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 |
|---|---|---|---|---|---|---|
12_25 | Our model does not make use of external information except for pre-trained embeddings and is able to outperform the Weka baseline for the development set, but not in the test set. In the shared task, it obtained the 13th place among 22 competitors.
Table 1: Data summary.
Table 2: Summary of the best results.
Table ... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 25 | 12,926 | 13,374 |
13_0 | Multilingual and Multi-Aspect Hate Speech Analysis
Current research on hate speech analysis is typically oriented towards monolingual and single classification tasks. In this paper, we present a new multilingual multi-aspect hate speech analysis dataset and use it to test the current state-of-the-art multilingual mult... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 0 | 0 | 699 |
13_1 | Most internet users involved in a study conducted by the Pew Research Center report having been subjected to offensive name calling online or witnessed someone being physically threatened or harassed online. Additionally, Amnesty International within Element AI have lately reported that many women politicians and jour... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 1 | 699 | 1,303 |
13_2 | Hate speech may not represent the general opinion, yet it promotes the dehumanization of people who are typically from minority groups BIBREF0, BIBREF1 and can incite hate crime BIBREF2.
Moreover, although people of various linguistic backgrounds are exposed to hate speech BIBREF3, BIBREF2, English is still at the ce... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 2 | 1,303 | 1,798 |
13_3 | However, such studies usually use monolingual corpora and do not contrast, or examine the correlations between online hate speech in different languages. On the other hand, tasks involving more than one language such as the hatEval task, which covers English and Spanish, include only separate classification tasks, nam... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 3 | 1,798 | 2,422 |
13_4 | For instance, the hateful tweets presented in Figure FIGREF5 show toxicity directed towards different targets, with or without using slurs, and generating several types of reactions. We believe that, in order to balance between truth and subjectivity, there are at least five important aspects in hate speech analysis. | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 4 | 2,422 | 2,741 |
13_5 | Hence, our annotations indicate (a) whether the text is direct or indirect; (b) if it is offensive, disrespectful, hateful, fearful out of ignorance, abusive, or normal; (c) the attribute based on which it discriminates against an individual or a group of people; (d) the name of this group; and (e) how the annotators ... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 5 | 2,741 | 3,310 |
13_6 | We claim that our multi-aspect annotation schema would provide a valuable insight into several linguistic and cultural differences and bias in hate speech.
We use Amazon Mechanical Turk to label around 13,000 potentially derogatory tweets in English, French, and Arabic based on the above mentioned aspects and, regard... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 6 | 3,310 | 3,925 |
13_7 | We adopt BIBREF8 as a learning algorithm adapted to loosely related tasks such as our five annotated aspects and, use the Babylon cross-lingual embeddings BIBREF10 to align the three languages. We compare the multilingual multitask learning settings with monolingual multitask, multilingual single-task, and monolingual... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 7 | 3,925 | 4,540 |
13_8 |
Related Work
There is little consensus on the difference between profanity and hate speech and, how to define the latter BIBREF17. As shown in Figure FIGREF11, slurs are not an unequivocal indicator of hate speech and can be part of a non-aggressive conversation, while some of the most offensive comments may come in... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 8 | 4,540 | 5,098 |
13_9 |
Given the subjectivity and the complexity of such data, annotation schemes have rarely been made fine-grained. Table TABREF10 compares different labelsets that exist in the literature. For instance, BIBREF12 use racist, sexist, and normal as labels; BIBREF13 label their data as hateful, offensive (but not hateful), a... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 9 | 5,098 | 5,680 |
13_10 | BIBREF15 label their data as offensive, abusive, hateful, aggressive, cyberbullying, spam, and normal. On the other hand, BIBREF20 have chosen to detect ideologies of hate speech counting 40 different hate ideologies among 13 extremist hate groups.
The detection of hate speech targets is yet another challenging aspec... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 10 | 5,680 | 6,286 |
13_11 | We address this false positive bias problem and the common ambiguity of target detection by asking the annotators to label target attributes such as origin, gender, or religious affiliation within 16 named target groups such as refugees, or immigrants.
Furthermore, BIBREF22 have reproduced the experiment of BIBREF12 ... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 11 | 6,286 | 6,907 |
13_12 | BIBREF6 present a dataset of Italian tweets, in which the annotations capture the degree of intensity of offensive and aggressive tweets, in addition to whether the tweets are ironic and contain stereotypes or not. BIBREF2 have collected more than 500 German tweets against refugees, and annotated them as hateful and n... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 12 | 6,907 | 7,515 |
13_13 | BIBREF24 present a dataset of code mixed Hindi-English tweets, while BIBREF25 report the presence of Hindi tokens in English data and use multilingual word embeddings to deal with this issue when detecting toxicity. Similarly, we use such embeddings to take advantage of the multilinguality and comparability of our cor... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 13 | 7,515 | 8,138 |
13_14 |
To fully exploit the collected annotations, we tested multitask learning on our dataset. Multitask learning BIBREF7 allows neural networks to share parameters with one another and, thus, learn from related tasks. It has been used in different NLP tasks such as parsing BIBREF9, dependency parsing BIBREF26, neural mach... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 14 | 8,138 | 8,792 |
13_15 |
Dataset
In this section, we present our data collection methodology and annotation process.
Dataset ::: Data Collection
Considering the cultural differences and commonly debated topics in the main geographic regions where English, French, and Arabic are spoken, searching for equivalent terms in the three languages... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 15 | 8,792 | 9,391 |
13_16 | In fact, we started our data collection by searching for common slurs and demeaning expressions such as “go back to where you come from”. Then, we observed that discussions about controversial topics, such as feminism in general, illegal immigrants in English, Islamo-gauchisme (“Islamic leftism") in French, or Iran in... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 16 | 9,391 | 9,868 |
13_17 |
Dataset ::: Linguistic Challenges
All of the annotated tweets include original tweets only, whose content has been processed by (1) deleting unarguably detectable spam tweets, (2) removing unreadable characters and emojis, and (3) masking the names of mentioned users using @user and potentially enclosed URLs using @... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 17 | 9,868 | 10,404 |
13_18 | Some French tweets also contain Romanized dialectal Arabic tokens generated by, most likely, bilingual North African Twitter users. Hence, although we eliminated most of these tweets in order to avoid misleading the annotators, the possibly remaining ones still added noise to the data.
One more challenge that the ann... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 18 | 10,404 | 11,038 |
13_19 | Therefore, we searched for derogatory terms adapted to different circumstances, and acquired an Arabic corpus that combines tweets written in MSA and Arabic dialects. For instance, the tweet shown in Figure FIGREF5 contains a dialectal slur that means “maiden.”
Dataset ::: Annotation Process
We rely on the general p... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 19 | 11,038 | 11,653 |
13_20 |
Our annotation guidelines explained the fact that offensive comments and hate do not necessarily come in the form of profanity. Since different degrees of discrimination work on the dehumanization of individuals or groups of people in distinct ways, we chose not to annotate the tweets within two or three classes. For... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 20 | 11,653 | 12,353 |
13_21 |
Dataset ::: Annotation Process ::: Avoiding scams
To prevent scams, we also prepared three annotation guideline forms and three aligned labelsets written in English, French, and Modern Standard Arabic with respect to the language of the tweets to be annotated.
We requested native speakers to annotate the data and c... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 21 | 12,353 | 12,928 |
13_22 | Since the rejection affects the reputation of the annotators and their chances to get new tasks on Amazon Mechanical Turk, well-reputed annotators are usually reliable. We have divided our corpora into smaller batches on Amazon Mechanical Turk in order to facilitate the analysis of the annotations of the workers and, ... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 22 | 12,928 | 13,508 |
13_23 |
Dataset ::: Pilot Dataset
We initially put samples of 100 tweets in each of the three languages on Amazon Mechanical Turk. | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 23 | 13,508 | 13,633 |
13_24 | We showed the annotators the tweet along with lists of labels describing (a) whether it is direct or indirect hate speech; (b) if the tweet is dangerous, offensive, hateful, disrespectful, confident or supported by some URL, fearful out of ignorance, or other; (c) the target attribute based on which it discriminates a... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 24 | 13,633 | 14,167 |
13_25 | (d) the name of its target group, and (e) whether the annotators feel anger, sadness, fear or nothing about the tweets.
Each tweet has been labeled by three annotators. We have provided them with additional text fields to fill in with labels or adjectives that would (1) better describe the tweet, (2) describe how the... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 25 | 14,167 | 14,582 |
13_26 | We kept the most commonly used labels from our initial labelset, took off some of the initial class names and added frequently introduced labels, especially the emotions of the annotators when reading the tweets and the names of the target groups. For instance, after this step, we have ended up merging race, ethnicity... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 26 | 14,582 | 15,127 |
13_27 |
Dataset ::: Final Dataset
The final dataset is composed of a pilot corpus of 100 tweets per language, and comparable corpora of 5,647 English tweets, 4,014 French tweets, and 3,353 Arabic tweets. Each of the annotated aspects represents a classification task of its own, that could either be evaluated independently, ... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 27 | 15,127 | 15,515 |
13_28 | The different labels are designed to facilitate the study of the correlations between the explicitness of the tweet, the type of hostility it conveys, its target attribute, the group it dehumanizes, how different people react to it, and the performance of multitask learning on the five tasks. We assigned each tweet to... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 28 | 15,515 | 16,181 |
13_29 | If there are two annotators agreeing on two labels respectively, we add both labels to the annotation.
The average Krippendorff scores for inter-annotator agreement (IAA) are 0.153, 0.244, and 0.202 for English, French, and Arabic respectively, which are comparable to existing complex annotations BIBREF6 given the na... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 29 | 16,181 | 16,658 |
13_30 |
Dataset ::: Final Dataset ::: Directness label
Annotators determine the explicitness of the tweet by labeling it as direct or indirect speech. This should be based on whether the target is explicitly named, or less easily discernible, especially if the tweet contains humor, metaphor, or figurative speech. Table TABR... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 30 | 16,658 | 17,131 |
13_31 |
Dataset ::: Final Dataset ::: Hostility type
To identify the hostility type of the tweet, we stick to the following conventions: (1) if the tweet sounds dangerous, it should be labeled as abusive; (2) according to the degree to which it spreads hate and the tone its author uses, it can be hateful, offensive or disre... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 31 | 17,131 | 17,671 |
13_32 | Table TABREF20 shows that hostility types are relatively consistent across different languages and offensive is the most frequent label.
Dataset ::: Final Dataset ::: Target attribute
After annotating the pilot dataset, we noticed common misconceptions regarding race, ethnicity, and nationality, therefore we merged ... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 32 | 17,671 | 18,253 |
13_33 | Table TABREF20 shows there are fewer tweets targeting disability in Arabic compared to English and French and no tweets insulting people based on their sexual orientation which may be due to the fact that the labels of gender, gender identity, and sexual orientation use almost the same wording. On the other hand, Fren... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 33 | 18,253 | 18,878 |
13_34 |
Dataset ::: Final Dataset ::: Target group
We determined 16 common target groups tagged by the annotators after the first annotation step. The annotators had to decide on whether the tweet is aimed at women, people of African descent, Hispanics, gay people, Asians, Arabs, immigrants in general, refugees; people of d... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 34 | 18,878 | 19,460 |
13_35 | In case the tweet targets more than one group of people, the annotators should choose the group which would be the most affected by it according to them. Table TABREF10 shows the counts of the five categories out of 16 that commonly occur in the three languages. In fact, most of the tweets target individuals or fall i... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 35 | 19,460 | 19,996 |
13_36 | English tweets tend to have more tweets targeting people with special needs, due to common language-specific demeaning terms used in conversations where people insult one another. Arabic tweets contain more hateful comments towards women for the same reason. On the other hand, the French corpus contains more tweets th... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 36 | 19,996 | 20,656 |
13_37 | After collecting the annotation results of the pilot dataset regarding how people feel about the tweets, and observing the added categories, we adopted a range of sentiments that are in the negative and neutral scales of the hourglass of emotions introduced by BIBREF29. This model includes sentiments that are connecte... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 37 | 20,656 | 21,274 |
13_38 |
Table TABREF20 shows more tweets making the annotators feel disgusted and angry in English, while annotators show more indifference in both French and Arabic. A relatively more frequent label in both French and Arabic is shock, therefore reflecting what some of the annotators were feeling during the labeling process.... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 38 | 21,274 | 21,841 |
13_39 |
Experiments ::: Models
We compare both traditional baselines using bag-of-words (BOW) as features on Logistic regression (LR), and deep learning based methods.
For deep learning based models, we run bidirectional LSTM (biLSTM) models with one hidden layer on each of the classification tasks. Deeper BiLSTM models pe... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 39 | 21,841 | 22,342 |
13_40 |
We test different models, namely single task single language (STSL), single task multilingual (STML), and multitask multilingual models (MTML) on our dataset. In multilingual settings, we tested Babylon multilingual word embeddings BIBREF10 and MUSE BIBREF30 on the different tasks. We use Babylon embeddings since the... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 40 | 22,342 | 22,702 |
13_41 |
Sluice networks BIBREF8 learn the weights of the neural networks sharing parameters (sluices) jointly with the rest of the model and share an embedding layer, Babylon embeddings in our case, that associates the elements of an input sequence. We use a standard 1-layer BiLSTM partitioned into two subspaces, a shared su... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 41 | 22,702 | 23,227 |
13_42 | The hidden layer has a dimension of 200, the learning rate is initially set to 0.1 with a learning rate decay, and we use the DyNet BIBREF31 automatic minibatch function to speed-up the computation. We initialize the cross-stitch unit to imbalanced, set the standard deviation of the Gaussian noise to 2, and use simple... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 42 | 23,227 | 23,716 |
13_43 | We use the dev set to tune the threshold for each binary classification problem in the multilabel classification settings of each task.
Experiments ::: Results and Analysis
We report both the micro and macro-F1 scores of the different classification tasks in Tables TABREF27 and TABREF28. Majority refers to labeling ... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 43 | 23,716 | 24,219 |
13_44 |
Experiments ::: Results and Analysis ::: STSL
STSL performs the best among all models on the directness classification, and it is also consistent in both micro and macro-F1 scores. This is due to the fact that the directness has only two labels and multilabeling is not allowed in this task. Tasks involving imbalance... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 44 | 24,219 | 24,646 |
13_45 |
Since macro-F1 is the average of all F1 scores of individual labels, all deep learning models have high macro-F1 scores in English which indicates that they are particularly good at classifying the direct class. STSL is also comparable or better than traditional BOW feature-based classifiers when performed on other t... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 45 | 24,646 | 25,209 |
13_46 | When we jointly train each task on the three languages, the performance decreases in most cases, other than the target group classification tasks. This may be due to the difference in label distributions across languages. Yet, multilingual training of the target group classification task improves in all languages. Sin... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 46 | 25,209 | 25,899 |
13_47 |
Experiments ::: Results and Analysis ::: MTML
MTML settings do not lead to a big improvement which may be due to the class imbalance, multilabel tasks, and the difference in the nature of the tasks. In order to inspect which tasks hurt or help one another, we trained multilingual models for pairwise tasks such as (g... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 47 | 25,899 | 26,378 |
13_48 | We noticed that when trained jointly, the target attribute slightly improves the performance of the tweet's hostility type classification by 0.03,0.05 and 0.01 better than the best reported scores in English, French, and Arabic, respectively. When target groups and attributes are trained jointly, the macro F-score of ... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 48 | 26,378 | 26,883 |
13_49 | We believe that we can take advantage of the correlations between target attributes and groups along with other tasks, to set logic rules and develop better multilingual and multitask settings.
Conclusion
In this paper, we presented a multilingual hate speech dataset of English, French, and Arabic tweets. We analyze... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 49 | 26,883 | 27,488 |
13_50 | Multilingual multitask learning also helped tasks where each label had less annotated data associated with it. Better tuned deep learning settings in our multilingual and multitask models would be expected to outperform the existing state-of-the-art embeddings and algorithms applied to our data. The different annotati... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 50 | 27,488 | 28,153 |
13_51 | 26206717) from Research Grants Council in Hong Kong, and by postgraduate studentships from the Computer Science and Engineering department of the Hong Kong University of Science and Technology.
Figure 1: Annotation examples in our dataset.
Figure 2: Three tweets in which (1) the first one accuses immigrants of harmi... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 51 | 28,153 | 28,715 |
13_52 |
Table 1: Comparative table of some of the available hate speech and abusive language corpora in terms of labels and sizes.
Table 2: The label distributions of each task. The counts of direct and indirect hate speech include all tweets except those that are single labeled as “normal”. Tweet and annotator’s sentiment ... | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 52 | 28,715 | 29,321 |
13_53 |
Table 4: Full evaluation of tasks where multilingual and multitask models outperform on average single task single language model on four different tasks. | https://arxiv.org/abs/1908.11049 | Multilingual and Multi-Aspect Hate Speech Analysis | 53 | 29,321 | 29,477 |
14_0 | Semantic Web for Machine Translation: Challenges and Directions
A large number of machine translation approaches have recently been developed to facilitate the fluid migration of content across languages. However, the literature suggests that many obstacles must still be dealt with to achieve better automatic translat... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 0 | 0 | 755 |
14_1 | Moreover, our research suggests that while Semantic Web technologies can enhance the quality of machine translation outputs for various problems, the combination of both is still in its infancy.
Introduction
Alongside increasing globalization comes a greater need for readers to understand texts in languages foreign ... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 1 | 755 | 1,417 |
14_2 |
However, translation is a difficult task due to the complexity of natural languages and their structure BIBREF0 . In addition, manual translation does not scale to the magnitude of the Web. One remedy for this problem is MT. The main goal of MT is to enable people to assess content in languages other than the languag... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 2 | 1,417 | 2,037 |
14_3 | Recently, Popović BIBREF3 has classified five types of errors that still remain in MT systems. According to research, the two main faults that are responsible for 40% and 30% of problems respectively, are reordering errors and lexical and syntactic ambiguity. Thus, addressing these barriers is a key challenge for mode... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 3 | 2,037 | 2,598 |
14_4 | However, RBMT systems have a critical drawback in their reliance on manually crafted rules, thus making the development of new translation modules for different languages even more difficult.
SMT and EBMT were developed to deal with the scalability issue in RBMT BIBREF4 , a necessary characteristic of MT systems that... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 4 | 2,598 | 3,262 |
14_5 | Nowadays, a novel SMT paradigm has arisen called NMT which relies on NN algorithms. NMT has been achieving impressive results and is now the state-of-the-art in MT approaches. However, NMT is still a statistical approach sharing some semantic drawbacks from other well-defined SMT approaches BIBREF5 .
One possible sol... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 5 | 3,262 | 3,780 |
14_6 | It is believed that explicit semantic knowledge made available through these technologies can empower MT systems to supply translations with significantly better quality while remaining scalable. In particular, the disambiguated knowledge about real-world entities, their properties and their relationships made availab... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 6 | 3,780 | 4,479 |
14_7 |
Related Works
The idea of using a structured KB in MT systems started in the 90s with the work of Knight and Luk BIBREF7 . Still, only a few researchers have designed different strategies for benefiting of structured knowledge in MT architectures BIBREF8 . Recently, the idea of using KG into MT systems has gained re... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 7 | 4,479 | 4,978 |
14_8 | In summary, they create additional training data and apply a post-editing technique, which replaces the OOV words while querying BabelNet. Shi et al. BIBREF11 have recently built a semantic embedding model reliant upon a specific KB to be used in NMT systems. The model relies on semantic embeddings to encode the key i... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 8 | 4,978 | 5,528 |
14_9 | This mapping results in a semantic embedding model containing KB triples, which are responsible for gathering the key information of each word in the sentences.
Open MT Challenges
The most problematic unresolved MT challenges, from our point of view, which are still experienced by the aforementioned MT approaches ar... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 9 | 5,528 | 6,191 |
14_10 | (2) The limitations of SMT approaches for translating across domains. Most MT systems exhibit good performance on law and the legislative domains due to the large amount of data provided by the European Union. In contrast, translations performed on sports and life-hacks commonly fail, because of the lack of training d... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 10 | 6,191 | 6,726 |
14_11 | This challenge shares the same problem with the first one, namely that most research work focuses on English as one of the involved languages. Therefore, MT systems which translate content between, for instance, Arabic and Spanish are rare. (5) For the speech translation task, the parallel data for training differs wi... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 11 | 6,726 | 7,198 |
14_12 | Since NMT has shown impressive results on reordering, the main problem turns out to be the disambiguation process (both syntactically and semantically) in SMT approaches BIBREF0 .
Suggestions and Possible Directions using SW
Based on the surveyed works on our research BIBREF6 , SWT have mostly been applied at the se... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 12 | 7,198 | 7,795 |
14_13 | They may be applied in different ways according to each MT approach.
Disambiguation. Human language is very ambiguous. Most words have multiple interpretations depending on the context in which they are mentioned. In the MT field, WSD techniques are concerned with finding the respective meaning and correct translatio... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 13 | 7,795 | 8,345 |
14_14 | Although the ambiguity problem has been lessened significantly since the contribution of Carpuat and subsequent works BIBREF13 , this problem still remains a challenge. As seen in Moussallem et al. BIBREF6 , MT systems still try to resolve this problem by using domain specific language models to prefer domain specific... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 14 | 8,345 | 8,912 |
14_15 | However, SWT were applied in two ways to support the semantic disambiguation in MT. First, the ambiguous words were recognized in the source text before carrying out the translation, applying a pre-editing technique. Second, SWT were applied to the output translation in the target language as a post-editing technique.... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 15 | 8,912 | 9,494 |
14_16 |
The real benefit of SW comes from its capacity to provide unseen knowledge about emergent data, which appears every day. Therefore, we suggest performing the topic-modelling technique over the source text to provide a necessary context before translation. Instead of applying the topic-modeling over the entire text, w... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 16 | 9,494 | 10,091 |
14_17 | For every word which has more than one translation, a SPARQL query would be required to find the best combination in the current context. Thus, at the translation phase, the disambiguation algorithm could search for an appropriate word using different SW resources such as DBpedia, in consideration of the context provi... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 17 | 10,091 | 10,638 |
14_18 |
On the other hand, there is also a syntactic disambiguation problem which as yet lacks good solutions. For instance, the English language contains irregular verbs like “set” or “put”. Depending on the structure of a sentence, it is not possible to recognize their verbal tense, e.g., present or past tense. Even statis... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 18 | 10,638 | 11,085 |
14_19 | Although this challenge has successfully been dealt with since NMT has been used for European languages, implementations of NMT for some non-European languages have not been fully exploited (e.g., Brazilian Portuguese, Latin-America Spanish, Zulu, Hindi) due to the lack of large bilingual data sets on the Web to be tr... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 19 | 11,085 | 11,706 |
14_20 | Thus, the verb “put", which is represented by a predicate that groups essential information about the verbal tense, may support the generation step of a given MT system. This sentence usually fails when translated to rich morphological languages, such as Brazilian-Portuguese and Arabic, for which the verb influences t... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 20 | 11,706 | 12,329 |
14_21 | Therefore, this run-time deficiency needs to be addressed or minimized before implementing reasoners successfully into MT systems.
Named Entities. Most NERD approaches link recognized entities with database entries or websites. This method helps to categorize and summarize text, but also contributes to the disambigua... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 21 | 12,329 | 12,805 |
14_22 | For instance, the word “Kiwi" is a family name in New Zealand which comes from the Māori culture, but it also can be a fruit, a bird, or a computer program. Named Entities are a common and difficult problem in both MT (see Koehn BIBREF0 ) and SW fields. The SW achieved important advances in NERD using structured data ... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 22 | 12,805 | 13,250 |
14_23 | In MT systems, however, this problem is directly related to the ambiguity problem and therefore has to be resolved in that wider context.
Although MT systems include good recognition methods, they still need improvement. When an MT system does not recognize an entity, the translation output often has poor quality, im... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 23 | 13,250 | 13,890 |
14_24 | For instance, in NMT, the idea is to include in the training set for the aforementioned word “Kiwi", “Kiwi.animal.link, Kiwi.person.link, Kiwi.food.link" then finally to align them with the translations in the target text. For example, in SMT, the additional information can be included by XML or by an additional model... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 24 | 13,890 | 14,376 |
14_25 | This idea is supported by BIBREF11 where the authors encoded the types of entities along with the words to improve the translation of sentences between Chinese-English. Recently, Moussallem et al. BIBREF15 have shown promising results by applying a multilingual entity linking algorithm along with knowledge graph embed... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 25 | 14,376 | 14,818 |
14_26 | Their approach achieved significant and consistent improvements of +3 BLEU, METEOR and CHRF3 on average on the newstest datasets between 2014 and 2018 for WMT English-German translation task.
Non-standard speech. The non-standard language problem is a rather important one in the MT field. Many people use the colloqui... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 26 | 14,818 | 15,198 |
14_27 | Thus, when MT systems are applied on this context, the input text frequently contains slang, MWE, and unreasonable abbreviations such as “Idr = I don't remember.” and “cya = see you”. Additionally, idioms contribute to this problem, decreasing the translation quality. Idioms often have an entirely different meaning th... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 27 | 15,198 | 15,736 |
14_28 | Some SMT systems like Google or Bing have recognition patterns over non-standard speech from old translations through the Web using SMT approaches. In rare cases SMT can solve this problem, but considering that new idiomatic expressions appear every day and most of them are isolated sentences, this challenge still rem... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 28 | 15,736 | 16,373 |
14_29 | These ontologies have properties which would help identify the birth place or the interests of a given user. For instance, the properties foaf:interest and sioc:topic can be used to describe a given person's topics of interest. If the person is a computer scientist and the model contains topics such as “Information Te... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 29 | 16,373 | 16,881 |
14_30 | Assuming that a user is located in a certain part of Russia and he is reading an English web page which contains some idioms, this property may be used to gather appropriate translations of idioms from English to Russian using a given RDF KB. Therefore, an MT system can be adapted to a user by using specific data abou... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 30 | 16,881 | 17,390 |
14_31 | The dataset contains idioms in 5 languages and are represented by knowledge graphs which facilitates the retrieval and inference of translations among the idioms.
Translating KBs. According to our research, it is clear that SWT may be used for translating KBs in order to be applied in MT systems. For instance, some c... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 31 | 17,390 | 17,975 |
14_32 | For example, the terms contained in triples would be translated to a given target language using a dictionary containing domain words. This dictionary may be acquired in two different ways. First, by performing localisation, as in the work by J. P. McCrae BIBREF17 which translates the terms contained in a monolingual ... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 32 | 17,975 | 18,547 |
14_33 | BIBREF18 , which creates bilingual embeddings using RDF based on Word2vec algorithms. Therefore, we suggest investigating an MT approach mainly based on SWT using NN for translating KBs. Once the KBs are translated, we suggest including them in the language models for improving the translation of entities.
Besides C.... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 33 | 18,547 | 19,189 |
14_34 | Thus, they translated two KBs belonging to medical and financial domains, along with the English Wikipedia, to German. Once translated, the KBs were used as external resources in the translation of German-English. The results were quite appealing and the further research into this area should be undertaken. Recently, ... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 34 | 19,189 | 19,736 |
14_35 | The authors applied their approach on the German DBpedia with the German translation of the English DBpedia on two tasks: fact checking and entity linking. THOTH showed promising results with a translation accuracy of 88.56 while being capable of improving two NLP tasks with its enriched-German KG .
conclusion
In th... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 35 | 19,736 | 20,340 |
14_36 | Considering the decision power of SWT, they cannot be ignored by future MT systems. As a next step, we intend to continue elaborating a novel MT approach which is capable of simultaneously gathering knowledge from different SW resources and consequently being able to address the ambiguity of named entities and also co... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 36 | 20,340 | 20,875 |
14_37 | Additionally, future works that can be expected from fellow researchers, include the creation of multilingual linguistic ontologies describing the syntax of rich morphologically languages for supporting MT approaches. Also, the creation of more RDF multilingual dictionaries which can improve some MT steps, such as ali... | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 37 | 20,875 | 21,373 |
14_38 | 19F2028A) as well as by the Brazilian National Council for Scientific and Technological Development (CNPq) (no. 206971/2014-1) | https://arxiv.org/abs/1907.10676 | Semantic Web for Machine Translation: Challenges and Directions | 38 | 21,373 | 21,500 |
15_0 | Advancing Speech Recognition With No Speech Or With Noisy Speech
In this paper we demonstrate end to end continuous speech recognition (CSR) using electroencephalography (EEG) signals with no speech signal as input. An attention model based automatic speech recognition (ASR) and connectionist temporal classification (... | https://arxiv.org/abs/1906.08871 | Advancing Speech Recognition With No Speech Or With Noisy Speech | 0 | 0 | 575 |
15_1 | In BIBREF0 we demonstrated deep learning based automatic speech recognition (ASR) using EEG signals for a limited English vocabulary of four words and five vowels. In this paper we extend our work for a much larger English vocabulary and we use state-of-art end-to-end continuous speech recognition models to perform re... | https://arxiv.org/abs/1906.08871 | Advancing Speech Recognition With No Speech Or With Noisy Speech | 1 | 575 | 1,212 |
15_2 | This limits technology accessibility to people with speaking disabilities and disorders. The research work presented in this paper tries to address this issue by investigating speech recognition using only EEG signals with no acoustic input and also by combining EEG features along with traditional acoustic features to... | https://arxiv.org/abs/1906.08871 | Advancing Speech Recognition With No Speech Or With Noisy Speech | 2 | 1,212 | 1,987 |
15_3 | Inspired from the unique robustness to environmental artifacts exhibited by the human auditory cortex BIBREF1 , BIBREF2 we used very noisy speech data for this work and demonstrated lower word error rate (WER) for smaller corpus using EEG features, concatenation of EEG features and acoustic features.
In BIBREF3 autho... | https://arxiv.org/abs/1906.08871 | Advancing Speech Recognition With No Speech Or With Noisy Speech | 3 | 1,987 | 2,504 |
15_4 | In BIBREF4 authors perform envisioned speech recognition using random forest classifier but in our case we use end to end state of art models and perform recognition for noisy speech. In BIBREF5 authors demonstrate speech recognition using electrocorticography (ECoG) signals, which are invasive in nature but in our wo... | https://arxiv.org/abs/1906.08871 | Advancing Speech Recognition With No Speech Or With Noisy Speech | 4 | 2,504 | 2,958 |
15_5 | In BIBREF6 the authors used classification approach for identifying phonological categories in imagined and silent speech but in our work we used continuous speech recognition state of art models and our models were predicting words, characters at each time step. Similarly in BIBREF7 neural network based classificatio... | https://arxiv.org/abs/1906.08871 | Advancing Speech Recognition With No Speech Or With Noisy Speech | 5 | 2,958 | 3,553 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.