▁IMAGINE - ▁ASKED - ▁MIS - ▁ENERGY - ▁PROJECT - ▁STUDY - ▁DREAM - ▁10 - ▁STORIES - ▁ALREADY - ▁TERM - ▁EFFECT - ▁KNEW - ▁SOCIETY - ▁PRODUCT - ▁PRETTY - ▁EVERYONE - ▁HEAD - ▁19 - ▁JA - ▁LIGHT - ▁LISTEN - ▁MUSIC - ▁LARGE - ▁QUITE - ▁J - ▁BOTH - ▁CHALLENGE - ▁SORT - ▁FELT - ▁TREAT - ▁EDUCATION - ▁WRONG - ▁YOURSELF - ▁MIL - ▁OURSELVES - ▁SOUND - ▁PROGRAM - ▁3 - ▁CLOSE - ▁QUA - ▁SINGLE - ▁MINUTE - ▁NOTHING - ▁ENVIRONMENT - ▁PUBLIC - ▁ORDER - ▁OB - ▁TRUE - ▁STEP - ▁WONDER - ▁NIGHT - ▁YET - ▁EYE - ▁LEFT - SHIP - ▁VALUE - ▁WHETHER - ▁MOTHER - ▁SIMPLE - ▁NU - ▁WOMAN - ▁LU - ▁CONTROL - ▁COMING - ▁SAW - ▁LEVEL - ▁TEST - ▁POSSIBLE - ▁ACROSS - ▁HOUSE - ▁WATCH - ▁GOVERNMENT - ▁PARENTS - ▁HALF - ▁TEN - ▁DEEP - ▁CANCER - ▁ISSUE - ▁LATER - ▁SOMETIMES - ▁ANIMAL - ▁SUPPORT - ▁EAT - ▁CULTURE - ▁FULL - ▁INSTEAD - ▁EARTH - ▁DISEASE - ▁MIN - ▁GAME - ▁DECIDED - ▁ALMOST - ▁SUCCESS - ▁AMAZING - ▁DRIVE - ▁DU - ▁EMOTION - ▁GLOBAL - ▁EQU - ▁PLANET - ▁CERTAIN - ▁HISTORY - ▁MEET - ▁TRAIN - ▁COMPUTER - ▁BECAME - ▁TEAM - ▁DISCOVER - ▁DIFFERENCE - WAY - ▁FOCUS - ▁PAST - ▁RESULT - ▁MONTHS - ▁MODEL - ▁YES - ▁VO - ▁COUNTRIES - ▁STUFF - ▁FIGURE - ▁30 - ▁PATIENT - ▁SPEND - ▁ENTIRE - ▁INDIVIDUAL - ▁UNTIL - ▁THOUGH - ▁DECISION - ▁CHOICE - ▁AFRICA - ▁RELATIONSHIP - ▁BREAK - ▁SOMEBODY - ▁FOLLOW - ▁CONVERSATION - ▁LEAVE - ▁THOUSAND - ▁SIGN - ▁SINCE - ▁DIFFICULT - ▁IMPACT - ▁HOURS - ▁COUPLE - ▁CAUSE - ▁PARTICULAR - ▁DOCTOR - ▁TAKING - ▁COMPANY - ▁EVERYBODY - ▁50 - ▁DIRECT - ▁EXPECT - ▁200 - ▁ORGAN - ▁EXACTLY - ▁THEMSELVES - ▁HAPPY - ▁MUST - ▁SAFE - ▁BASED - ▁BEAUTIFUL - ▁PHONE - ▁AGAINST - ▁WRITE - ▁DRUG - ▁PICTURE - ▁MEDIA - ▁WAIT - ▁FRONT - ▁RISK - ▁BEHAVIOR - ▁BLACK - ▁100 - ▁NATURE - ▁ORGANIZATION - ▁HUNDRED - ▁EASY - ▁ACCESS - ▁HOLD - ▁COMMON - ▁MARKET - ▁GRAND - ▁VOICE - ▁DEATH - ▁PIECE - ▁BILLION - ▁LEAST - ▁DURING - '3' - ▁NATURAL - ▁TYPE - ▁INVEST - ▁GENERATION - ENCY - ▁STRONG - OLOGICAL - ▁CLEAR - ▁PRESENT - ▁INTERNET - ▁KILL - OLOGY - ▁SUPER - ▁UNITED - ▁IMAGE - ▁RATHER - ▁SOLUTION - ▁ECONOMIC - ▁PROTECT - ▁BEHIND - ▁COLLECT - ▁SCIENTIST - UDE - ▁PRODUCE - ▁PERFECT - ▁DOLLARS - ▁VIEW - ▁CONSIDER - ▁THIRD - ▁MACHINE - ▁OUTSIDE - ▁SKILL - ▁EXPERIMENT - ▁COLLEGE - ▁QUI - ▁OPPORTUNITY - ▁LOCAL - ▁SIMPLY - ▁EARLY - ▁MAJOR - ▁CANNOT - ▁PHYSICAL - ▁WHATEVER - ▁MIDDLE - ▁VIDEO - ▁ALONG - OGRAPH - ▁SOLVE - ▁KEY - ▁TRUST - ▁FIELD - HOOD - ▁ATTENTION - ▁MICRO - ▁SHORT - ▁SITUATION - ▁STREET - ▁COMPANIES - ▁POLITICAL - ▁NORMAL - ▁AMOUNT - ▁SERVICE - ▁OBJECT - ▁POTENTIAL - ▁COLOR - ▁KNOWLEDGE - ▁MORNING - ▁TRUTH - ▁UNIVERSITY - ▁PROVIDE - ▁RESOURCE - ▁POSITIVE - ▁EUROPE - ▁SPECIAL - ▁CONTINUE - ▁BASICALLY - ▁SMART - ▁PRACTICE - ▁POPULATION - ▁TRAVEL - ▁AFFECT - ▁FINALLY - ▁APPROACH - ▁COUNT - ▁PERHAPS - ▁INTERACT - ▁EXPLAIN - ▁ENGINEER - ▁ENGAGE - ▁SITTING - ▁OFFICE - ▁COMPLEX - ▁WHITE - ▁GENDER - ▁MESSAGE - ▁WORTH - ▁ITSELF - IZATION - ▁BUILT - ▁IMPROVE - ▁OKAY - ▁PRISON - ▁MATERIAL - ▁NETWORK - ▁EITHER - ▁GIVING - ▁LIMIT - ▁MEASURE - ▁DARK - ▁AUDIENCE - ▁ACCEPT - ▁RECORD - ▁OCEAN - ▁CHOOSE - ▁SPECIES - ▁YORK - ▁SUSTAIN - ▁SLEEP - ▁OBVIOUS - ▁HOSPITAL - ▁PERSPECTIVE - ▁INCREASE - ▁OPERA - ▁TAUGHT - ▁MULTI - ▁CHANGING - ▁JOURNEY - ▁INDUSTRY - ▁NEURO - ▁REQUIRE - ▁DECADE - ▁CURRENT - ▁PUSH - ▁BENEFIT - ▁YEAH - ▁BLOOD - ▁SCALE - ▁ESPECIALLY - ▁COMMUNITIES - ▁ADULT - ▁CHARACTER - ▁REPRESENT - IFIED - ▁SUFFER - ▁RECOGNIZE - ▁CENTURY - ▁SUDDEN - ▁FUNCTION - ▁ACHIEVE - ▁SIMILAR - ▁BROUGHT - ▁TRADITION - ▁UNIVERSE - ▁CLIMATE - ▁BREATH - ▁EXTREME - ▁REPORT - ▁DAUGHTER - ▁COMFORT - ▁CONCEPT - ▁ECONOMY - ▁INNOVATION - ▁QUICKLY - ▁SUGGEST - ▁SPECIFIC - ▁CRAZY - ▁CONSCIOUS - ▁SPREAD - ▁TRULY - '{' - init: xavier_uniform input_size: 2048 ctc_conf: dropout_rate: 0.0 ctc_type: builtin reduce: true ignore_nan_grad: null zero_infinity: true joint_net_conf: null use_preprocessor: true token_type: bpe bpemodel: data/en_token_list/bpe_unigram1000/bpe.model non_linguistic_symbols: null cleaner: null g2p: null speech_volume_normalize: null rir_scp: null rir_apply_prob: 1.0 noise_scp: null noise_apply_prob: 1.0 noise_db_range: '13_15' short_noise_thres: 0.5 aux_ctc_tasks: [] frontend: null frontend_conf: {} specaug: null specaug_conf: {} normalize: global_mvn normalize_conf: stats_file: exp/asr_stats_extracted_en_bpe1000/train/feats_stats.npz model: espnet model_conf: ctc_weight: 0.3 lsm_weight: 0.1 length_normalized_loss: false preencoder: null preencoder_conf: {} encoder: avhubert encoder_conf: avhubert_url: https://dl.fbaipublicfiles.com/avhubert/model/lrs3_vox/noise-pretrain/large_vox_iter5.pt avhubert_dir_path: ./local/pre-trained encoder_embed_dim: 1024 encoder_attention_heads: 16 encoder_ffn_embed_dim: 4096 encoder_layers: 24 dropout: 0.1 dropout_features: 0.1 encoder_layerdrop: 0.05 attention_dropout: 0.1 extracted: true freeze_finetune_updates: 10000 feature_grad_mult: 1.0 postencoder: null postencoder_conf: {} decoder: transformer decoder_conf: attention_heads: 4 linear_units: 4096 num_blocks: 6 dropout_rate: 0.1 positional_dropout_rate: 0.1 self_attention_dropout_rate: 0.1 src_attention_dropout_rate: 0.1 preprocessor: default preprocessor_conf: {} required: - output_dir - token_list version: '202308' distributed: true ```
### Citing ESPnet ```BibTex @inproceedings{watanabe2018espnet, author={Shinji Watanabe and Takaaki Hori and Shigeki Karita and Tomoki Hayashi and Jiro Nishitoba and Yuya Unno and Nelson Yalta and Jahn Heymann and Matthew Wiesner and Nanxin Chen and Adithya Renduchintala and Tsubasa Ochiai}, title={{ESPnet}: End-to-End Speech Processing Toolkit}, year={2018}, booktitle={Proceedings of Interspeech}, pages={2207--2211}, doi={10.21437/Interspeech.2018-1456}, url={http://dx.doi.org/10.21437/Interspeech.2018-1456} } ``` or arXiv: ```bibtex @misc{watanabe2018espnet, title={ESPnet: End-to-End Speech Processing Toolkit}, author={Shinji Watanabe and Takaaki Hori and Shigeki Karita and Tomoki Hayashi and Jiro Nishitoba and Yuya Unno and Nelson Yalta and Jahn Heymann and Matthew Wiesner and Nanxin Chen and Adithya Renduchintala and Tsubasa Ochiai}, year={2018}, eprint={1804.00015}, archivePrefix={arXiv}, primaryClass={cs.CL} } ```