{ "paper_id": "2020", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T01:26:48.653489Z" }, "title": "Multilingual Universal Sentence Encoder for Semantic Retrieval", "authors": [ { "first": "Yinfei", "middle": [], "last": "Yang", "suffix": "", "affiliation": { "laboratory": "", "institution": "CA Google Cambridge", "location": { "region": "MA" } }, "email": "yinfeiy@google.com" }, { "first": "Daniel", "middle": [], "last": "Cer", "suffix": "", "affiliation": { "laboratory": "", "institution": "CA Google Cambridge", "location": { "region": "MA" } }, "email": "cer@google.com" }, { "first": "Amin", "middle": [], "last": "Ahmad", "suffix": "", "affiliation": { "laboratory": "", "institution": "CA Google Cambridge", "location": { "region": "MA" } }, "email": "" }, { "first": "Mandy", "middle": [], "last": "Guo", "suffix": "", "affiliation": { "laboratory": "", "institution": "CA Google Cambridge", "location": { "region": "MA" } }, "email": "" }, { "first": "Jax", "middle": [], "last": "Law", "suffix": "", "affiliation": { "laboratory": "", "institution": "CA Google Cambridge", "location": { "region": "MA" } }, "email": "" }, { "first": "Noah", "middle": [], "last": "Constant", "suffix": "", "affiliation": { "laboratory": "", "institution": "CA Google Cambridge", "location": { "region": "MA" } }, "email": "" }, { "first": "Gustavo", "middle": [ "Hernandez" ], "last": "Abrego", "suffix": "", "affiliation": { "laboratory": "", "institution": "CA Google Cambridge", "location": { "region": "MA" } }, "email": "" }, { "first": "Steve", "middle": [], "last": "Yuan", "suffix": "", "affiliation": { "laboratory": "", "institution": "CA Google Cambridge", "location": { "region": "MA" } }, "email": "" }, { "first": "Chris", "middle": [], "last": "Tar", "suffix": "", "affiliation": { "laboratory": "", "institution": "CA Google Cambridge", "location": { "region": "MA" } }, "email": "" }, { "first": "Yun-Hsuan", "middle": [], "last": "Sung", "suffix": "", "affiliation": { "laboratory": "", "institution": "CA Google Cambridge", "location": { "region": "MA" } }, "email": "" }, { "first": "Brian", "middle": [], "last": "Strope", "suffix": "", "affiliation": { "laboratory": "", "institution": "CA Google Cambridge", "location": { "region": "MA" } }, "email": "" }, { "first": "Ray", "middle": [], "last": "Kurzweil", "suffix": "", "affiliation": { "laboratory": "", "institution": "CA Google Cambridge", "location": { "region": "MA" } }, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "We present easy-to-use retrieval focused multilingual sentence embedding models, made available on TensorFlow Hub. The models embed text from 16 languages into a shared semantic space using a multi-task trained dualencoder that learns tied cross-lingual representations via translation bridge tasks (Chidambaram et al., 2018). The models achieve a new state-of-the-art in performance on monolingual and cross-lingual semantic retrieval (SR). Competitive performance is obtained on the related tasks of translation pair bitext retrieval (BR) and retrieval question answering (ReQA). On transfer learning tasks, our multilingual embeddings approach, and in some cases exceed, the performance of English only sentence embeddings.", "pdf_parse": { "paper_id": "2020", "_pdf_hash": "", "abstract": [ { "text": "We present easy-to-use retrieval focused multilingual sentence embedding models, made available on TensorFlow Hub. The models embed text from 16 languages into a shared semantic space using a multi-task trained dualencoder that learns tied cross-lingual representations via translation bridge tasks (Chidambaram et al., 2018). The models achieve a new state-of-the-art in performance on monolingual and cross-lingual semantic retrieval (SR). Competitive performance is obtained on the related tasks of translation pair bitext retrieval (BR) and retrieval question answering (ReQA). On transfer learning tasks, our multilingual embeddings approach, and in some cases exceed, the performance of English only sentence embeddings.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "We introduce three new multilingual members in the universal sentence encoder (USE) family of sentence embedding models. The models target performance on tasks that involve multilingual semantic similarity and achieve a new state-of-the-art in performance on monolingual and cross-lingual semantic retrieval (SR). One model targets efficient resource usage with a CNN model architecture (Kim, 2014) . Another targets accuracy using the Transformer architecture (Vaswani et al., 2017 ). The third model provides an alternative interface to our multilingual Transformer model for use in retrieval question answering (ReQA). The 16 languages supported by our multilingual models are given in Table 1. 1 Table 1 : Multilingual universal sentence encoder's supported languages (ISO 639-1). Multilingual sentences are mapped to a shared semantic space.", "cite_spans": [ { "start": 387, "end": 398, "text": "(Kim, 2014)", "ref_id": "BIBREF16" }, { "start": 461, "end": 482, "text": "(Vaswani et al., 2017", "ref_id": "BIBREF22" } ], "ref_spans": [ { "start": 689, "end": 708, "text": "Table 1. 1 Table 1", "ref_id": null } ], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Our multilingual models are implemented in Ten-sorFlow (Abadi et al., 2016) and made publicly available on TensorFlow Hub. 2 Listing 1 illustrates the easy-to-use generation of multilingual sentence embeddings. The models conveniently only rely on TensorFlow without requiring additional libraries or packages. Listing 2 demonstrates using the question answering interface. Responses are encoded with additional context information such that the resulting context aware embeddings have a high dot product similarity score with the questions they answer. This allows for retrieval of indexed candidates using efficient nearest neighbor search. 3 3 Encoder Architecture", "cite_spans": [ { "start": 55, "end": 75, "text": "(Abadi et al., 2016)", "ref_id": null }, { "start": 643, "end": 644, "text": "3", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Model Toolkit", "sec_num": "2" }, { "text": "Similar to and Chidambaram et al. (2018) Figure 1 : Multilingual universal sentence encoder model training architecture using multi-task training over: (i) retrieval question-answering (ReQA), natural language inference (NLI) and translation ranking. Transformer or CNN based sentence embedding models provide a shared encoder across all tasks.", "cite_spans": [ { "start": 15, "end": 40, "text": "Chidambaram et al. (2018)", "ref_id": "BIBREF9" } ], "ref_spans": [ { "start": 41, "end": 49, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Multi-task Dual Encoder Training", "sec_num": "3.1" }, { "text": "import tensorflow_hub as hub module = hub.Module(\"https://tfhub.dev/google/\" \"universal-sentence-encoder-multilingual/1\") multilingual_embeddings = module([ \"Hola Mundo!\", \"Bonjour le monde!\", \"Ciao mondo!\" \"Hello World!\", \"Hallo Welt!\", \"Hallo Wereld!\", \"\u4f60\u597d\u4e16\u754c!\", \"\u041f\u0440\u0438\u0432\u0435\u0442, \u043c\u0438\u0440!\", \u202b\ufe91\u202aA\u202c\ufedf\ufecc\u202aA\u202c\ufedf\ufee2!\"\u202c Ab\u202b\ufea3\u202cr\u202b)]\"\ufee3\u202c Listing 1: Python code mapping multilingual sentences into a shared semantic embedding space. module = hub.Module(\"https://tfhub.dev/google/\" \"universal-sentence-encoder-multilingual-qa/1\") query_embeddings = module( dict(text=[\"What is your age?\"]), signature=\"question_encoder\", as_dict=True) candidate_embeddings = module( dict(text=[\"I am 20 years old.\"], context=[\"I will be 21 next year.\"]), signature=\"response_encoder\", as_dict=True)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Multi-task Dual Encoder Training", "sec_num": "3.1" }, { "text": "Listing 2: Python code embedding a question and answer for retrieval Question-Answering (ReQA). multi-task dual-encoder training framework, with a single shared encoder supporting multiple downstream tasks. The training tasks include: a multifeature question-answer prediction task, 4 a translation ranking task, and a natural language inference (NLI) task. Additional task specific hidden layers for the question-answering and NLI tasks are added after the shared encoder to provide representational specialization for each type of task. The model training architecture is illustrated at figure 1.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Multi-task Dual Encoder Training", "sec_num": "3.1" }, { "text": "4 Question-answer prediction is similar to conversationalresponse prediction . We treat the question as the conversational input and the answer as the response. For improved answer selection, we provide a bag-of-words (BoW) context feature as an additional input to the answer encoder. For our models, we use the entire paragraph containing the answer as context. The context feature is encoded using a separate DAN encoder.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Multi-task Dual Encoder Training", "sec_num": "3.1" }, { "text": "SentencePiece tokenization (Kudo and Richardson, 2018) is used for all of the 16 languages supported by our models. 5 A single 128k SentencePiece vocabulary is trained from 8 million sentences sampled from our training corpus and balanced across the 16 languages. For validation, the vocab is used to process a development set, separately sampled from the sentence encoding model training corpus. We find the development set character coverage is higher than 99% for all languages, with less than 1% out-of-vocabulary tokens. Each token in the vocab is mapped to a fixed length embedding vector. 6", "cite_spans": [ { "start": 27, "end": 54, "text": "(Kudo and Richardson, 2018)", "ref_id": "BIBREF17" } ], "ref_spans": [], "eq_spans": [], "section": "SentencePiece", "sec_num": "3.2" }, { "text": "Two distinct architectures for the sentence encoding models are provided: (i) transformer (Vaswani et al., 2017) , targeted at higher accuracy at the cost of resource consumption; (ii) convolutional neural network (CNN) (Kim, 2014) , designed for efficient inference but obtaining reduced accuracy.", "cite_spans": [ { "start": 90, "end": 112, "text": "(Vaswani et al., 2017)", "ref_id": "BIBREF22" }, { "start": 220, "end": 231, "text": "(Kim, 2014)", "ref_id": "BIBREF16" } ], "ref_spans": [], "eq_spans": [], "section": "Shared Encoder", "sec_num": "3.3" }, { "text": "Transformer The transformer encoding model embeds sentences using the encoder component of the transformer architecture (Vaswani et al., 2017) . Bi-directional self-attention is used to compute context-aware representations of tokens in a sentence, taking into account both the ordering and the identity of the tokens. The context-aware token representations are then averaged together to obtain a sentence-level embedding.", "cite_spans": [ { "start": 120, "end": 142, "text": "(Vaswani et al., 2017)", "ref_id": "BIBREF22" } ], "ref_spans": [], "eq_spans": [], "section": "Shared Encoder", "sec_num": "3.3" }, { "text": "CNN The CNN sentence encoding model feeds the input token sequence embeddings into a con- volutional neural network (Kim, 2014) . Similar to the transformer encoder, average pooling is used to turn the token-level embeddings into a fixedlength representation. Sentence embeddings are then obtained by passing the averaged representation through additional feedforward layers. We also translate a portion of question-answer pairs to ensure each language has a minimum of 60M training pairs. For each of our datasets, we use 90% of the data for training, and the remaining 10% for development/validation. Table 2 and 3 lists the details of data used for each task / langauge.", "cite_spans": [ { "start": 116, "end": 127, "text": "(Kim, 2014)", "ref_id": "BIBREF16" } ], "ref_spans": [ { "start": 603, "end": 610, "text": "Table 2", "ref_id": "TABREF2" } ], "eq_spans": [], "section": "Shared Encoder", "sec_num": "3.3" }, { "text": "Input sentences are truncated to 256 tokens for the CNN model and 100 tokens for transformer. The CNN encoder uses 2 CNN layers with filter width of [1, 2, 3, 5] and 256 filters per width. The Transformer encoder employs 6 transformer layers, with 8 attentions heads, hidden size 512, and filter size 2048. Similar to our prior work , we configure our models with the intention of making them small and fast enough to be used directly within many downstream applications without the need for model distillation. Model hyperparameters are tuned on development data sampled from the same sources as the training data. We export sentence encoding modules for our two encoder architectures: USE Trans and USE CNN . We also export a larger graph for QA tasks from our Transformer based model that includes QA specific layers and support providing context information from the larger document as USE QA Trans+Cxt . 10", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Model Configuration", "sec_num": "4.2" }, { "text": "In this section we evaluate our multilingual encoding models on semantic retrieval, bitext and Yang et al. (2019) 89.0 86.1 89.2 87.9 Table 5 : P@1 on UN translation pair bitext retrieval (BR). Yang et al. 2019is a specialized translation retrieval model and the current state-of-the-art.", "cite_spans": [ { "start": 95, "end": 113, "text": "Yang et al. (2019)", "ref_id": "BIBREF24" } ], "ref_spans": [ { "start": 134, "end": 141, "text": "Table 5", "ref_id": null } ], "eq_spans": [], "section": "Experiments on Retrieval Tasks", "sec_num": "5" }, { "text": "retrieval question answer tasks.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experiments on Retrieval Tasks", "sec_num": "5" }, { "text": "Following Gillick et al. 2018, we construct semantic retrieval (SR) tasks from the Quora questionpairs (Hoogeveen et al., 2015) and AskUbuntu (Lei et al., 2016) datasets. The SR task is to identify all sentences in the retrieval corpus that are semantically similar to a query sentence. 11 For each dataset, we first build a graph connecting each of the positive pairs, and then compute its transitive closure. Each sentence then serves as a test query that should retrieve all of the other sentences it is connected to within the transitive closure. Mean average precision (MAP) is employed to evaluate the models. More details on the constructed datasets can be found in Gillick et al. (2018) . Both datasets are English only. Table 4 shows the MAP@100 on the Quora and AskUbuntu retrieval tasks. We use Gillick et al. (2018) as the baseline model, which is trained using a similar dual encoder architecture. The numbers provided here are for models without focused in-domain training data. 12 Both USE CNN and USE Trans outperform the prior state-of-theart. USE Trans and USE CNN perform comparably on Quora. However, USE Trans performs notably better than USE CNN on AskUbuntu, suggesting the AskUbuntu data could be more challenging.", "cite_spans": [ { "start": 103, "end": 127, "text": "(Hoogeveen et al., 2015)", "ref_id": "BIBREF14" }, { "start": 142, "end": 160, "text": "(Lei et al., 2016)", "ref_id": null }, { "start": 287, "end": 289, "text": "11", "ref_id": null }, { "start": 673, "end": 694, "text": "Gillick et al. (2018)", "ref_id": "BIBREF13" }, { "start": 806, "end": 827, "text": "Gillick et al. (2018)", "ref_id": "BIBREF13" } ], "ref_spans": [ { "start": 729, "end": 736, "text": "Table 4", "ref_id": "TABREF5" } ], "eq_spans": [], "section": "Semantic Retrieval (SR)", "sec_num": "5.1" }, { "text": "Bitext retrieval performance is evaluated on the United Nation (UN) Parallel Corpus (Ziemski et al., 2016) , containing 86,000 bilingual document pairs matching English (en) documents with with their translations in five other languages: French (fr), 11 The task is related to paraphrase identification (Dolan et al., 2004) and Semantic Textual Similarity (STS) (Cer et al., 2017) , but with the identification of meaning similarity being assessed in the context of a retrieval task.", "cite_spans": [ { "start": 84, "end": 106, "text": "(Ziemski et al., 2016)", "ref_id": "BIBREF26" }, { "start": 251, "end": 253, "text": "11", "ref_id": null }, { "start": 303, "end": 323, "text": "(Dolan et al., 2004)", "ref_id": "BIBREF12" }, { "start": 362, "end": 380, "text": "(Cer et al., 2017)", "ref_id": "BIBREF6" } ], "ref_spans": [], "eq_spans": [], "section": "Bitext Retrieval (BR)", "sec_num": "5.2" }, { "text": "12 The model for Quora is trained on Paralex (http: //knowitall.cs.washington.edu/paralex) and AskUbuntu data. The model for AskUbuntu is trained on Paralex and Quora. Spanish (es), Russian (ru), Arabic (ar) and Chinese (zh). Document pairs are aligned at the sentencelevel, which results in 11.3 million aligned sentence pairs for each language pair. Table 5 shows sentence-level retrieval precision@1 (P@1) for the proposed models as well as the current state-of-the-art results from Yang et al. (2019) , which uses a specialized translation pair retrieval model. USE Trans is generally better than USE CNN , performing lower than the SOTA but not by too much with the exception of en-zh. 13 ", "cite_spans": [ { "start": 486, "end": 504, "text": "Yang et al. (2019)", "ref_id": "BIBREF24" } ], "ref_spans": [ { "start": 352, "end": 359, "text": "Table 5", "ref_id": null } ], "eq_spans": [], "section": "Bitext Retrieval (BR)", "sec_num": "5.2" }, { "text": "Similar to the data set construction used for the SR tasks, the SQuAD v1.0 dataset (Rajpurkar et al., 2016 ) is transformed into a retrieval question answering (ReQA) task. 14 We first break all documents in the dataset into sentences using the sentence splitter distributed with the ReQA evaluation suite. 15 Each question of the (question, answer spans) tuples in the dataset is treated as a query. The task is to retrieve the sentence designated by the tuple answer span. Search is performed on a retrieval corpus consisting of all of the sentences within the corpus. We contrast sentence and paragraph-level retrieval using our models, with the later allowing for comparison against a BM25 baseline (Jones et al., 2000) . 16", "cite_spans": [ { "start": 83, "end": 106, "text": "(Rajpurkar et al., 2016", "ref_id": "BIBREF20" }, { "start": 173, "end": 175, "text": "14", "ref_id": null }, { "start": 307, "end": 309, "text": "15", "ref_id": null }, { "start": 703, "end": 723, "text": "(Jones et al., 2000)", "ref_id": "BIBREF15" } ], "ref_spans": [], "eq_spans": [], "section": "Retrieval Question Answering (ReQA)", "sec_num": "5.3" }, { "text": "13 Performance is degraded from Yang et al. (2019) due to using a single sentencepiece vocabulary to cover 16 languages. Languages like Chinese, Korean, Japanese have much more characters. To ensure the vocab coverage, sentencepiece tends to split the text of these languages into single characters, which increases the difficulty of the task.", "cite_spans": [ { "start": 32, "end": 50, "text": "Yang et al. (2019)", "ref_id": "BIBREF24" } ], "ref_spans": [], "eq_spans": [], "section": "Retrieval Question Answering (ReQA)", "sec_num": "5.3" }, { "text": "14 The retrieval question answering task was suggested by Chen et al. (2017) and then recently explored further by Cakaloglu et al. (2018) . However, Cakaloglu et al. (2018) 's use of sampling makes it difficult to directly compare with their results and we provide our own baseline based on BM25.", "cite_spans": [ { "start": 58, "end": 76, "text": "Chen et al. (2017)", "ref_id": "BIBREF8" }, { "start": 115, "end": 138, "text": "Cakaloglu et al. (2018)", "ref_id": "BIBREF5" }, { "start": 150, "end": 173, "text": "Cakaloglu et al. (2018)", "ref_id": "BIBREF5" } ], "ref_spans": [], "eq_spans": [], "section": "Retrieval Question Answering (ReQA)", "sec_num": "5.3" }, { "text": "15 https://github.com/google/ retrieval-qa-eval 16 BM25 is a strong baseline for text retrieval tasks. Paragraph-level experiments use the BM25 implementa- Table 7 : Cross-lingual performance on Quora/AskUbuntu cl-SR (MAP) and SQuAD cl-ReQA (P@1). Queries/questions are machine translated, while retrieval candidates remain in English.", "cite_spans": [], "ref_spans": [ { "start": 156, "end": 163, "text": "Table 7", "ref_id": null } ], "eq_spans": [], "section": "Retrieval Question Answering (ReQA)", "sec_num": "5.3" }, { "text": "We evaluated ReQA using the SQuAD dev and train sets and without training on the SQuAD data. 17 The sentence and paragraph retrieval P@1 are shown in table 6. For sentence retrieval, we compare encodings produced using context from the text surrounding the retrieval candidate, USE QA Trans+Cxt , to sentence encodings produced without contextual cues, USE Trans . Paragraph retrieval contrasts USE QA Trans+Cxt with BM25.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Retrieval Question Answering (ReQA)", "sec_num": "5.3" }, { "text": "Our English retrieval experiments are extended to explore cross-lingual semantic retrieval (cl-SR) and cross-lingual retrieval question answering (cl-ReQA). SR queries and ReQA questions are machine translated into other languages, while keeping the retrieval candidates in English. 18 Table 7 provides our cross-lingual retrieval results for our transformer and CNN multilingual sentence encoding models. We compare against the state-ofthe-art LASER multilingual sentence embedding tion: https://github.com/nhirakawa/BM25, with default parameters. We exclude sentence-level BM25, as BM25 generally performs poorly at this granularity.", "cite_spans": [ { "start": 283, "end": 285, "text": "18", "ref_id": null } ], "ref_spans": [ { "start": 286, "end": 294, "text": "Table 7", "ref_id": null } ], "eq_spans": [], "section": "Cross-lingual Retrieval", "sec_num": "5.4" }, { "text": "17 For sentences, the resulting retrieval task for development set consists of 11,425 questions and 10,248 candidates, and the retrieval task for train set is consists of 87,599 questions and 91,703 candidates. For paragraph retrieval, there are 2,067 retrieval candidates in the development set and 18,896 in the training set. To retrieve paragraphs with our model, we first run sentence retrieval and use the retrieved nearest sentence to select the enclosing paragraph.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Cross-lingual Retrieval", "sec_num": "5.4" }, { "text": "18 Poor translations are detected and rejected when the original English text and English back translation have a cosine similarity < 0.5 according our previously released English USETrans model .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Cross-lingual Retrieval", "sec_num": "5.4" }, { "text": "library (Artetxe and Schwenk, 2019) . 19 On both the Quora and AskUbuntu cl-SR tasks, USE Trans outperforms USE CNN and LASER on all datasets, except the Polish (pl) Quora data where LASER achieves slightly better performance. 20 USE CNN tends to outperform LASER on Quora and always outperforms LASER by a sizable margin on AskUbuntu. We note that our CNN based model not only outperforms LASER, but also relies on simpler model architecture than LASER's LSTM based archtitecture. Given the similar level of performance on Quora between USE CNN and LASER, we suspect the notably better performance on AskUbuntu over LASER is due to differences in the training data provided to encoding models.", "cite_spans": [ { "start": 8, "end": 35, "text": "(Artetxe and Schwenk, 2019)", "ref_id": "BIBREF2" }, { "start": 38, "end": 40, "text": "19", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Cross-lingual Retrieval", "sec_num": "5.4" }, { "text": "For comparison with prior USE models, English task transfer performance is evaluated on SentEval (Conneau and Kiela, 2018) . For sentence classification transfer tasks, the output of the sentence encoders are provided to a task specific DNN. For the pairwise semantic similarity task, the similarity of sentence embeddings and is assessed using \u2212 arccos 72.2 78.5 92.1 86.9 88.1 77.5 0.760 / 0.717 USETransformer for English 82 , on some tasks.", "cite_spans": [ { "start": 97, "end": 122, "text": "(Conneau and Kiela, 2018)", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "Experiments on Transfer Tasks", "sec_num": "6" }, { "text": "Figure 2 provides compute and memory usage benchmarks for our models. 21 Inference times on GPU are 2 to 3 times faster than CPU. Our CNN models have the smallest memory footprint and are the fastest on both CPU and GPU. The memory requirements increase with sentence length, with the Transformer model increasing more than twice as fast as the CNN model. 22 While this makes CNNs an attractive choice for efficiently encoding longer texts, this comes with a corresponding drop in accuracy on many retrieval and transfer tasks.", "cite_spans": [ { "start": 356, "end": 358, "text": "22", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Resource Usage", "sec_num": "7" }, { "text": "Easy-to-use retrieval focused multilingual models for embedding sentence-length text are made avail- 21 CPU benchmarks are run on Intel(R) Xeon(R) Platinum 8173M CPU @ 2.00GHz. GPU benchmarks were run on an NVidia v100. Memory footprint was measured on CPU.", "cite_spans": [ { "start": 101, "end": 103, "text": "21", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "8" }, { "text": "22 Transformer models are ultimately governed by a time and space complexity of ( 2 ). The benchmarks show for shorter sequence lengths the time and space requirements are dominated by computations that scale linearly with length and have a larger constant factor than the quadratic terms. able on TensorFlow Hub. Our models embed text from 16 languages into a shared semantic embedding space and achieve a new state-of-the-art in performance on monolingual and cross-lingual semantic retrieval (SR). The models achieve good performance on the related tasks of translation pair bitext retrieval (BR) and retrieval question answering (ReQA). Monolingual transfer task performance approaches, and in some cases exceeds, English only sentence embedding models. Our models are freely available under an Apache license with additional documentation and tutorial colaboratory notebooks at:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "8" }, { "text": "https://tfhub.dev/s?q=universalsentence-encoder-multilingual", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "8" }, { "text": "https://www.tensorflow.org/hub/, Apache 2.0 license, with models available as saved TF graphs.3 Popular efficient search tools include FAISS https: //github.com/facebookresearch/faiss, Annoy https://github.com/spotify/annoy, or FLANN https://www.cs.ubc.ca/research/flann.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "https://github.com/google/ sentencepiece 6 Out-of-vocabulary characters map to an token.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "QA pairs are mined from online forums and QA websites, including Reddit, StackOverflow, and YahooAnswers.8 The translation pairs are mined using a system similar to the approach described inUszkoreit et al. (2010).9 MultiNLI(Williams et al., 2018), a more extensive corpus, contains examples from multiple sources but with different licences. Employing SNLI avoids navigating the licensing complexity of using MultiNLI to training public models.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "While USEQA Trans+Cxt uses the same underlying shared encoder as USETrans but with additional task specific layers, we anticipate that the models could diverge in the future.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "https://github.com/facebookresearch/ LASER 20 Results are not presented for LASER on ja and zh due unicode errors.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [ { "text": "We thank our teammates from Descartes and other Google groups for their feedback and suggestions. Special thanks goes to Muthu Chidambaram for his early exploration of multilingual training, Taku Kudo for the SentencePiece model support, Chen Chen for the templates used to perform the transfer learning experiments and Mario Guajardo for an early version of the ReQA tutorial Colab.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgments", "sec_num": null } ], "bib_entries": { "BIBREF1": { "ref_id": "b1", "title": "Tensorflow: A system for large-scale machine learning", "authors": [ { "first": "Benoit", "middle": [], "last": "Murray", "suffix": "" }, { "first": "Paul", "middle": [], "last": "Steiner", "suffix": "" }, { "first": "Vijay", "middle": [], "last": "Tucker", "suffix": "" }, { "first": "Pete", "middle": [], "last": "Vasudevan", "suffix": "" }, { "first": "Martin", "middle": [], "last": "Warden", "suffix": "" }, { "first": "Yuan", "middle": [], "last": "Wicke", "suffix": "" }, { "first": "Xiaoqiang", "middle": [], "last": "Yu", "suffix": "" }, { "first": "", "middle": [], "last": "Zheng", "suffix": "" } ], "year": 2016, "venue": "Proceedings of USENIX OSDI'16, OSDI'16", "volume": "", "issue": "", "pages": "265--283", "other_ids": {}, "num": null, "urls": [], "raw_text": "Murray, Benoit Steiner, Paul Tucker, Vijay Vasudevan, Pete Warden, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. 2016. Tensorflow: A system for large-scale machine learn- ing. In Proceedings of USENIX OSDI'16, OSDI'16, pages 265-283.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Massively multilingual sentence embeddings for zero-shot cross-lingual transfer and beyond", "authors": [ { "first": "Mikel", "middle": [], "last": "Artetxe", "suffix": "" }, { "first": "Holger", "middle": [], "last": "Schwenk", "suffix": "" } ], "year": 2019, "venue": "Transactions of the Association for Computational Linguistics", "volume": "7", "issue": "", "pages": "597--610", "other_ids": { "DOI": [ "10.1162/tacl_a_00288" ] }, "num": null, "urls": [], "raw_text": "Mikel Artetxe and Holger Schwenk. 2019. Massively mul- tilingual sentence embeddings for zero-shot cross-lingual transfer and beyond. Transactions of the Association for Computational Linguistics, 7:597-610.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "A large annotated corpus for learning natural language inference", "authors": [ { "first": "R", "middle": [], "last": "Samuel", "suffix": "" }, { "first": "Gabor", "middle": [], "last": "Bowman", "suffix": "" }, { "first": "Christopher", "middle": [], "last": "Angeli", "suffix": "" }, { "first": "Christopher", "middle": [ "D" ], "last": "Potts", "suffix": "" }, { "first": "", "middle": [], "last": "Manning", "suffix": "" } ], "year": 2015, "venue": "Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "632--642", "other_ids": { "DOI": [ "10.18653/v1/D15-1075" ] }, "num": null, "urls": [], "raw_text": "Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632-642.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Text embeddings for retrieval from a large knowl", "authors": [ { "first": "Tolgahan", "middle": [], "last": "Cakaloglu", "suffix": "" }, { "first": "Christian", "middle": [], "last": "Szegedy", "suffix": "" }, { "first": "Xiaowei", "middle": [], "last": "Xu", "suffix": "" } ], "year": 2018, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Tolgahan Cakaloglu, Christian Szegedy, and Xiaowei Xu. 2018. Text embeddings for retrieval from a large knowl- edge base. CoRR, abs/1810.10176.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "SemEval-2017 task 1: Semantic textual similarity multilingual and crosslingual focused evaluation", "authors": [ { "first": "Daniel", "middle": [], "last": "Cer", "suffix": "" }, { "first": "Mona", "middle": [], "last": "Diab", "suffix": "" }, { "first": "Eneko", "middle": [], "last": "Agirre", "suffix": "" }, { "first": "I\u00f1igo", "middle": [], "last": "Lopez-Gazpio", "suffix": "" }, { "first": "Lucia", "middle": [], "last": "Specia", "suffix": "" } ], "year": 2017, "venue": "Proceedings of the 11th International Workshop on Semantic Evaluation", "volume": "", "issue": "", "pages": "1--14", "other_ids": { "DOI": [ "10.18653/v1/S17-2001" ] }, "num": null, "urls": [], "raw_text": "Daniel Cer, Mona Diab, Eneko Agirre, I\u00f1igo Lopez-Gazpio, and Lucia Specia. 2017. SemEval-2017 task 1: Seman- tic textual similarity multilingual and crosslingual focused evaluation. In Proceedings of the 11th International Work- shop on Semantic Evaluation (SemEval-2017), pages 1-14.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Universal sentence encoder for English", "authors": [ { "first": "Daniel", "middle": [], "last": "Cer", "suffix": "" }, { "first": "Yinfei", "middle": [], "last": "Yang", "suffix": "" }, { "first": "Sheng-Yi", "middle": [], "last": "Kong", "suffix": "" }, { "first": "Nan", "middle": [], "last": "Hua", "suffix": "" }, { "first": "Nicole", "middle": [], "last": "Limtiaco", "suffix": "" }, { "first": "Rhomni", "middle": [], "last": "St", "suffix": "" }, { "first": "Noah", "middle": [], "last": "John", "suffix": "" }, { "first": "Mario", "middle": [], "last": "Constant", "suffix": "" }, { "first": "Steve", "middle": [], "last": "Guajardo-Cespedes", "suffix": "" }, { "first": "Chris", "middle": [], "last": "Yuan", "suffix": "" }, { "first": "Brian", "middle": [], "last": "Tar", "suffix": "" }, { "first": "Ray", "middle": [], "last": "Strope", "suffix": "" }, { "first": "", "middle": [], "last": "Kurzweil", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations", "volume": "", "issue": "", "pages": "169--174", "other_ids": {}, "num": null, "urls": [], "raw_text": "Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, Rhomni St. John, Noah Constant, Mario Guajardo-Cespedes, Steve Yuan, Chris Tar, Brian Strope, and Ray Kurzweil. 2018. Universal sentence encoder for English. In Proceedings of the 2018 Conference on Em- pirical Methods in Natural Language Processing: System Demonstrations, pages 169-174.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Reading Wikipedia to answer open-domain questions", "authors": [ { "first": "Danqi", "middle": [], "last": "Chen", "suffix": "" }, { "first": "Adam", "middle": [], "last": "Fisch", "suffix": "" }, { "first": "Jason", "middle": [], "last": "Weston", "suffix": "" }, { "first": "Antoine", "middle": [], "last": "Bordes", "suffix": "" } ], "year": 2017, "venue": "Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics", "volume": "1", "issue": "", "pages": "1870--1879", "other_ids": { "DOI": [ "10.18653/v1/P17-1171" ] }, "num": null, "urls": [], "raw_text": "Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bor- des. 2017. Reading Wikipedia to answer open-domain questions. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1870-1879.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Learning cross-lingual sentence representations via a multi-task dual-encoder model", "authors": [ { "first": "Muthuraman", "middle": [], "last": "Chidambaram", "suffix": "" }, { "first": "Yinfei", "middle": [], "last": "Yang", "suffix": "" }, { "first": "Daniel", "middle": [], "last": "Cer", "suffix": "" }, { "first": "Steve", "middle": [], "last": "Yuan", "suffix": "" }, { "first": "Yun-Hsuan", "middle": [], "last": "Sung", "suffix": "" }, { "first": "Brian", "middle": [], "last": "Strope", "suffix": "" }, { "first": "Ray", "middle": [], "last": "Kurzweil", "suffix": "" } ], "year": 2018, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Muthuraman Chidambaram, Yinfei Yang, Daniel Cer, Steve Yuan, Yun-Hsuan Sung, Brian Strope, and Ray Kurzweil. 2018. Learning cross-lingual sentence representations via a multi-task dual-encoder model. CoRR, abs/1810.12836.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "SentEval: An evaluation toolkit for universal sentence representations", "authors": [ { "first": "Alexis", "middle": [], "last": "Conneau", "suffix": "" }, { "first": "Douwe", "middle": [], "last": "Kiela", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the Eleventh International Conference on Language Resources and Evaluation", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Alexis Conneau and Douwe Kiela. 2018. SentEval: An eval- uation toolkit for universal sentence representations. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC-2018).", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Supervised learning of universal sentence representations from natural language inference data", "authors": [ { "first": "Alexis", "middle": [], "last": "Conneau", "suffix": "" }, { "first": "Douwe", "middle": [], "last": "Kiela", "suffix": "" }, { "first": "Holger", "middle": [], "last": "Schwenk", "suffix": "" } ], "year": 2017, "venue": "Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "670--680", "other_ids": { "DOI": [ "10.18653/v1/D17-1070" ] }, "num": null, "urls": [], "raw_text": "Alexis Conneau, Douwe Kiela, Holger Schwenk, Lo\u00efc Bar- rault, and Antoine Bordes. 2017. Supervised learning of universal sentence representations from natural language inference data. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 670-680, Copenhagen, Denmark. Association for Computational Linguistics.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Unsupervised construction of large paraphrase corpora: Exploiting massively parallel news sources", "authors": [ { "first": "Bill", "middle": [], "last": "Dolan", "suffix": "" }, { "first": "Chris", "middle": [], "last": "Quirk", "suffix": "" }, { "first": "Chris", "middle": [], "last": "Brockett", "suffix": "" } ], "year": 2004, "venue": "COLING 2004: Proceedings of the 20th International Conference on Computational Linguistics", "volume": "", "issue": "", "pages": "350--356", "other_ids": {}, "num": null, "urls": [], "raw_text": "Bill Dolan, Chris Quirk, and Chris Brockett. 2004. Unsuper- vised construction of large paraphrase corpora: Exploiting massively parallel news sources. In COLING 2004: Pro- ceedings of the 20th International Conference on Compu- tational Linguistics, pages 350-356.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "End-to-end retrieval in continuous space", "authors": [ { "first": "Daniel", "middle": [], "last": "Gillick", "suffix": "" }, { "first": "Alessandro", "middle": [], "last": "Presta", "suffix": "" }, { "first": "Gaurav Singh", "middle": [], "last": "Tomar", "suffix": "" } ], "year": 2018, "venue": "CoRR", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Daniel Gillick, Alessandro Presta, and Gaurav Singh Tomar. 2018. End-to-end retrieval in continuous space. CoRR, abs/1811.08008.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Cqadupstack: A benchmark data set for community question-answering research", "authors": [ { "first": "Doris", "middle": [], "last": "Hoogeveen", "suffix": "" }, { "first": "Karin", "middle": [ "M" ], "last": "Verspoor", "suffix": "" }, { "first": "Timothy", "middle": [], "last": "Baldwin", "suffix": "" } ], "year": 2015, "venue": "Proceedings of the 20th Australasian Document Computing Symposium, ADCS '15", "volume": "3", "issue": "", "pages": "1--3", "other_ids": { "DOI": [ "10.1145/2838931.2838934" ] }, "num": null, "urls": [], "raw_text": "Doris Hoogeveen, Karin M. Verspoor, and Timothy Baldwin. 2015. Cqadupstack: A benchmark data set for community question-answering research. In Proceedings of the 20th Australasian Document Computing Symposium, ADCS '15, pages 3:1-3:8.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "A probabilistic model of information retrieval: Development and comparative experiments", "authors": [ { "first": "K", "middle": [], "last": "Jones", "suffix": "" }, { "first": "S", "middle": [], "last": "Walker", "suffix": "" }, { "first": "S", "middle": [ "E" ], "last": "Robertson", "suffix": "" } ], "year": 2000, "venue": "Inf. Process. Manage", "volume": "36", "issue": "6", "pages": "779--808", "other_ids": { "DOI": [ "10.1016/S0306-4573(00)00015-7" ] }, "num": null, "urls": [], "raw_text": "K. Sparck Jones, S. Walker, and S. E. Robertson. 2000. A probabilistic model of information retrieval: Develop- ment and comparative experiments. Inf. Process. Manage., 36(6):779-808.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Convolutional neural networks for sentence classification", "authors": [ { "first": "Yoon", "middle": [], "last": "Kim", "suffix": "" } ], "year": 2014, "venue": "Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)", "volume": "", "issue": "", "pages": "1746--1751", "other_ids": { "DOI": [ "10.3115/v1/D14-1181" ] }, "num": null, "urls": [], "raw_text": "Yoon Kim. 2014. Convolutional neural networks for sentence classification. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1746-1751.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "SentencePiece: A simple and language independent subword tokenizer and detokenizer for neural text processing", "authors": [ { "first": "Taku", "middle": [], "last": "Kudo", "suffix": "" }, { "first": "John", "middle": [], "last": "Richardson", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations", "volume": "", "issue": "", "pages": "66--71", "other_ids": {}, "num": null, "urls": [], "raw_text": "Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66- 71.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Semi-supervised question retrieval with gated convolutions", "authors": [ { "first": "Tao", "middle": [], "last": "Lei", "suffix": "" }, { "first": "Hrishikesh", "middle": [], "last": "Joshi", "suffix": "" }, { "first": "Regina", "middle": [], "last": "Barzilay", "suffix": "" }, { "first": "Tommi", "middle": [], "last": "Jaakkola", "suffix": "" }, { "first": "Kateryna", "middle": [], "last": "Tymoshenko", "suffix": "" }, { "first": "Alessandro", "middle": [], "last": "Moschitti", "suffix": "" }, { "first": "Llu\u00eds", "middle": [], "last": "M\u00e0rquez", "suffix": "" } ], "year": 2016, "venue": "Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies", "volume": "", "issue": "", "pages": "1279--1289", "other_ids": { "DOI": [ "10.18653/v1/N16-1153" ] }, "num": null, "urls": [], "raw_text": "Tao Lei, Hrishikesh Joshi, Regina Barzilay, Tommi Jaakkola, Kateryna Tymoshenko, Alessandro Moschitti, and Llu\u00eds M\u00e0rquez. 2016. Semi-supervised question retrieval with gated convolutions. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Com- putational Linguistics: Human Language Technologies, pages 1279-1289.", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "An efficient framework for learning sentence representations", "authors": [ { "first": "Lajanugen", "middle": [], "last": "Logeswaran", "suffix": "" }, { "first": "Honglak", "middle": [], "last": "Lee", "suffix": "" } ], "year": 2018, "venue": "International Conference on Learning Representations (ICLR", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lajanugen Logeswaran and Honglak Lee. 2018. An efficient framework for learning sentence representations. In Inter- national Conference on Learning Representations (ICLR).", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "SQuAD: 100,000+ questions for machine comprehension of text", "authors": [ { "first": "Pranav", "middle": [], "last": "Rajpurkar", "suffix": "" }, { "first": "Jian", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "Konstantin", "middle": [], "last": "Lopyrev", "suffix": "" }, { "first": "Percy", "middle": [], "last": "Liang", "suffix": "" } ], "year": 2016, "venue": "Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "2383--2392", "other_ids": { "DOI": [ "10.18653/v1/D16-1264" ] }, "num": null, "urls": [], "raw_text": "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Confer- ence on Empirical Methods in Natural Language Process- ing, pages 2383-2392.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "Large scale parallel document mining for machine translation", "authors": [ { "first": "Jakob", "middle": [], "last": "Uszkoreit", "suffix": "" }, { "first": "Jay", "middle": [], "last": "Ponte", "suffix": "" }, { "first": "Ashok", "middle": [], "last": "Popat", "suffix": "" }, { "first": "Moshe", "middle": [], "last": "Dubiner", "suffix": "" } ], "year": 2010, "venue": "Proceedings of the 23rd International Conference on Computational Linguistics", "volume": "", "issue": "", "pages": "1101--1109", "other_ids": {}, "num": null, "urls": [], "raw_text": "Jakob Uszkoreit, Jay Ponte, Ashok Popat, and Moshe Dubiner. 2010. Large scale parallel document mining for machine translation. In Proceedings of the 23rd International Con- ference on Computational Linguistics (Coling 2010), pages 1101-1109.", "links": null }, "BIBREF22": { "ref_id": "b22", "title": "Attention is all you need", "authors": [ { "first": "Ashish", "middle": [], "last": "Vaswani", "suffix": "" }, { "first": "Noam", "middle": [], "last": "Shazeer", "suffix": "" }, { "first": "Niki", "middle": [], "last": "Parmar", "suffix": "" }, { "first": "Jakob", "middle": [], "last": "Uszkoreit", "suffix": "" }, { "first": "Llion", "middle": [], "last": "Jones", "suffix": "" }, { "first": "Aidan", "middle": [ "N" ], "last": "Gomez", "suffix": "" }, { "first": "Illia", "middle": [], "last": "Kaiser", "suffix": "" }, { "first": "", "middle": [], "last": "Polosukhin", "suffix": "" } ], "year": 2017, "venue": "Proceedings of NIPS", "volume": "", "issue": "", "pages": "6000--6010", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkor- eit, Llion Jones, Aidan N Gomez, \u0141 ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Proceed- ings of NIPS, pages 6000-6010.", "links": null }, "BIBREF23": { "ref_id": "b23", "title": "A broad-coverage challenge corpus for sentence understanding through inference", "authors": [ { "first": "Adina", "middle": [], "last": "Williams", "suffix": "" }, { "first": "Nikita", "middle": [], "last": "Nangia", "suffix": "" }, { "first": "Samuel", "middle": [], "last": "Bowman", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies", "volume": "1", "issue": "", "pages": "1112--1122", "other_ids": { "DOI": [ "10.18653/v1/N18-1101" ] }, "num": null, "urls": [], "raw_text": "Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence under- standing through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Asso- ciation for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112-1122.", "links": null }, "BIBREF24": { "ref_id": "b24", "title": "Improving multilingual sentence embedding using bi-directional dual encoder with additive margin softmax", "authors": [ { "first": "Yinfei", "middle": [], "last": "Yang", "suffix": "" }, { "first": "Gustavo", "middle": [ "Hern\u00e1ndez" ], "last": "\u00c1brego", "suffix": "" }, { "first": "Steve", "middle": [], "last": "Yuan", "suffix": "" }, { "first": "Mandy", "middle": [], "last": "Guo", "suffix": "" }, { "first": "Qinlan", "middle": [], "last": "Shen", "suffix": "" }, { "first": "Daniel", "middle": [], "last": "Cer", "suffix": "" }, { "first": "Yun-Hsuan", "middle": [], "last": "Sung", "suffix": "" }, { "first": "Brian", "middle": [], "last": "Strope", "suffix": "" }, { "first": "Ray", "middle": [], "last": "Kurzweil", "suffix": "" } ], "year": 2019, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yinfei Yang, Gustavo Hern\u00e1ndez \u00c1brego, Steve Yuan, Mandy Guo, Qinlan Shen, Daniel Cer, Yun-Hsuan Sung, Brian Strope, and Ray Kurzweil. 2019. Improving multilingual sentence embedding using bi-directional dual encoder with additive margin softmax. CoRR, abs/1902.08564.", "links": null }, "BIBREF25": { "ref_id": "b25", "title": "Learning semantic textual similarity from conversations", "authors": [ { "first": "Yinfei", "middle": [], "last": "Yang", "suffix": "" }, { "first": "Steve", "middle": [], "last": "Yuan", "suffix": "" }, { "first": "Daniel", "middle": [], "last": "Cer", "suffix": "" }, { "first": "Sheng-Yi", "middle": [], "last": "Kong", "suffix": "" }, { "first": "Noah", "middle": [], "last": "Constant", "suffix": "" }, { "first": "Petr", "middle": [], "last": "Pilar", "suffix": "" }, { "first": "Heming", "middle": [], "last": "Ge", "suffix": "" }, { "first": "Yun-Hsuan", "middle": [], "last": "Sung", "suffix": "" }, { "first": "Brian", "middle": [], "last": "Strope", "suffix": "" }, { "first": "Ray", "middle": [], "last": "Kurzweil", "suffix": "" } ], "year": 2018, "venue": "Proceedings of The Third Workshop on Representation Learning for NLP", "volume": "", "issue": "", "pages": "164--174", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yinfei Yang, Steve Yuan, Daniel Cer, Sheng-yi Kong, Noah Constant, Petr Pilar, Heming Ge, Yun-Hsuan Sung, Brian Strope, and Ray Kurzweil. 2018. Learning semantic textual similarity from conversations. In Proceedings of The Third Workshop on Representation Learning for NLP, pages 164- 174.", "links": null }, "BIBREF26": { "ref_id": "b26", "title": "The united nations parallel corpus v1.0", "authors": [ { "first": "Micha\u0142", "middle": [], "last": "Ziemski", "suffix": "" }, { "first": "Marcin", "middle": [], "last": "Junczys-Dowmunt", "suffix": "" }, { "first": "Bruno", "middle": [], "last": "Pouliquen", "suffix": "" } ], "year": 2016, "venue": "Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC 2016)", "volume": "", "issue": "", "pages": "3530--3534", "other_ids": {}, "num": null, "urls": [], "raw_text": "Micha\u0142 Ziemski, Marcin Junczys-Dowmunt, and Bruno Pouliquen. 2016. The united nations parallel corpus v1.0. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC 2016), pages 3530-3534.", "links": null } }, "ref_entries": { "TABREF0": { "html": null, "content": "
Dot ProductSoftmaxDot Product
Fully Connected
FullyFully
ConnectedConnected(P enc , H enc , |P enc -H enc |, P enc *H enc )
Q encA encP encH encS encT enc
SharedEncoderEncoderEncoderEncoderEncoderEncoder
QuestionAnswerPremieseHypothesisSourceTarget
Question AnswerNLITranslation
(en, es, zh \u2026)(en, es, zh \u2026)(en-es, en-zh, \u2026)
", "num": null, "type_str": "table", "text": ", we target broad coverage using a" }, "TABREF2": { "html": null, "content": "
QATranslationNLI
LangNative + TranslatedNative Translated
ar60M158M570K
de75M517M570K
en2.7B-570K
es340M416M570K
fr92M586M570K
it103M261M570K
ja384M69M570K
ko60M57M570K
nl60M574M570K
pt180M536M570K
pl60M292M570K
ru112M148M570K
th60M70M570K
tr69M415M570K
zh1B112M570K
zh-t147M112M570K
", "num": null, "type_str": "table", "text": "Training tasks for the multilingual sentence encoder. For better coverage across languages, we combine native text with machine translated (MT) data. For NLI, native data is only used for English (en)." }, "TABREF3": { "html": null, "content": "", "num": null, "type_str": "table", "text": "" }, "TABREF4": { "html": null, "content": "
4 Training and Configuration 4.1 Training Corpus Training data consists of mined question-answer pairs, 7 mined translation pairs, 8 and the Stanford Natural Language Inference (SNLI) corpus (Bow-man et al., 2015). 9 SNLI only contains English data. Model USETrans USECNN Gillick et al. (2018)Quora AskUbuntu Average 89.1 42.3 65.7 89.2 39.9 64.6 87.5 37.3 62.4
", "num": null, "type_str": "table", "text": "The number of mined questions-answer pairs also varies across languages with a bias toward a handful of top tier languages. To balance training across languages, we use Google's translation system to translate SNLI to the other 15 languages." }, "TABREF5": { "html": null, "content": "", "num": null, "type_str": "table", "text": "MAP@100 on SR (English). Models are compared with the best models fromGillick et al. (2018) that exclude in-domain training data." }, "TABREF8": { "html": null, "content": "
", "num": null, "type_str": "table", "text": "" }, "TABREF10": { "html": null, "content": "
(\ufe01 In table 8, our multilingual models show competi-|| || || || )\ufe01 , following Yang et al. (2018). tive transfer performance when compared to state-CR SUBJ MPQA TREC SST STS Bench (dev / test) 73.8 83.2 MR 90.1 87.7 96.4 78.1 0.829 / 0.809 78.1 87.0 92.1 89.9 96.6 80.9 0.837 / 0.825 81.1 86.3 92.4 90.2 88.2 84.6 0.801 / 0.758 79.4 83.1 93.7 89.3 ---82.4 86.0 94.8 90.2 92.4 87.6 -of-theModel USE mutlilingual models USECNN USETransformer The state-of-the-art English embedding models InferSent (Conneau et al., 2017) Skip-Thought LN (Ba et al., 2016) Quick-Thought (Logeswaran and Lee, 2018) USEDAN for English
", "num": null, "type_str": "table", "text": "-art sentence embedding models. USE Trans performs better than USE CNN on all tasks. Our new" }, "TABREF12": { "html": null, "content": "
(a) CPU Inference Time(b) GPU Inference Time(c) Memory Footprint
Figure 2: Resource usage for the multilingual Transformer and CNN encoding models.
multilingual USE Trans model outperforms our best
previously released English only model, USE Trans
for English
", "num": null, "type_str": "table", "text": "Performance on English transfer tasks from SentEval(Conneau and Kiela, 2018)." } } } }