chunk_id stringlengths 3 7 | chunk stringlengths 1 823 | source_url stringclasses 416
values | title stringclasses 416
values | chunk_idx int64 0 294 | chunk_start_char int64 0 139k | chunk_end_char int64 303 139k |
|---|---|---|---|---|---|---|
38_30 | After that, they assigned a INLINEFORM0 label to a returned document if it is a SAT click and also assigned INLINEFORM1 labels to the remaining top-10 documents. The rank position of the INLINEFORM2 labeled documents is used as the ground truth to evaluate the search performance before and after re-ranking.
The datas... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 30 | 14,739 | 15,248 |
38_31 | The training, validation and test sets consist of 5,658, 1,184 and 1,210 relevant (i.e., valid) triples; and 40,239, 7,882 and 8,540 irrelevant (i.e., invalid) triples, respectively.
Evaluation protocol: Our CapsE is used to re-rank the original list of documents returned by a search engine as follows: (i) We train o... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 31 | 15,248 | 15,657 |
38_32 | (ii) We then sort the scores in the descending order to obtain a new ranked list. To evaluate the performance of our proposed model, we use two standard evaluation metrics: mean reciprocal rank (MRR) and Hits@1. For each metric, the higher value indicates better ranking performance.
We compare CapsE with the followin... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 32 | 15,657 | 16,202 |
38_33 | (3) SP BIBREF9 , BIBREF11 : A search personalization method makes use of the session-based user profiles. (4) Following BIBREF12 , we use TransE as a strong baseline model for the search personalization task. Previous work shows that the well-known embedding model TransE, despite its simplicity, obtains very competiti... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 33 | 16,202 | 16,701 |
38_34 |
Embedding initialization: We follow BIBREF12 to initialize user profile, query and document embeddings for the baselines TransE and ConvKB, and our CapsE.
We train a LDA topic model BIBREF31 with 200 topics only on the relevant documents (i.e., SAT clicks) extracted from the query logs. We then use the trained LDA m... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 34 | 16,701 | 17,203 |
38_35 | In particular, the INLINEFORM1 element ( INLINEFORM2 ) of the vector embedding for document INLINEFORM3 is: INLINEFORM4 where INLINEFORM5 is the probability of the topic INLINEFORM6 given the document INLINEFORM7 .
We also represent each query by a probability distribution vector over topics. Let INLINEFORM0 be the s... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 35 | 17,203 | 17,616 |
38_36 | The INLINEFORM4 element of the vector embedding for query INLINEFORM5 is defined as in BIBREF12 : INLINEFORM6 , where INLINEFORM7 is the exponential decay function of INLINEFORM8 which is the rank of INLINEFORM9 in INLINEFORM10 . And INLINEFORM11 is the decay hyper-parameter ( INLINEFORM12 ). Following BIBREF12 , we u... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 36 | 17,616 | 18,107 |
38_37 | Thus, after the initialization process, we fix (i.e., not updating) query and document embeddings during training for TransE, ConvKB and CapsE.
In addition, as mentioned by BIBREF9 , the more recently clicked document expresses more about the user current search interest. Hence, we make use of the user clicked docume... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 37 | 18,107 | 18,576 |
38_38 |
Hyper-parameter tuning: For our CapsE model, we set batch size to 128, and also the number of neurons within the capsule in the second capsule layer to 10 ( INLINEFORM0 ). The number of iterations in the routing algorithm is set to 1 ( INLINEFORM1 ). For the training model, we use the Adam optimizer with the initial ... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 38 | 18,576 | 19,087 |
38_39 | We run the model up to 200 epochs and perform a grid search to choose optimal hyper-parameters on the validation set. We monitor the MRR score after each training epoch and obtain the highest MRR score on the validation set when using INLINEFORM10 and the initial learning rate at INLINEFORM11 .
We employ the TransE a... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 39 | 19,087 | 19,569 |
38_40 | We also monitor the MRR score after each training epoch and attain the highest MRR score on the validation set when using margin = 5, INLINEFORM0 -norm and SGD learning rate at INLINEFORM1 for TransE; and INLINEFORM2 and the Adam initial learning rate at INLINEFORM3 for ConvKB.
Main experimental results
Table TABREF... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 40 | 19,569 | 20,024 |
38_41 | Our CapsE performs better than its closely related CNN-based model ConvKB on both experimental datasets (except Hits@10 on WN18RR and MR on FB15k-237), especially on FB15k-237 where our CapsE gains significant improvements of INLINEFORM0 in MRR (which is about 25.1% relative improvement), and INLINEFORM1 % absolute im... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 41 | 20,024 | 20,365 |
38_42 | Table TABREF10 also shows that our CapsE obtains the best MR score on WN18RR and the highest MRR and Hits@10 scores on FB15k-237.
Following BIBREF3 , for each relation INLINEFORM0 in FB15k-237, we calculate the averaged number INLINEFORM1 of head entities per tail entity and the averaged number INLINEFORM2 of tail en... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 42 | 20,365 | 20,793 |
38_43 | If INLINEFORM6 1.5 and INLINEFORM7 1.5, INLINEFORM8 is categorized one-to-many (1-M). If INLINEFORM9 1.5 and INLINEFORM10 1.5, INLINEFORM11 is categorized many-to-one (M-1). If INLINEFORM12 1.5 and INLINEFORM13 1.5, INLINEFORM14 is categorized many-to-many (M-M). | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 43 | 20,793 | 21,057 |
38_44 | As a result, 17, 26, 81 and 113 relations are labelled 1-1, 1-M, M-1 and M-M, respectively. And 0.9%, 6.3%, 20.5% and 72.3% of the test triples in FB15k-237 contain 1-1, 1-M, M-1 and M-M relations, respectively. | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 44 | 21,057 | 21,269 |
38_45 |
Figure FIGREF11 shows the Hits@10 and MRR results for predicting head and tail entities w.r.t each relation category on FB15k-237. | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 45 | 21,269 | 21,401 |
38_46 | CapsE works better than ConvKB in predicting entities on the “side M” of triples (e.g., predicting head entities in M-1 and M-M; and predicting tail entities in 1-M and M-M), while ConvKB performs better than CapsE in predicting entities on the “side 1” of triples (i.e., predicting head entities in 1-1 and 1-M; and pr... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 46 | 21,401 | 21,841 |
38_47 | INLINEFORM0 , INLINEFORM1 , INLINEFORM2 and INLINEFORM3 are symmetric relations which can be considered as M-M relations. Our CapsE also performs better than ConvKB on these 4 M-M relations. Thus, results shown in Figures FIGREF11 and FIGREF12 are consistent. These also imply that our CapsE would be a potential candid... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 47 | 21,841 | 22,246 |
38_48 |
We see that the length and orientation of each capsule in the first layer can also help to model the important entries in the corresponding dimension, thus CapsE can work well on the “side M” of triples where entities often appear less frequently than others appearing in the “side 1” of triples. Additionally, existin... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 48 | 22,246 | 22,711 |
38_49 | These are reasons why our CapsE can be considered as the best one on FB15k-237 and it outperforms most existing models on WN18RR.
Effects of routing iterations: We study how the number of routing iterations affect the performance. Table TABREF13 shows the Hits@10 scores on the WN18RR validation set for a comparison w... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 49 | 22,711 | 23,181 |
38_50 | We see that the best performance for each setup over each 10 epochs is obtained by setting the number INLINEFORM2 of routing iterations to 1. This indicates the opposite side for knowledge graphs compared to images. In the image classification task, setting the number INLINEFORM3 of iterations in the routing process h... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 50 | 23,181 | 23,616 |
38_51 | In contrast, this property from images may be only right for the 1-1 relations, but not for the 1-M, M-1 and M-M relations in the KGs because of the high variant of each relation type (e.g., symmetric relations) among different entities.
Search personalization application
Given a user, a submitted query and the docu... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 51 | 23,616 | 24,096 |
38_52 | Following BIBREF12 , we represent the relationship between the submitted query, the user and the returned document as a (s, r, o)-like triple (query, user, document). The triple captures how much interest a user puts on a document given a query. Thus, we can evaluate the effectiveness of our CapsE for the search perso... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 52 | 24,096 | 24,680 |
38_53 | This indicates a prospective strategy of expanding the triple embedding models to improve the ranking quality of the search personalization systems. In particular, our MRR and Hits@1 scores are higher than those of TransE (with relative improvements of 14.5% and 22% over TransE, respectively). Specifically, our CapsE ... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 53 | 24,680 | 25,294 |
38_54 | We observe that the performance is improved with the increase in the number of filters since capsules can encode more useful properties for a large embedding size.
Related work
Other transition-based models extend TransE to additionally use projection vectors or matrices to translate embeddings of INLINEFORM0 and IN... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 54 | 25,294 | 25,865 |
38_55 | Moreover, ConvKB BIBREF15 applies convolutional neural network, in which feature maps are concatenated into a single feature vector which is then computed with a weight vector via a dot product to produce the score for the input triple. ConvKB is the most closely related model to our CapsE. See an overview of embeddin... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 55 | 25,865 | 26,493 |
38_56 | That historical information can be used to build the user profile, which is crucial to an effective search personalization system. Widely used approaches consist of two separated steps: (1) building the user profile from the interactions between the user and the search system; and then (2) learning a ranking function ... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 56 | 26,493 | 27,049 |
38_57 | In this case, apart from the user profile, dozens of other features have been proposed as the input of a learning-to-rank algorithm BIBREF9 , BIBREF33 . Alternatively, BIBREF12 modeled the potential user-oriented relationship between the submitted query and the returned document by applying TransE to reward higher sco... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 57 | 27,049 | 27,574 |
38_58 |
Conclusion
We propose CapsE—a novel embedding model using the capsule network to model relationship triples for knowledge graph completion and search personalization. Experimental results show that our CapsE outperforms other state-of-the-art models on two benchmark datasets WN18RR and FB15k-237 for the knowledge gr... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 58 | 27,574 | 28,175 |
38_59 | Our code is available at: https://github.com/daiquocnguyen/CapsE.
Acknowledgement
This research was partially supported by the ARC Discovery Projects DP150100031 and DP160103934. The authors thank Yuval Pinter for assisting us in running his code.
Figure 1: An example illustration of our CapsE with k = 4, N = 5, an... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 59 | 28,175 | 28,615 |
38_60 |
Table 2: Experimental results on the WN18RR and FB15k-237 test sets. Hits@10 (H@10) is reported in %. Results of DISTMULT, ComplEx and ConvE are taken from Dettmers et al. (2018). Results of TransE on FB15k237 are taken from Nguyen et al. (2018). Our CapsE Hits@1 scores are 33.7% on WN18RR and 48.9% on FB15k-237. | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 60 | 28,615 | 28,931 |
38_61 | Formulas of MRR and Hits@1 show a strong correlation, so using Hits@1 does not really reveal any additional information for this task. The best score is in bold, while the second best score is in underline. ? denotes our new results for TransE and ConvKB, which are better than those published by Nguyen et al. (2018).
... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 61 | 28,931 | 29,340 |
38_62 |
Figure 3: Hits@10 and MRR on the WN18RR test set w.r.t each relation. The right y-axis is the percentage of triples corresponding to relations.
Table 3: Hits@10 on the WN18RR validation set with N = 50 and the initial learning rate at 1e−5 w.r.t each number of iterations in the routing algorithm m and each 10 traini... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 62 | 29,340 | 29,773 |
38_63 | Hits@1 (H@1) is reported in %. In information retrieval, Hits@1 is also referred to as P@1. The subscripts denote the relative improvement over our TransE results.
Figure 4: Learning curves on the validation set with the initial learning rate at 5e−5. | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 63 | 29,773 | 30,026 |
39_0 | To Tune or Not To Tune? How About the Best of Both Worlds?
The introduction of pre-trained language models has revolutionized natural language research communities. However, researchers still know relatively little regarding their theoretical and empirical properties. In this regard, Peters et al. perform several expe... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 0 | 0 | 591 |
39_1 | In this paper, we propose a new adaptation method which we first train the task model with the BERT parameters frozen and then fine-tune the entire model together. Our experimental results show that our model adaptation method can achieve 4.7% accuracy improvement in semantic similarity task, 0.99% accuracy improvemen... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 1 | 591 | 1,209 |
39_2 | The contribution of these models can be categorized into two aspects. First, pre-trained language models allow modelers to achieve reasonable accuracy without the need an excessive amount of manually labeled data. This strategy is in contrast with the classical deep learning methods, which requires a multitude more da... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 2 | 1,209 | 1,832 |
39_3 |
In the post pre-trained language model era, to pursue new state-of-art, two directions can be followed. The first method, is to improve the pre-training process, such as in the work of ERNIE BIBREF8 , GPT2.0 BIBREF2 and MT-DNN BIBREF9 . The second method is to stand on the shoulder of the pre-trained language models.... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 3 | 1,832 | 2,275 |
39_4 |
In principles, there are three ways to train the networks with stacked neural networks on top of pre-trained language models, as shown in Table TABREF1 . In Peters et al . BIBREF0 , the authors compare the possibility of option stack-only and finetune-only, and conclude that option finetune-only is better than option... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 4 | 2,275 | 2,813 |
39_5 | BIBREF0 did not compare option stack-and-finetune and finetune-only. On the other hand, before pre-trained deep language models became popular, researchers often use a strategy analog to option stack-and-finetune. That is, modelers first train the model until convergence, and then fine-tune the word embeddings with a ... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 5 | 2,813 | 3,335 |
39_6 |
In this study, we aim to compare the strategy stack-and-finetune and strategy finetune-only. More specifically, we perform three NLP tasks, sequence labeling, text classification, and question similarity. In the first tasks, we demonstrate that even without modifying the network structures, building networks on top o... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 6 | 3,335 | 3,859 |
39_7 | Finally, in the last task, we demonstrate that if one can tailor-made a neural network that specifically fit the characteristics of the pre-trained language models, one can improve the accuracy even further. All the results indicate the strategy stack-and-finetune is superior to strategy finetune-only. This leads us t... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 7 | 3,859 | 4,420 |
39_8 | BIBREF0 , this allows us to achieve better results, at least on the selected tasks. More importantly, the results of this study demonstrate the importance of neural networks design, even in the presence of all-powerful pre-trained language models. Second, during the experiment, we have found that although simply using... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 8 | 4,420 | 5,085 |
39_9 | Especially, since word embeddings can be viewed as shallow transfer learning, while pre-trained model should be viewed as deep transfer learning, one must try to combat over-fitting problems with more care due to the enormous number of parameters presented in the pre-trained models. Besides, we also find that in order... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 9 | 5,085 | 5,676 |
39_10 | First, we review the relevant literature on pre-trained deep neural networks, the argument in Peter et al. BIBREF0 as well as fine-tuning strategies with word embeddings. Second, we present three experiments and showed the superiority of strategy stack-and-finetune compared to strategy finetune-only. Finally, we concl... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 10 | 5,676 | 6,187 |
39_11 | Among all of them, one of the most famous is the word embeddings, which maps each word into a continuous vector, instead of one-hot encodings BIBREF10 . By doing so, not only are we able to reduce the dimensionality of the input features, which helps to avoid over-fitting, but also capture, at least partially, the int... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 11 | 6,187 | 6,719 |
39_12 | For example, consider the word ”bank” sentences “I am walking on the bank of the river.” with “I am going to rob the bank”. It is obvious that the word “bank” represents completely different meaning, which the word embeddings techniques fail to capture.
The aforementioned deficiencies prompt researchers to propose de... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 12 | 6,719 | 7,200 |
39_13 | Some early attempts include pre-trained models includes, CoVe BIBREF11 , CVT BIBREF12 , BIBREF13 , ELMo BIBREF14 and ULMFiT BIBREF15 . However, the most successful ones are BERT BIBREF1 and Open-GPT BIBREF2 . Unlike standard NLP deep learning model, BERT and Open-GPT are built on top of transformer BIBREF16 structures... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 13 | 7,200 | 7,564 |
39_14 | The difference between BERT and Open-GPT is that BERT uses bi-directional self-attentions while Open-GPT uses only unidirectional ones, as shown in Figure FIGREF2 . The transformer structures differ from the LSTM's in the two important aspects. First, it allows for stacking of multiple layers with residual connections... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 14 | 7,564 | 8,097 |
39_15 | After training on a large corpus, both BERT and Open-GPT are able to renew the SOTA of many important natural language tasks, such as such as SQuAD BIBREF3 , CoQA BIBREF4 , named entity recognition BIBREF5 , Glue BIBREF6 , machine translation BIBREF7 .
In the presence of the success of pre-trained language models, es... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 15 | 8,097 | 8,587 |
39_16 | BIBREF20 investigated the linguistic knowledge and transferability of contextual representations by comparing BERT BIBREF1 with ELMo BIBREF14 , and concluded that while the higher levels of LSTM's are more task-specific, this trend does not exhibit in transformer based models. Stickland and Murray BIBREF21 invented pr... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 16 | 8,587 | 9,101 |
39_17 | BIBREF22 propose a “post-training” algorithms, which does not directly fine-tune BERT, but rather first “post-train” BERT on the task related corpus using the masked language prediction task next sentence prediction task, which helps to reduce the bias in the training corpus. Finally, Sun et al. BIBREF23 added additio... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 17 | 9,101 | 9,654 |
39_18 | That is, whether it is possible to top BERT with the commonly used or task specific layers, and if this is possible, how to best utilize the pre-trained language models in this situation. In this regards, Peters et al. BIBREF0 investigated how to best adapt the pre-trained model to a specific task, and focused on two ... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 18 | 9,654 | 10,195 |
39_19 | BIBREF0 performs five experiments, including: (1) named entity recognition BIBREF5 ; (2) sentiment analysis BIBREF24 ; (3) natural language inference BIBREF25 ; (4) paraphrase detection BIBREF26 ; (5) semantic textual similarity BIBREF27 . By the results of these tasks, Peters et al. BIBREF0 concludes that adding a li... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 19 | 10,195 | 10,647 |
39_20 |
Methodology
Under our strategy stack-and-finetune, the model training process is divided into two phases, which are described in detail below. In the first phase, the parameters of the pre-training model are fixed, and only the upper-level models added for a specific task is learned. In the second phase, we fine-tun... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 20 | 10,647 | 11,281 |
39_21 | BIBREF1 , the author directly trained BERT along with with a light-weighted task-specific head. In our case though, we top BERT with a more complex network structure, using Kaiming initialization BIBREF28 . | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 21 | 11,281 | 11,488 |
39_22 | If one would fine-tune directly the top models along with the weights in BERT, one is faced with the following dilemma: on the one hand, if the learning rate is too large, it is likely to disturb the structure innate to the pre-trained language models; on the other hand, if the learning rate is too small, since we top... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 22 | 11,488 | 12,022 |
39_23 |
Another aspect that is worth commenting in the first phase is that it is most beneficial that one does not train the top model until it reaches the highest accuracy on the training or validation data sets, but rather only up to a point where the prediction accuracy of the training and validation data sets do not diff... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 23 | 12,022 | 12,560 |
39_24 | Therefore, if one were to train the top models until they reach the highest prediction accuracy in the training or validation data sets, it would likely cause the models to over-fit. Therefore, in our experiment, we found that this leads to the highest performance increase in the fine-tuning stage.
Overview
We perfo... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 24 | 12,560 | 13,208 |
39_25 | Second, we perform a text classification experiments, in this experiments, we trained three models, and perform a model ensemble. We hope to show that even the added network has not contributed to significantly in improving the accuracy, it does provide opportunities for model ensembles. Finally, we perform the textua... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 25 | 13,208 | 13,864 |
39_26 | In the sequence labeling task, the BERT word embedding of each word passes through two fully connected layers, and the prediction probability of named entity can be obtained. In the next two verification tasks, we use “[CLS]” for prediction and add two fully connected layers subsequently. Under our strategy stack-and-... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 26 | 13,864 | 14,245 |
39_27 | We tried to set the learning rate of the first stage to INLINEFORM0 , INLINEFORM1 , INLINEFORM2 , INLINEFORM3 and INLINEFORM4 , and set it to a smaller number in the latter stage, such as INLINEFORM5 , INLINEFORM6 , INLINEFORM7 and INLINEFORM8 . After our experiments, we found that it gets better results while the lea... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 27 | 14,245 | 14,680 |
39_28 | Since BERT-Adam BIBREF1 has excellent performance, in our experiments, we use it as an optimizer with INLINEFORM10 , INLINEFORM11 -weight decay of INLINEFORM12 .We apply a dropout trick on all layers and set the dropout probability as 0.1. | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 28 | 14,680 | 14,920 |
39_29 |
Experiment A: Sequence Labeling
In the sequence labeling task,we explore sub-task named entity recognition using CoNLL03 dataset BIBREF5 , which is a public available used in many studies to test the accuracy of their proposed methods BIBREF29 , BIBREF30 , BIBREF31 , BIBREF32 , BIBREF1 . For strategy finetune-only a... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 29 | 14,920 | 15,400 |
39_30 |
As is shown in Table 2, even without modifying the networks to specifically adapt to the pre-trained model, our training strategy still brought improvement towards overall accuracy of 0.99% for the accuracy and 0.068 on the F1 score, proving the success of our proposed methods.
Experiment B: Text Classification
In ... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 30 | 15,400 | 15,913 |
39_31 | As for the upper model,we choose DenseNet BIBREF33 and HighwayLSTM BIBREF34 .
The DenseNet structure contains four independent blocks and each block has four CNNs connected by residual. We initialize word embedding in the word representation layer with BERT. We initialize each character as a 768-dimension vector. In ... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 31 | 15,913 | 16,335 |
39_32 |
We find the ensembled model enjoys a 0.72% improvements compared to the fine-tune only model and 0.005 improvement for the F1 score.
Experiment C: Semantic Similarity Tasks
We use “Quora-Question-Pair” dataset 1. This is a commonly used dataset containing 400k question pairs, annotated manually to be semantically e... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 32 | 16,335 | 16,780 |
39_33 | Various models which are tested on this data set are proposed, including but not limited to BIBREF35 , BIBREF36 , BIBREF37 , BIBREF38 .
Apart from the BERT fine-tuning only model and BERT+ BIMPM model, we also devise two new network structures by modifying the BIMPM model. In the first model is to remove the first bi... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 33 | 16,780 | 17,173 |
39_34 | In the second model, we combine the matching layer of BIMPM and with a transformer BIBREF16 , a model we call Sim-Transformer by replacing the output layer of the matching layer, originally a bi-LSTM model, with a transformer model. From the experimental results shown in Table 4, we can see that due to the strong expr... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 34 | 17,173 | 17,619 |
39_35 | In addition, we also find that Sim-Transformer's performance without fine-tuning is nearly four percentage points lower than BIMPM, but it out-performs BIMPM after fine-tuning. In general, the results show that BERT + Sim-Transformer out-performs BERT-only model by 4.7%, thus confirming our hypotheses again.
Discussi... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 35 | 17,619 | 18,127 |
39_36 | However, we would like to caution the readers in two aspects when reading the conclusion of this study. First, this study does not argue that our proposed methods are always superior to fine-tuning only methods. For example, all the experiments in our study are based on data sets of relatively large size. In the other... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 36 | 18,127 | 18,601 |
39_37 | If this is the case, then it is possible that deep domain adaptation BIBREF39 might be a better choice if one desires to stack neural networks on top of pre-trained language models. However, most domain adaptation applications belong to the field of computer vision, therefore, a call for domain adaptations research in... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 37 | 18,601 | 19,227 |
39_38 | In classical embedding + training networks, the general training method is to fix the word-embeddings, then train the top model until it converges, and finally fine-tuning the word-embeddings for a few epochs. This training strategy does not work when we replace pre-trained language models with word-embeddings. In our... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 38 | 19,227 | 19,799 |
39_39 | After that, we fine-tune the pre-trained language model as well as our models on top together. This allows us to achieve better results on the experimentation. However, it is not yet clear to us when to stop the training of top neural networks. This poses an even more essential question for Auto ML researchers in the ... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 39 | 19,799 | 20,321 |
39_40 | While if Auto ML is to be performed on NLP tasks successfully, it might be essential that the gap between training accuracy and test accuracy to be incorporated when one evaluates the model.
Finally, it is not yet clear what is the most proper way to build networks that tops the pre-trained language models. However, ... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 40 | 20,321 | 20,891 |
39_41 | Second, this also means, one does not need explicitly to build extremely complex networks on top of pre-trained language models unless it complements the mechanisms of self-attention. Finally, a challenge remains as to how to use the depth of pre-trained language models. The process of our experiment shows that utiliz... | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 41 | 20,891 | 21,443 |
39_42 | [2], Figure 1
Table 2: Results for named entity recognition
Table 3: Results for text classification
Table 4: Results for semantic similarity task | https://arxiv.org/abs/1907.05338 | To Tune or Not To Tune? How About the Best of Both Worlds? | 42 | 21,443 | 21,593 |
40_0 | A Corpus of Adpositional Supersenses for Mandarin Chinese
Adpositions are frequent markers of semantic relations, but they are highly ambiguous and vary significantly from language to language. Moreover, there is a dearth of annotated corpora for investigating the cross-linguistic variation of adposition semantics, or... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 0 | 0 | 592 |
40_1 | Our approach adapts a framework that defined a general set of supersenses according to ostensibly language-independent semantic criteria, though its development focused primarily on English prepositions (Schneider et al., 2018). We find that the supersense categories are well-suited to Chinese adpositions despite synt... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 1 | 592 | 1,135 |
40_2 | prepositions and postpositions) include some of the most frequent words in languages like Chinese and English, and help convey a myriad of semantic relations of space, time, causality, possession, and other domains of meaning. They are also a persistent thorn in the side of second language learners owing to their extr... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 2 | 1,135 | 1,683 |
40_3 | Semantically annotated corpora of adpositions in multiple languages, including parallel data, would facilitate broader empirical study of adposition variation than is possible today, and could also contribute to NLP applications such as machine translation BIBREF3, BIBREF4, BIBREF5, BIBREF6, BIBREF7, BIBREF8, BIBREF9 ... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 3 | 1,683 | 2,191 |
40_4 | For this corpus we have adapted schneider-etal-2018-comprehensive Semantic Network of Adposition and Case Supersenses annotation scheme (SNACS; see sec:snacs) to Chinese. Though other languages were taken into consideration in designing SNACS, no serious annotation effort has been undertaken to confirm empirically tha... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 4 | 2,191 | 2,547 |
40_5 | After developing new guidelines for syntactic phenomena in Chinese (subsec:adpositioncriteria), we apply the SNACS supersenses to a translation of The Little Prince (3 2 3), finding the supersenses to be robust and achieving high inter-annotator agreement (sec:corpus-annotation). We analyze the distribution of adposit... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 5 | 2,547 | 3,135 |
40_6 | The annotated corpus and the Chinese guidelines for SNACS will be made freely available online.
Related Work
To date, most wide-coverage semantic annotation of prepositions has been dictionary-based, taking a word sense disambiguation perspective BIBREF16, BIBREF17, BIBREF18. BIBREF19 proposed a supersense-based (un... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 6 | 3,135 | 3,640 |
40_7 | Previous SNACS annotation efforts have been mostly focused on English—particularly STREUSLE BIBREF20, BIBREF0, the semantically annotated corpus of reviews from the English Web Treebank BIBREF21. We present the first adaptation of SNACS for Chinese by annotating an entire Chinese translation of The Little Prince.
Rel... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 7 | 3,640 | 4,222 |
40_8 | Most previous semantic projects for Mandarin Chinese focused on content words and did not directly annotate the semantic relations signaled by functions words such as prepositions BIBREF23, BIBREF24, BIBREF25, BIBREF26. For example, in Chinese PropBank, BIBREF27 argued that the head word and its part of speech are cle... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 8 | 4,222 | 4,666 |
40_9 | BIBREF28 annotated the Tsinghua Corpus BIBREF29 from People’s Daily where the content words were selected as the headwords, i.e., the object is the headword of the prepositional phrase. In these prepositional phrases, the nominal headwords were labeled with one of the 59 semantic relations (e.g. Location, LocationIni,... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 9 | 4,666 | 5,130 |
40_10 | Similarly, in Semantic Dependency Relations (SDR, BIBREF30, BIBREF31), prepositions and localizers were labeled as semantic markers mPrep and mRange, whereas semantic roles, e.g., Location, Patient, are assigned to the governed nominal phrases.
BIBREF32 compared PropBank parsing performance on Chinese and English, an... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 10 | 5,130 | 5,641 |
40_11 | The high frequency of prepositions as head words in PropBank reflects their importance in context. However, very few annotation scheme attempted to directly label the semantics of these adposition words.
BIBREF33 is the most relevant adposition annotation effort, categorizing Chinese prepositions into 66 types of sen... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 11 | 5,641 | 6,227 |
40_12 |
Related Work ::: SNACS: Adposition Supersenses
BIBREF0 proposed the Semantic Network of Adposition and Case Supersenses (SNACS), a hierarchical inventory of 50 semantic labels, i.e., supersenses, that characterize the use of adpositions, as shown in fig:supersenses. Since the meaning of adpositions is highly affecte... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 12 | 6,227 | 6,800 |
40_13 | In single-label-amb, however, the same preposition about takes a measurement in the context, expressing an approximation.
. I gave a presentation about:Topic urbanization.
. We have about:Approximator 3 eggs left.
Though assigning a single label to each adposition can help capture its lexical contribution to the se... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 13 | 6,800 | 7,347 |
40_14 |
For instance, eg:stimulustopic blends the domains of emotion (principally reflected in care, which licenses a Stimulus), and cognition (principally reflected in about, which often marks non-emotional Topics). Thus, SNACS incorporates the construal analysis BIBREF34 wherein the lexical semantic contribution of an adpo... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 14 | 7,347 | 7,876 |
40_15 |
Another motivation for incorporating the construal analysis, as pointed out by BIBREF34, is its capability to adapt the English-centric supersense labels to other languages, which is the main contribution of this paper. The construal analysis can give us insights into the similarities and differences of function and ... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 15 | 7,876 | 8,408 |
40_16 | The English SNACS guidelines (we use version 2.3) broadly define the set of SNACS annotation targets to include canonical prepositions (taking an noun phrase (NP) complement) and their subordinating (clausal complement) uses. Possessives, intransitive particles, and certain uses of the infinitive marker to are also in... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 16 | 8,408 | 8,865 |
40_17 | However, since no overt derivational morphology occurs on these prepositional tokens (previously referred to as coverbs), we need to filter non-prepositional uses of these words. Secondly, post-nominal particles, i.e., localizers, though not always considered adpositions in Chinese, deliver rich semantic information.
... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 17 | 8,865 | 9,452 |
40_18 | In some cases, coverbs can also occur as the main predicate. For example, the coverb 4 heads the predicate phrase in zho:pred.
. 1 4:Locus 24 4:TopicLocus 3342.
3sg p:at academia lc:on-top-of successful
`He succeeded in academia.’
. 3 4 de 2 4 4 34.
2sg want de sheep res at inside
`The sheep you wanted is in the... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 18 | 9,452 | 9,778 |
40_19 | (zh_lpp_1943.92)
In this project, we only annotate coverbs when they do not function as the main predicate in the sentence, echoing the view that coverbs modify events introduced by the predicates, rather than establishing multiple events in a clause BIBREF37. Therefore, lexical items such as 4 are annotated when fun... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 19 | 9,778 | 10,330 |
40_20 | For example, 4 in zho:shangtopic denotes a contextual meaning, `in a particular area,' whereas the co-occurring coverb 4 only conveys a generic location. It is unclear whether localizers are syntactically postpositions, but we annotate all localizers because of their semantic significance. Though coverbs frequently co... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 20 | 10,330 | 10,808 |
40_21 | As a result, we treat them as separate targets for SNACS annotation: for example, 4 and 4 receive Locus and TopicLocus respectively in zho:shangtopic.
Setting aside the syntactic controversies of coverbs and localizers in Mandarin Chinese, we regard both of them as adpositions that merit supersense annotations. As in... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 21 | 10,808 | 11,289 |
40_22 | In this paper, we take the stance that coverbs co-occur with the main predicate and precede an NP, whereas localizers follow a noun phrase and add semantic information to the clause.
Corpus Annotation
We chose to annotate the novella The Little Prince because it has been translated into hundreds of languages and dia... | https://arxiv.org/abs/2003.08437 | A Corpus of Adpositional Supersenses for Mandarin Chinese | 22 | 11,289 | 11,863 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.