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 |
|---|---|---|---|---|---|---|
27_41 | We denote the monolingual encoder-decoder model as Bert2Bert ($\sim $220M parameters) and causal decoder model as CausalBert ($\sim $110M parameters). Then we fine-tune each model in each language independently for the same number of epoch and optimizer as the multilingual model.
Experiments ::: Implementation Detail... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 41 | 20,712 | 21,260 |
27_42 | We adapt this model to the other languages by using the Google Translate API to translate target languages (e.g., Chinese) query to English as the input to the model, then translate the English response back to the target language. Thus, the response generation flow is: target query $\rightarrow $ English query $\righ... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 42 | 21,260 | 21,823 |
27_43 | Then, we follow the second pre-training stage of XNLG BIBREF4 for pre-training Italian, Japanese, Korean, Indonesia cross-lingual transferable models. For Chinese and French, we directly apply the pre-trained XNLG BIBREF4 weights. Then, the pre-trained models are fine-tune on English PersonaChat training set and early... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 43 | 21,823 | 22,207 |
27_44 |
Experiments ::: Results and Discussion ::: Quantitative Analysis
Table TABREF20 compares monolingual, multilingual, and cross-lingual models in terms of BLEU and perplexity in the human-translated test set. On both evaluation matrices, the causal decoder models outperform the encoder-decoder models. We observe that ... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 44 | 22,207 | 22,622 |
27_45 | (Generated samples are available in Appendix D) We hypothesize that this is because the one-to-many problem BIBREF76 in open-domain conversation weakens the relation between encoder and decoder; thus the well pre-trained decoder (Bert) easily converges to a locally-optimal, and learns to ignore the dialogue context fr... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 45 | 22,622 | 23,081 |
27_46 | On the other hand, M-CausalBert achieves a comparable or slightly better performance compared to CausalBert, which suggests that M-CausalBert leverages the data from other languages. As expected, we observe a significant gap between the cross-lingual model and other models, which indicates that cross-lingual zero-shot... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 46 | 23,081 | 23,621 |
27_47 | The results illustrate that Multi outperforms Mono in English and Chinese, and is on par with Mono in other languages. On the other hand, Poly shows a strong performance in English as it was pre-trained with a large-scale English conversation corpus. In contrast, the performance of Poly drops in other languages, which... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 47 | 23,621 | 24,017 |
27_48 |
Experiments ::: Results and Discussion ::: Qualitative Analysis and Discussion
We randomly sample 7 self-chat dialogues for each baseline model in the seven languages and report them in Appendix D., And we summarize the generation of each model as follows:
Experiments ::: Results and Discussion ::: Qualitative Anal... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 48 | 24,017 | 24,596 |
27_49 |
Experiments ::: Results and Discussion ::: Qualitative Analysis and Discussion ::: Monolingual & Multilingual
We observe that both the monolingual and multilingual models can generate fluent responses. Compared to Bert2Bert and M-Bert2Bert, CausalBert and M-CausalBert can generate more on-topic responses but sometim... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 49 | 24,596 | 25,185 |
27_50 | Therefore, we experiment on M-CausalBert with two settings: 1) many-to-one, in which users converse with the model in 6 languages, and the model generate responses in English, 2) one-to-many, in which users converse with the model using English, and the model generates responses in 6 languages using language embedding... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 50 | 25,185 | 25,801 |
27_51 | Understanding the mixed-language dialogue context is a desirable skill for end-to-end chit-chat systems, and a systematic study of this research question is needed in future.
Experiments ::: Results and Discussion ::: Qualitative Analysis and Discussion ::: Cross-lingual.
The current state-of-the-art cross-lingual g... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 51 | 25,801 | 26,457 |
27_52 |
Conclusion
In this paper, we studied both cross-lingual and multilingual approaches in end-to-end personalized dialogue modeling. We presented the XPersona dataset, a multilingual extension of Persona-Chat, for evaluating the multilingual personalized chatbots. We further provided both cross-lingual and multilingual... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 52 | 26,457 | 26,974 |
27_53 | The experimental results showed that multilingual trained models, with a single model across multiple languages, can outperform the two-stage translation approach and is on par with monolingual models. On the other hand, the current state-of-the-art cross-lingual approach XNLG achieved lower performance than other bas... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 53 | 26,974 | 27,480 |
27_54 |
Dataset Collection ::: Annotation Instructions
In this section, we show the instructions for French annotation:
There are two existing columns of conversations: the first column (en) is the original conversations in English, the second column (fr) is the conversations translated by an automatic system (e.g., Google... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 54 | 27,480 | 28,142 |
27_55 | Hence you can customize dialogues and persona sentences to make them fluent and coherent in the target language, including by deviating from the original translation. However, you should retain persona and conversation consistency.
Dataset Collection ::: Training Set Statistics
We report our iterative revised traini... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 55 | 28,142 | 28,796 |
27_56 | Then the annotators are asked the following questions:
Who would you talk to for a long conversation?
If you had to say one of these speakers is interesting and one is boring, who would you say is more interesting?
Which speaker sounds more human?
Generated Samples ::: Mixed-language Samples
We report more the mi... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 56 | 28,796 | 29,194 |
27_57 |
Generated Samples ::: Model Comparison Samples
We randomly sample one self-chat dialogue examples for each model in each language and report them in figure 5-32.
in CausalBert,M-CausalBert,PolyEncoder,M-Bert2Bert
in CausalBert,M-CausalBert,PolyEncoder,M-Bert2Bert
in CausalBert,M-CausalBert,PolyEncoder,M-Bert2Bert... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 57 | 29,194 | 29,608 |
27_58 | -Bert2Bert
in CausalBert,M-CausalBert,PolyEncoder,M-Bert2Bert
in CausalBert,M-CausalBert,PolyEncoder,M-Bert2Bert
Table 1: Multi-turn annotated dialogue samples from test set in seven languages. For simplicity, we only show three turns for each dialogue and the persona in English.
Table 2: The statistics of the coll... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 58 | 29,608 | 30,057 |
27_59 | Edit distance per dialogue (Edit) and BLEU score are computed to show the difference between the human-annotated dataset and auto-translated dataset. (Training set is reported in Appendix A)
Figure 1: (a) Multilingual Encoder-Decoder model. (b) Multilingual Causal Decoder model. (Detailed illustration is reported in ... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 59 | 30,057 | 30,573 |
27_60 |
Table 4: Results of ACUTE-EVAL human evaluation. Tests are conducted pairwise between M-CausalBert (Multi.) and other models (Human, Poly-encoder (Poly), Monolingual CausalBert (Mono)). Numbers indicate the winning rate of Multi. Numbers in bold are statistically significant (p < 0.05). | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 60 | 30,573 | 30,862 |
27_61 |
Table 5: Many-to-one: understand mixed-language dialogue context in multiple languages and generate response in one language
Table 6: One-to-many: response one dialogue context with 7 different languages
Figure 2: Human evaluation interface modified from ACUTE-EVAL(Li et al., 2019)
Table 7: The number of dialogues... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 61 | 30,862 | 31,397 |
27_62 |
Figure 4: Multilingual Encoder-Decoder model. | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 62 | 31,397 | 31,444 |
27_63 |
Table 8: One-to-many by M-CausalBert
Table 9: Many-to-one by M-CausalBert
Figure 5: English CausalBert
Figure 12: Chinese M-Bert2Bert
Figure 13: Chinese CrossLingual
Figure 19: France CausalBert
Figure 23: France CrossLingual
Figure 27: Japanese M-Bert2Bert
Figure 31: Korean PolyEncoder
Figure 29: Korean Cau... | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 63 | 31,444 | 31,814 |
27_64 | 37: Indonesian CrossLingual | https://arxiv.org/abs/2003.07568 | XPersona: Evaluating Multilingual Personalized Chatbot | 64 | 31,814 | 31,842 |
28_0 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation
The translation of pronouns presents a special challenge to machine translation to this day, since it often requires context outside the current sentence. Recent work on models that have access to information a... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 0 | 0 | 694 |
28_1 | This paper therefore presents a test suite of contrastive translations focused specifically on the translation of pronouns. Furthermore, we perform experiments with several context-aware models. We show that, while gains in BLEU are moderate for those systems, they outperform baselines by a large margin in terms of ac... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 1 | 694 | 1,341 |
28_2 | They are notoriously hard to translate since they often require context outside the current sentence.
As an example, consider the sentences in Figure FIGREF1 . In both languages, there is a pronoun in the second sentence that refers to the European Central Bank. When the second sentence is translated from English to ... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 2 | 1,341 | 1,942 |
28_3 | In this German sentence, the antecedent Europäische Zentralbank dictates the feminine gender of the pronoun sie.
It is unfortunate, then, that current NMT systems generally operate on the sentence level BIBREF2 , BIBREF3 , BIBREF4 . Documents are translated sentence-by-sentence for practical reasons, such as line-bas... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 3 | 1,942 | 2,327 |
28_4 | Furthermore, improvements of larger-context models over baselines in terms of document-level metrics such as BLEU or RIBES have been moderate, so that their computational overhead does not seem justified, and so that it is hard to develop more effective context-aware architectures and empirically validate them.
To ad... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 4 | 2,327 | 2,834 |
28_5 | The test suite consists of pairs of source and target sentences, in combination with contrastive translation variants (for evaluation by model scoring) and additional linguistic and contextual information (for further analysis). The resource is freely available. Additionally, we evaluate several context-aware models t... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 5 | 2,834 | 3,486 |
28_6 | The context-aware models we use in our experiments are detailed in Section SECREF4 . We discuss our experiments in Section SECREF5 and the results in Section SECREF6 .
Related Work
Two lines of work are related to our paper: research on context-aware translation (described in Section SECREF8 ) and research on focuse... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 6 | 3,486 | 3,875 |
28_7 |
Context-Aware NMT Models
If the translation of a pronoun requires context beyond the current sentence (see the example in Figure FIGREF1 ), a natural extension of sentence-level NMT models is to condition the model prediction on this necessary context. In the following, we describe a number of existing approaches to... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 7 | 3,875 | 4,437 |
28_8 | All of their models are standard sequence-to-sequence models built with recurrent neural networks (RNNs), since the method does not require any architectural change. BIBREF11 use the same concatenation technique with a Transformer architecture BIBREF2 , and experiment with wider context.
A number of works do propose ... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 8 | 4,437 | 5,066 |
28_9 | While BIBREF6 , BIBREF9 extend an RNN encoder-decoder with a second encoder that the decoder attends to, BIBREF8 extend the Transformer architecture with an encoder that is attended to by the main encoder. BIBREF8 also introduce parameter sharing between the main encoder and the context encoder, but do not empirically... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 9 | 5,066 | 5,414 |
28_10 |
While the number of encoded sentences in the previous work is fixed, BIBREF7 , BIBREF10 explore the integration of variable-size context through a hierarchical architecture, where a first-level RNN reads in words to produce sentence vectors, which are then fed into a second-level RNN to produce a document summary.
A... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 10 | 5,414 | 5,933 |
28_11 | Some work considers only source context, but for pronoun translation, target-side context is intuitively important for disambiguation, especially if the antecedent itself is ambiguous. In our evaluation, we therefore emphasize models that take into account both source and target context.
Our experiments are based on ... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 11 | 5,933 | 6,516 |
28_12 |
This section describes several context-aware NMT models that we use in our experiments. They fall into two major categories: models based on RNNs and models based on the Transformer architecture BIBREF2 . We experiment with additional context on the source side and target side.
Evaluation of Pronoun Translation
Pro... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 12 | 6,516 | 7,064 |
28_13 | While suitable for analysis, the small size of the test suites makes it hard to make statistically confident comparisons between systems, and the hand-picked nature of the test suites introduces biases. To overcome these problems, we opted for a fully automatic approach to constructing a large-scale test suite.
Conce... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 13 | 7,064 | 7,635 |
28_14 |
The first edition of the task focused on English INLINEFORM0 French, and it was found that local context (such as the verb group) was a strong signal for pronoun prediction. Hence, future editions only provided target-side lemmas instead of fully inflected forms, which makes the task less suitable to evaluate end-to-... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 14 | 7,635 | 8,057 |
28_15 |
BIBREF17 do not report on the proportion of intra-sentential and inter-sentential anaphora in their test set, but the two top-performing systems only made use of intra-sentential information. Our test suite focuses on allowing the comparison of end-to-end context-aware NMT systems, and we thus extract a large number... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 15 | 8,057 | 8,533 |
28_16 | Our focus on evaluating end-to-end NMT systems also relieves us from having to provide annotated training sets, and reduces pressure to achieve balance and full coverage of phenomena.
An alternative approach to automatically evaluate pronoun translation are reference-based methods that produce a score based on word a... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 16 | 8,533 | 9,008 |
28_17 | BIBREF21 perform a human meta-evaluation and show substantial disagreement between reference-based metrics and human judges, especially because there often exist valid alternative translations that use different pronouns than the reference. Our test set, and our protocol of generating contrastive examples, is focused ... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 17 | 9,008 | 9,585 |
28_18 | As additional goals, our test set is designed to 1) focus on hard cases, so that it can be used as a benchmark to track progress in context-aware translation and 2) allow for fine-grained analysis.
Section SECREF14 describes how we extract our data set. Section SECREF26 explains how, given a set of contrastive exampl... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 18 | 9,585 | 10,074 |
28_19 | The goal is to provide a large number of difficult test cases where an English pronoun has to be translated to a German pronoun.
The most challenging cases are translating it to either er, sie or es, depending on the grammatical gender of the antecedent. Not only is the translation of it ambiguous, there is also clas... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 19 | 10,074 | 10,678 |
28_20 |
To provide a basis for filtering with constraints, we tokenize the whole data set with the Moses tokenizer, generate symmetric word alignments with fast_align BIBREF23 , parse the English text with CoreNLP BIBREF24 , parse the German text with ParZu BIBREF25 and perform coreference resolution on both sides. The coref... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 20 | 10,678 | 11,121 |
28_21 |
Then we opt for high-precision, aggressive filtering, according to the following protocol: for each pair of sentences INLINEFORM0 in English and German, extract iff
INLINEFORM0 contains the English pronoun it, and INLINEFORM1 contains a German pronoun that is third person singular (er, sie or es), as indicated by t... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 21 | 11,121 | 11,631 |
28_22 |
This removes most candidate pairs, but is necessary to overcome the noise introduced by our preprocessing pipeline, most notably coreference resolution. From the filtered set, we create a balanced test set by randomly sampling 4000 instances of each of the three translations of it under consideration (er, sie, es). W... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 22 | 11,631 | 12,176 |
28_23 | A contrastive translation is a translation variant where the correct pronoun is swapped with an incorrect one. For an example, see Table TABREF19 , where the pronoun it in the original translation corresponds to sie because the antecedent bat is a feminine noun in German (Fledermaus). We produce wrong translations by ... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 23 | 12,176 | 12,786 |
28_24 | Those examples do not require any additional context for disambiguation and we therefore expect the sentence-level baseline to perform well on them.
We take extra care to ensure that the resulting contrastive translations are grammatically correct, because ungrammatical sentences are easily dismissed by an NMT system... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 24 | 12,786 | 13,262 |
28_25 |
The German coreference resolution system does not resolve es because most instances of es in German are either non-referential forms, or they refer to a clause instead of a nominal antecedent. We limit the test set to nominal antecedents, as these are the only ambiguous cases with respect to translation. For this rea... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 25 | 13,262 | 13,823 |
28_26 |
Our extraction process respects document boundaries, to ensure we always search for the right context. We extract additional information from the annotated documents, such as the distance (in sentences) between pronouns and their antecedents, the document of origin, lemma, morphology and dependency information if ava... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 26 | 13,823 | 14,457 |
28_27 |
We exploit the fact that NMT systems are in fact language models of the target language, conditioned on source text. Like language models, NMT systems can be used to compute a model score (the negative log probability) for an existing translation. Contrastive evaluation, then, means to compare the model score of two ... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 27 | 14,457 | 15,087 |
28_28 | An entirely different target sequence might rank higher in the system's beam during decoding. The only conclusion permitted by contrastive evaluation is whether or not the reference translation is more probable than a contrastive variant.
If the model score of the reference is indeed higher, we refer to this outcome ... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 28 | 15,087 | 15,701 |
28_29 |
During scoring, the model is provided with reference translations as target context, while during translation, the model needs to predict the full sequence. It is an open question to what extent performance deteriorates when context is itself predicted, and thus noisy. We highlight that the same problem arises for se... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 29 | 15,701 | 16,355 |
28_30 | The input and output embeddings of the decoder are tied, encoder embeddings are not.
concat22 We concatenate each sentence with one preceding sentence, for both the source and target side of the corpus. Then we train on this new data set without any changes to the model architecture. This very simple method is inspir... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 30 | 16,355 | 16,947 |
28_31 | This model has access to one additional context: the previous source sentence. It is read by a separate encoder, and attended to by an additional attention network. The output of the resulting two attention vectors is combined with yet another attention network.
s-t-hier Identical to s-hier, except that it considers ... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 31 | 16,947 | 17,460 |
28_32 |
s-hier-to-2 The model has an additional encoder for source context, whereas the target side of the corpus is concatenated, in the same way as for concat22. This model achieved the best results in BIBREF9 .
For each variant, we also introduce and test weight tying: we share the parameters of embedding matrices betwee... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 32 | 17,460 | 17,941 |
28_33 | A Transformer avoids recurrence completely: it follows an encoder-decoder architecture using stacked self-attention and fully connected layers for both the encoder and decoder.
baseline A standard context-agnostic Transformer. All model parameters are identical to a Transformer-base in BIBREF2 .
concat22 A simple co... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 33 | 17,941 | 18,529 |
28_34 |
BIBREF8 A more sophisticated context-aware Transformer that uses source context only. It has a separate encoder for source context, but all layers except the last one are shared between encoders. A source and context sentence are first encoded independently, and then a single attention layer and a gating function ar... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 34 | 18,529 | 19,098 |
28_35 |
Experiments
We train all models on the data from the WMT 2017 English INLINEFORM0 German news translation shared task ( INLINEFORM1 5.8 million sentence pairs). These corpora do not have document boundaries, therefore a small fraction of sentences will be paired with wrong context, but we expect the model to be robu... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 35 | 19,098 | 19,636 |
28_36 | We learn a joint BPE model with 89.5k merge operations BIBREF29 . We train shallow models with an embedding size of 512, a hidden layer size of 1024 and layer normalization. Models are trained with Adam BIBREF30 , with an initial learning rate of 0.0001. We apply early stopping based on validation perplexity. The batc... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 36 | 19,636 | 20,117 |
28_37 |
For our Transformer-based experiments, we use a custom implementation and follow the hyperparameters from BIBREF2 , BIBREF8 . Systems are trained on lowercased text that was encoded using BPE (32k merge operations). Models consist of 6 encoder and decoder layers with 8 attention heads. The hidden state size is 512, t... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 37 | 20,117 | 20,605 |
28_38 | We compute scores with SacreBLEU BIBREF31 . Evaluation with BLEU is done mainly to control for overall translation quality.
To evaluate pronoun translation, we perform contrastive evaluation and report the accuracy of models on our contrastive test set.
Evaluation
The BLEU scores in Table TABREF30 show a moderate i... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 38 | 20,605 | 21,085 |
28_39 | The contrastive evaluation on our test set on the other hand shows a clear increase in the accuracy of pronoun translation: The best model s-hier-to-2.tied achieves a total of +16 percentage points accuracy on the test set over the baseline, see Table TABREF31 .
Table TABREF32 shows that context-aware models perform ... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 39 | 21,085 | 21,571 |
28_40 | The evaluation according to the distance of the antecedent in Table TABREF35 confirms that the subset of sentences with antecedent distance 1 benefits most from the tested context-aware models (up to +20 percentage points accuracy). However, we note two surprising patterns:
The first observation can be explained by t... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 40 | 21,571 | 22,154 |
28_41 | We speculate that this is due to our more permissive extraction heuristics for it INLINEFORM1 es.
We attribute the second observation to the existence of coreference chains where the preceding sentence contains a pronoun that refers to the same nominal antecedent as the pronoun in the current sentence. Consider the e... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 41 | 22,154 | 22,598 |
28_42 | The nominal antecedent occurs two sentences before the current sentence, but the German sentence in between contains the pronoun sie, which is a useful signal for the context-aware models, even though they cannot know the nominal antecedent.
Note that only models aware of target-side context can benefit from such cir... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 42 | 22,598 | 23,126 |
28_43 |
While such coreference chains complicate the interpretation of the results, we note that improvements on inter-sentential anaphora with antecedent distance INLINEFORM0 are relatively small (compared to distance 1), and that performance is still relatively poor (especially for the minority classes er and sie). We enco... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 43 | 23,126 | 23,761 |
28_44 | We observe an improvement of 5 percentage points from s-hier-to-2 to s-hier-to-2.tied, and 4 percentage points from s-t-hier to s-t-hier.tied. Context encoders introduce a large number of extra parameters, while inter-sentential context is only relevant for a relatively small number of predictions. We hypothesize that... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 44 | 23,761 | 24,207 |
28_45 | Our results also confirm the finding by BIBREF9 that multi-encoder architectures, specifically s-hier-to-2(.tied), can outperform a simple concatenation system in the translation of coreferential pronouns.
The Transformer-based models perform strongest on pronouns with intra-segmental antecedent, outperforming the re... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 45 | 24,207 | 24,794 |
28_46 | However, the Transformer-based concat22 slightly underperforms the RNN-based concat22, and we consider it future research how to better exploit target context with Transformer-based models.
Conclusions
We present a large-scale test suite to specifically test the capacity of NMT models to translate pronouns correctly... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 46 | 24,794 | 25,366 |
28_47 | Even though the increase in BLEU score is moderate for all context-aware models, the improvement in the translation of pronouns is considerable: The best model (s-hier-to-2.tied) achieves a +16 percentage points gain in accuracy over the baseline.
Our experiments confirm the importance of careful architecture design,... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 47 | 25,366 | 25,997 |
28_48 | So far, we have only evaluated models that consider one sentence of context, but the nominal antecedent is more distant for a sizable proportion of the test set, and the evaluation of variable-size context models BIBREF7 , BIBREF10 is interesting future work.
Acknowledgements
We are grateful to the Swiss National Sc... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 48 | 25,997 | 26,475 |
28_49 |
Table 2: Frequency and probability of alignments of it in the training data of our systems (all data from the WMT 2017 news translation task). Alignments are produced by a fast_align model.
Table 3: Example sentence pair with contrastive translations. An antecedent distance of 1 means that the antecedent is in the i... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 49 | 26,475 | 26,922 |
28_50 |
Table 5: English→German BLEU scores on newstest2017, newstest2018 and all sentence pairs from our pronoun test set. Case-sensitive and case-insensitive (uncased) scores are reported. Higher is better, and the best scores are marked in bold.
Table 6: Accuracy on contrastive test set (N=4000 per pronoun) with regard t... | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 50 | 26,922 | 27,378 |
28_51 |
Table 8: Accuracy on contrastive test set with regard to antecedent distance of antecedent (in sentences).
Table 9: Example where 1) antecedent distance is >1 and 2) the context given contains another pronoun as an additional hint. | https://arxiv.org/abs/1810.02268 | A Large-Scale Test Set for the Evaluation of Context-Aware Pronoun Translation in Neural Machine Translation | 51 | 27,378 | 27,612 |
29_0 | Improving Fine-grained Entity Typing with Entity Linking
Fine-grained entity typing is a challenging problem since it usually involves a relatively large tag set and may require to understand the context of the entity mention. In this paper, we use entity linking to help with the fine-grained entity type classificatio... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 0 | 0 | 675 |
29_1 |
Introduction
Given a piece of text and the span of an entity mention in this text, fine-grained entity typing (FET) is the task of assigning fine-grained type labels to the mention BIBREF0. The assigned labels should be context dependent BIBREF1. For example, in the sentence “Trump threatens to pull US out of World ... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 1 | 675 | 1,179 |
29_2 |
This task is challenging because it usually uses a relatively large tag set, and some mentions may require the understanding of the context to be correctly labeled. Moreover, since manual annotation is very labor-intensive, existing approaches have to rely on distant supervision to train models BIBREF0, BIBREF2.
Thu... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 2 | 1,179 | 1,650 |
29_3 | EL is helpful for a model to make typing decisions because if a mention is correctly linked to its target entity, we can directly obtain the type information about this entity in the knowledge base (KB). For example, in the sentence “There were some great discussions on a variety of issues facing Federal Way,” the men... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 3 | 1,650 | 2,238 |
29_4 | In the aforementioned example where “Trump” is the mention, obtaining all the types of Donald Trump in the knowledge base (e.g., politician, businessman, TV personality, etc.) is still informative for inferring the correct type (i.e., politician) that fits the context, since they narrows the possible labels down.
How... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 4 | 2,238 | 2,791 |
29_5 | Thus, in this paper, we propose a deep neural fine-grained entity typing model that flexibly predicts labels based on the context, the mention string, and the type information from KB obtained with EL.
Using EL also introduces a new problem for the training process. Currently, a widely used approach to create FET tra... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 5 | 2,791 | 3,343 |
29_6 | Our approach, when links the mention correctly, also uses all the types of the referred entity in KB as extra information. This may cause the trained model to overfit the weakly labeled data. We design a variant of the hinge loss and introduce noise during training to address this problem.
We conduct experiments on t... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 6 | 3,343 | 3,963 |
29_7 |
Our contributions are summarized as follows:
We propose a deep neural fine-grained entity typing model that utilizes type information from KB obtained through entity linking.
We address the problem that our model may overfit the weakly labeled data by using a variant of the hinge-loss and introducing noise during t... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 7 | 3,963 | 4,466 |
29_8 |
Related Work
An early effort of classifying named entities into fine-grained types can be found in BIBREF4, which only focuses on person names. Latter, datasets with larger type sets are constructed BIBREF5, BIBREF0, BIBREF6. These datasets are more preferred by recent studies BIBREF3, BIBREF7.
Most of the existing... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 8 | 4,466 | 4,982 |
29_9 | Since FET systems usually use distant supervision for training, the labels of the training samples can be noisy, erroneous or overly specific. Several studies BIBREF11, BIBREF12, BIBREF9 address these problems by separating clean mentions and noisy mentions, modeling type correction BIBREF3, using a hierarchy-aware lo... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 9 | 4,982 | 5,473 |
29_10 | But they use EL to help with clustering and type name selection, which is very different from how we use it to improve the performance of a supervised FET model. BIBREF14 finds related entities based on the context instead of directly applying EL. The types of these entities are then used for inferring the type of the... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 10 | 5,473 | 6,033 |
29_11 | Thus, this is a multi-class, multi-label classification problem BIBREF0. Next, we will introduce our approach for this problem in detail, including the neural model, the training of the model, and the entity linking algorithm we use.
Method ::: Fine-grained Entity Typing Model ::: Input
Each input sample to our FET ... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 11 | 6,033 | 6,412 |
29_12 | We denote $w_1,w_2,...,w_n$ as the words in the current sentence, $w_{p_1},w_{p_2},...,w_{p_l}$ as the words in the mention string, where $n$ is the number of words in the sentence, $p_1,...,p_l$ are the indices of the words in the mention string, $l$ is the number of words in the mention string. We also use a set of ... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 12 | 6,412 | 6,811 |
29_13 | It first constructs three representations: context representation, mention string representation, and KB type representation. Note that the KB type representation is obtained from a knowledge base through entity linking and is independent of the context of the mention.
Method ::: Fine-grained Entity Typing Model ::: ... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 13 | 6,811 | 7,386 |
29_14 | Their corresponding word embeddings are fed into two layers of BiLSTMs. Let $\mathbf {h}_m^1$ and $\mathbf {h}_m^2$ be the output of the first and the second layer of BiLSTMs for $w_m$, respectively. We use $\mathbf {f}_c=\mathbf {h}_m^1+\mathbf {h}_m^2$ as the context representation vector. | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 14 | 7,386 | 7,679 |
29_15 |
Method ::: Fine-grained Entity Typing Model ::: Mention String Representation
Let $\mathbf {x}_1,...,\mathbf {x}_l$ be the word embeddings of the mention string words $w_{p_1},...,w_{p_l}$. Then the mention string representation $\mathbf {f}_s=(\sum _{i=1}^l \mathbf {x}_i)/l$.
Method ::: Fine-grained Entity Typing ... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 15 | 7,679 | 8,118 |
29_16 | If the EL algorithm returns an entity, we retrieve the types of of this entity from the KB. We use Freebase as our KB. Since the types in Freebase is different from $T$, the target type set, they are mapped to the types in $T$ with rules similar to those used in BIBREF14. Afterwards, we perform one hot encoding on the... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 16 | 8,118 | 8,625 |
29_17 |
Method ::: Fine-grained Entity Typing Model ::: Prediction
Apart from the three representations, we also obtain the score returned by our entity linking algorithm, which indicates its confidence on the linking result. We denote it as a one dimensional vector $\mathbf {g}$. Then, we get $\mathbf {f}=\mathbf {f}_c\opl... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 17 | 8,625 | 9,038 |
29_18 | $\mathbf {f}$ is then fed into an MLP that contains three dense layers to obtain $\mathbf {u}_m$, out final representation for the current mention sample $m$. Let $t_1,t_2,...,t_k$ be all the types in $T$, where $k=|T|$. We embed them into the same space as $\mathbf {u}_m$ by assigning each of them a dense vector BIBR... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 18 | 9,038 | 9,427 |
29_19 | Then the score of the mention $m$ having the type $t_i\in T$ is calculated as the dot product of $\mathbf {u}_m$ and $\mathbf {t}_i$:
We predict $t_i$ as a type of $m$ if $s(m,t_i)>0$.
Method ::: Model Training
Following existing studies, we also generate training data by using the anchor links in Wikipedia. Each a... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 19 | 9,427 | 9,890 |
29_20 |
Since the KB type representations we use in our FET model are also obtained through mapping Freebase types, they will perfectly match the automatically generated labels for the mentions that are correctly linked (i.e., when the entity returned by the EL algorithm and the target entry of the anchor link are the same).... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 20 | 9,890 | 10,402 |
29_21 | After mapping the Freebase types of Donald Trump to the target tag set, this sample will be weakly annotated as /person/politician, /person/tv_personality, and /person/business, which is exactly the same as the type information (the “Types From KB” in Figure FIGREF4) obtained through EL. Thus, during training, when th... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 21 | 10,402 | 10,924 |
29_22 | For most types of entities such as locations and organizations, it is fine since they usually have the same types in different contexts. But it is problematic for person mentions, as their types can be context dependent.
To address this problem, during training, if a mention is linked to a person entity by our entity... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 22 | 10,924 | 11,534 |
29_23 | This will force the model to still infer the type labels from the context even when the mention is correctly linked, since the KB type representation no longer perfectly match the weak labels.
To make it more flexible, we also propose to use a variant of the hinge loss used by BIBREF16 to train our model:
where $\ta... | https://arxiv.org/abs/1909.12079 | Improving Fine-grained Entity Typing with Entity Linking | 23 | 11,534 | 11,943 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.