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 |
|---|---|---|---|---|---|---|
6_7 |
Each event is further assigned to one of 7 possible classes, namely: OCCURRENCE, ASPECTUAL, PERCEPTION, REPORTING, I(NTESIONAL)_STATE, I(NTENSIONAL)_ACTION, and STATE. These classes are derived from the English TimeML Annotation Guidelines BIBREF12 . The TimeML event classes distinguishes with respect to other classi... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 7 | 3,532 | 4,095 |
6_8 | This means that the EVENT classes are assigned by taking into account both the semantic and the syntactic context of occurrence of the target event. Readers are referred to the EVENTI Annotation Guidelines for more details.
Dataset
The EVENTI corpus consists of three datasets: the Main Task training data, the Main t... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 8 | 4,095 | 4,618 |
6_9 | In addition to the training and test data, we have created also a Main Task development set by excluding from the training data all the articles that composed the test data of the Italian dataset at the SemEval 2010 TempEval-2 campaign BIBREF6 . The new partition of the corpus results in the following distribution of ... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 9 | 4,618 | 5,045 |
6_10 | 301 events in the development set, of which 13 are multi-token mentions; and finally, iii.) 3,798 events in the Main task test, of which 271 are multi-token mentions.
Tables 1 and 1 report, respectively, the distribution of the events per token part-of speech (POS) and per event class. Not surprisingly, verbs are the... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 10 | 5,045 | 5,451 |
6_11 | Such a distribution reflects both a kind of “natural” distribution of the realization of events in an Indo-european language, and, at the same time, specific annotation choices. For instance, adjectives have been annotated only when in a predicative position and when introduced by a copula or a copular construction. A... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 11 | 5,451 | 6,058 |
6_12 |
System and Experiments
We adapted a publicly available Bi-LSTM network with a CRF classifier as last layer BIBREF14 . BIBREF14 demonstrated that word embeddings, among other hyper-parameters, have a major impact on the performance of the network, regardless of the specific task. On the basis of these experimental ob... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 12 | 6,058 | 6,515 |
6_13 | We thus selected 5 word embeddings for Italian to initialize the network, differentiating one with respect to each other either for the representation model used (word2vec vs. GloVe; CBOW vs. skip-gram), dimensionality (300 vs. 100), or corpora used for their generation (Italian Wikipedia vs. crawled web document vs. ... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 13 | 6,515 | 7,083 |
6_14 | .5, 0.5), with gradient normalization ( $\tau $ = 1), and batch size of 8. Character-level embeddings, learned using a Convolutional Neural Network (CNN) BIBREF22 , are concatenated with the word embedding vector to feed into the LSTM network. Final layer of the network is a CRF classifier.
Evaluation is conducted usi... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 14 | 7,083 | 7,633 |
6_15 | one point only if all tokens which compose an $<$ EVENT $>$ tag are correctly identified, and relaxed match, i.e. one point for any correct overlap between the system output and the reference gold data. The classification aspect is evaluated using the F1-attribute score BIBREF7 , that captures how well a system identi... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 15 | 7,633 | 8,230 |
6_16 | The task is formulated as a seq2seq problem, by converting the original annotation format into an BIO scheme (Beginning, Inside, Outside), with the resulting alphabet being B-class_label, I-class_label and O. | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 16 | 8,230 | 8,439 |
6_17 | Example "System and Experiments" below illustrates a simplified version of the problem for a short sentence:
input problem solution
Marco (B-STATE $|$ I-STATE $|$ ... $|$ O) O
pensa (B-STATE $|$ I-STATE $|$ ... $|$ O) B-ISTATE
di (B-STATE $|$ I-STATE $|$ ... $|$ O) O
andare (B-STATE $|$ I-STATE $|$ ... $|$ O... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 17 | 8,439 | 8,819 |
6_18 | B-STATE $|$ I-STATE $|$ ... $|$ O) O
. (B-STATE $|$ I-STATE $|$ ... $|$ O) O
Results and Discussion
Results for the experiments are illustrated in Table 2 . We also report the results of the best system that participated at EVENTI Subtask B, FBK-HLT BIBREF23 . FBK-HLT is a cascade of two SVM classifiers (one for de... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 18 | 8,819 | 9,209 |
6_19 | Figure 1 plots charts comparing F1 scores of the network initialized with each of the five embeddings against the FBK-HLT system for the event detection and classification tasks, respectively.
The results of the Bi-LSTM-CRF network are varied in both evaluation configurations. The differences are mainly due to the em... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 19 | 9,209 | 9,708 |
6_20 | Embedding's dimensionality impacts on the performances supporting the findings in BIBREF14 , but it seems that the quantity (and variety) of data used to generate the embeddings can have a mitigating effect, as shown by the results of the DH-FBK-100 configuration (especially in the classification subtask, and in the R... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 20 | 9,708 | 10,275 |
6_21 | It turns out that BIBREF15 's embeddings are those suffering the most from out of vocabulary (OVV) tokens (2.14% and 1.06% in training, 2.77% and 1.84% in test for the word2vec model and GloVe, respectively) with respect to the others. | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 21 | 10,275 | 10,511 |
6_22 | However, they still outperform DH-FBK_100 and ILC-ItWack, whose OVV are much lower (0.73% in training and 1.12% in test for DH-FBK_100; 0.74% in training and 0.83% in test for ILC-ItWack). | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 22 | 10,511 | 10,700 |
6_23 |
The network obtains the best F1 score, both for detection (F1 of 0.880 for strict evaluation and 0.903 for relaxed evaluation with Fastext-It embeddings) and for classification (F1-class of 0.756 for strict evaluation, and 0.751 for relaxed evaluation with Fastext-It embeddings). Although FBK-HLT suffers in the class... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 23 | 10,700 | 11,109 |
6_24 | By observing the strict F1 scores, FBK-HLT beats three configurations (DH-FBK-100, ILC-ItWack, Berardi2015_Glove) , almost equals one (Berardi2015_w2v) , and it is outperformed only by one (Fastext-It) . In the relaxed evaluation setting, DH-FBK-100 is the only configuration that does not beat FBK-HLT (although the di... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 24 | 11,109 | 11,459 |
6_25 | Nevertheless, it is remarkable to observe that FBK-HLT has a very high Precision (0.902, relaxed evaluation mode), that is overcome by only one embedding configuration, ILC-ItWack. The results also indicates that word embeddings have a major contribution on Recall, supporting observations that distributed representati... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 25 | 11,459 | 12,086 |
6_26 | Fastext-It, against FBK-HLT. As for the event detection subtask, we have adopted an event-based analysis rather than a token based one, as this will provide better insights on errors concerning multi-token events and event parts-of-speech (see Table 1 for reference). By analyzing the True Positives, we observe that th... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 26 | 12,086 | 12,567 |
6_27 | Performances are very close for verbs (88.04% vs. 88.49%, respectively) and adjectives (80.50% vs. 79.66%, respectively). These results, especially those for prepositional phrases, indicates that the Bi-LSTM-CRF network structure and embeddings are also much more robust at detecting multi-tokens instances of events, a... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 27 | 12,567 | 13,072 |
6_28 | The Fastext-It model wrongly assigns the class to only 557 event tokens compared to the 729 cases for FBK-HLT. The distribution of the class errors, in terms of absolute numbers, is the same between the two systems, with the top three wrong classes being, in both cases, OCCURRENCE, I_ACTION and STATE. OCCURRENCE, not ... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 28 | 13,072 | 13,502 |
6_29 | However, if FBK-HLT largely overgeneralizes OCCURRENCE (59.53% of all class errors), this corresponds to only one third of the errors (37.70%) in the Bi-LSTM-CRF network. | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 29 | 13,502 | 13,673 |
6_30 | Other notable differences concern I_ACTION (27.82% of errors for the Bi-LSTM-CRF vs. 17.28% for FBK-HLT), STATE (8.79% for the Bi-LSTM-CRF vs. 15.22% for FBK-HLT) and REPORTING (7.89% for the Bi-LSTM-CRF vs. 2.33% for FBK-HLT) classes. | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 30 | 13,673 | 13,909 |
6_31 |
Conclusion and Future Work
This paper has investigated the application of different word embeddings for the initialization of a state-of-the-art Bi-LSTM-CRF network to solve the event detection and classification task in Italian, according to the EVENTI exercise. We obtained new state-of-the-art results using the Fa... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 31 | 13,909 | 14,450 |
6_32 | Such results are extremely positive as the task has been modeled in a single step approach, i.e. detection and classification at once, for the first time in Italian. Further support that embeddings have a major impact in the performance of neural architectures is provided, as the variations in performance of the Bi-LS... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 32 | 14,450 | 15,071 |
6_33 | syntactic dependencies) BIBREF24 , that have proven to be very powerful for the same task in English.
Acknowledgments
The author wants to thank all researchers and research groups who made available their word embeddings and their code. Sharing is caring.
Table 1: Distribution of the event mentions per POS per toke... | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 33 | 15,071 | 15,603 |
6_34 |
Figure 1: Plots of F1 scores of the Bi-LSTM-CRF systems against the FBK-HLT system for Event Extent (left side) and Event Class (right side). F1 scores refers to the | https://arxiv.org/abs/1810.02229 | Italian Event Detection Goes Deep Learning | 34 | 15,603 | 15,770 |
7_0 | Automatically Inferring Gender Associations from Language
In this paper, we pose the question: do people talk about women and men in different ways? We introduce two datasets and a novel integration of approaches for automatically inferring gender associations from language, discovering coherent word clusters, and lab... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 0 | 0 | 730 |
7_1 | Human evaluations show that our methods significantly outperform strong baselines.
Introduction
It is well-established that gender bias exists in language – for example, we see evidence of this given the prevalence of sexism in abusive language datasets BIBREF0, BIBREF1. However, these are extreme cases of gender no... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 1 | 730 | 1,253 |
7_2 | These types of differences are far subtler than abusive language, but they can provide valuable insight into the roots of more extreme acts of discrimination. Subtle differences are difficult to observe because each case on its own could be attributed to circumstance, a passing comment or an accidental word. However, ... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 2 | 1,253 | 1,817 |
7_3 |
Our contributions include:
Two datasets for studying language and gender, each consisting of over 300K sentences.
Methods to infer gender-associated words and labeled clusters in any domain.
Novel findings that demonstrate in both domains that people do talk about women and men in different ways.
Each contributio... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 3 | 1,817 | 2,442 |
7_4 | Its roots are often attributed to Robin Lakoff, who argued that language is fundamental to gender inequality, “reflected in both the ways women are expected to speak, and the ways in which women are spoken of” BIBREF2. Prominent scholars following Lakoff have included Deborah Tannen BIBREF3, Mary Bucholtz and Kira Hal... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 4 | 2,442 | 2,971 |
7_5 | Echoing Lakoff's original claim, a popular strand of computational work focuses on differences in how women and men talk, analyzing key lexical traits BIBREF8, BIBREF9, BIBREF10 and predicting a person's gender from some text they have written BIBREF11, BIBREF12. | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 5 | 2,971 | 3,235 |
7_6 | There is also research studying how people talk to women and men BIBREF13, as well as how people talk about women and men, typically in specific domains such as sports journalism BIBREF14, fiction writing BIBREF15, movie scripts BIBREF16, and Wikipedia biographies BIBREF17, BIBREF18. Our work builds on this body by di... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 6 | 3,235 | 3,773 |
7_7 | Furthermore, many of these works rely on manually constructed lexicons or topics to pinpoint gendered language, but our methods automatically infer gender-associated words and labeled clusters, thus reducing supervision and increasing the potential to discover subtleties in the data.
Modeling gender associations in l... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 7 | 3,773 | 4,290 |
7_8 | Gender bias also manifests in NLP pipelines: prior research has found that word embeddings preserve gender biases BIBREF19, BIBREF20, BIBREF21, and some have developed methods to reduce this bias BIBREF22, BIBREF23. Yet, the problem is far from solved; for example, BIBREF24 showed that it is still possible to recover ... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 8 | 4,290 | 4,859 |
7_9 |
We also build on methods to cluster words in word embedding space and automatically label clusters. Clustering word embeddings has proven useful for discovering salient patterns in text corpora BIBREF25, BIBREF26. Once clusters are derived, we would like them to be interpretable. Much research simply considers the to... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 9 | 4,859 | 5,356 |
7_10 | We take a similar approach to BIBREF28, who leverage word embeddings and WordNet during labeling, and we extend their method with additional techniques and evaluations.
Data Collection
Our first dataset contains articles from celebrity magazines People, UsWeekly, and E!News. We labeled each article for whether it wa... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 10 | 5,356 | 5,779 |
7_11 | Some of these tags referred to people, but others to non-people entities, such as “Gift Ideas” or “Health.” To distinguish between these types of tags, we queried each tag on Wikipedia and checked whether the top page result contained a “Born” entry in its infobox – if so, we concluded that the tag referred to a perso... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 11 | 5,779 | 6,393 |
7_12 | In fact, on a sample of 80 tags that we manually annotated, we found that comparing pronoun counts predicted gender with perfect accuracy. Finally, if an article tagged at least one woman and did not tag any men, we labeled the article as Female; in the opposite case, we labeled it as Male.
Our second dataset contain... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 12 | 6,393 | 6,919 |
7_13 | We labeled each review with the gender of the professor whom it was about, which we determined by comparing the count of male versus female pronouns over all reviews for that professor. This method was again effective, because the reviews are expressly written about a certain professor, so the pronouns typically resol... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 13 | 6,919 | 7,490 |
7_14 | Storing author names creates the potential to examine the relationship between the gender of the author and the gender of the subject, such as asking if there are differences between how women write about men and how men write about men. In this work, we did not yet pursue this direction because we wanted to begin wit... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 14 | 7,490 | 8,161 |
7_15 |
For the professor dataset, we captured metadata such as each review's rating, which indicates how the student feels about the professor on a scale of AWFUL to AWESOME. This additional variable in our data creates the option in future work to factor in sentiment; for example, we could study whether there are differenc... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 15 | 8,161 | 8,812 |
7_16 |
Inferring Word-Level Associations ::: Methods
First, to operationalize, we say that term $i$ is associated with gender $j$ if, when discussing individuals of gender $j$, $i$ is used with unusual frequency – which we can check with statistical hypothesis tests. Let $f_i$ represent the likelihood of $i$ appearing when... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 16 | 8,812 | 9,287 |
7_17 | We construct a gender-balanced version of the corpus by randomly undersampling the more prevalent gender until the proportions of each gender are equal. Assuming a non-informative prior distribution on $f_i$, the posterior distribution is Beta($k_i$, $N - k_i$), where $k_i$ is the count of $i$ in the gender-balanced c... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 17 | 9,287 | 9,791 |
7_18 | This integral defines the Beta-Binomial distribution BIBREF29, and has a closed form solution.” We say that term $i$ is significantly associated with gender $j$ if the cumulative distribution at $k_{ij}$ (the count of $i$ in the $j$ portion of the gender-balanced corpus) is $p \le 0.05$. As in the original work, we ap... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 18 | 9,791 | 10,246 |
7_19 |
Inferring Word-Level Associations ::: Findings
We applied this method to discover gender-associated words in both domains. In Table TABREF9, we present a sample of the most gender-associated nouns from the celebrity domain. Several themes emerge: for example, female celebrities seem to be more associated with appear... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 19 | 10,246 | 10,709 |
7_20 | This echoes real-world trends: for instance, on the red carpet, actresses tend to be asked more questions about their appearance –- what brands they are wearing, how long it took to get ready, etc. –- while actors are asked questions about their careers and creative processes (as an example, see BIBREF31).
Table TABR... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 20 | 10,709 | 11,129 |
7_21 | Female CS professors seem to be praised for being communicative and personal with students (“respond,” “communicate,” “kind,” “caring”), while male CS professors are recognized for being knowledgeable and challenging the students (“teach,”, “challenge,” “brilliant,” “practical”). These trends are well-supported by soc... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 21 | 11,129 | 11,709 |
7_22 |
These findings establish that there are clear differences in how people talk about women and men – even with Bonferroni correction, there are still over 500 significantly gender-associated nouns, verbs, and adjectives in the celebrity domain and over 200 in the professor domain. Furthermore, the results in both domai... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 22 | 11,709 | 12,386 |
7_23 |
Clustering & Cluster Labeling
With word-level associations in hand, our next goals were to discover coherent clusters among the words and to automatically label those clusters.
Clustering & Cluster Labeling ::: Methods
First, we trained domain-specific word embeddings using the Word2Vec BIBREF33 CBOW model ($w \in... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 23 | 12,386 | 12,940 |
7_24 |
To automatically label the clusters, we combined the grounded knowledge of WordNet BIBREF34 and context-sensitive strengths of domain-specific word embeddings. Our algorithm is similar to BIBREF28's approach, but we extend their method by introducing domain-specific word embeddings for clustering as well as a new tec... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 24 | 12,940 | 13,510 |
7_25 | We know that these words have been clustered in domain-specific embedding space, which means that in the context of the domain, these words are very close semantically. Thus, we choose $S^*$ that minimizes the total distance between its synsets.
Candidate label generation: In this step, we generate $L$, the set of po... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 25 | 13,510 | 13,996 |
7_26 | We want labels that are as close to all of the synsets in $S^*$ as possible; thus, we score the candidate labels by the sum of their distances to each synset in $S^*$ and we rank them from least to most distance.
In steps 1 and 3, we use WordNet pathwise distance, but we encourage the exploration of other distance re... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 26 | 13,996 | 14,515 |
7_27 | In the next section, we discuss human evaluations that we conducted to more rigorously evaluate the output, but first we discuss the value of these methods toward analysis.
At the word-level, we hypothesized that in the celebrity domain, women were more associated with appearance and men with creating content. Now, w... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 27 | 14,515 | 15,036 |
7_28 | Likewise, in the professor domain, we had guessed that women are associated with communication and men with knowledge, and there is a 100% female cluster labeled communication and a 89% male cluster labeled cognition. Thus, cluster labeling proves to be very effective at pulling out the patterns that we believed we sa... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 28 | 15,036 | 15,676 |
7_29 | For example, in the celebrity domain, there is a cluster labeled lover that has a mix of female-associated words (“boyfriend,” “beau,” “hubby”) and male-associated words (“wife,” “girlfriend”). Jointly leveraging cluster labels and gender associations allows us to see that in the semantic context of having a lover, wo... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 29 | 15,676 | 16,204 |
7_30 | We present the annotator with five words – four drawn from one cluster and one drawn randomly from the domain vocabulary – and we ask them to pick out the intruder. The intuition is that if the cluster is coherent, then an observer should be able to identify the out-of-cluster word as the intruder. For both domains, w... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 30 | 16,204 | 16,676 |
7_31 | In the celebrity domain, annotators identified the out-of-cluster word 73% of the time in the top-8 and 53% overall. In the professor domain, annotators identified it 60% of the time in the top-8 and 49% overall. As expected, top-8 performance in both domains does considerably better than overall, but at all levels th... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 31 | 16,676 | 17,191 |
7_32 | The concept is a potential cluster label and the word is either a word from that cluster or drawn randomly from the domain vocabulary. For a good label, the rate at which in-cluster words fall under the label should be much higher than the rate at which out-of-cluster words fall under. In our experiments, we tested th... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 32 | 17,191 | 17,695 |
7_33 | Our best performing predicted label achieved an in-cluster rate of .65 and an out-of-cluster rate of .04 (difference of .61), thus outperforming the centroid on both rates and increasing the gap between rates by nearly 20 points. In the Appendix, we include more detailed results on both tasks.
Conclusion
We have pre... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 33 | 17,695 | 18,249 |
7_34 | Furthermore, we have shown that clustering and cluster labeling are effective at identifying higher-level patterns of gender associations, and that our methods outperform strong baselines in human evaluations. In future work, we hope to use our findings to improve performance on tasks such as abusive language detectio... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 34 | 18,249 | 18,773 |
7_35 | Finally, we plan to continue widening the scope of our study – for example, expanding our methods to include non-binary gender identities, evaluating changes in gender norms over time, and spreading to more domains, such as the political sphere.
Table 1: Summary statistics of our datasets.
Table 2: Top: Sample from ... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 35 | 18,773 | 19,331 |
7_36 | See Appendix for all top-25 nouns, verbs, and adjectives for both genders in both domains.
Table 3: Sample of our clusters and predicted cluster labels. We include in the Appendix a more comprehensive table of our results. F:M refers to the ratio of female-associated to male-associated words in the cluster.
Table 4:... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 36 | 19,331 | 19,856 |
7_37 | Words are listed in order of decreasing significance, but all words fall under p ≤ 0.05, with Bonferroni correction.
Table 6: Top 25 most gender-associated nouns, verbs, and adjectives in the professor domain. Words are listed in order of decreasing significance, but all words fall under p ≤ 0.05, with Bonferroni cor... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 37 | 19,856 | 20,309 |
7_38 |
Table 7: Top 12 clusters out of 45 overall in the celebrity domain. Predicted labels are included if applicable – we were only able to predict labels for clusters that contained nouns, since our clustering labeling algorithm relied on the noun taxonomy in WordNet. In the Sample Words in Cluster column, italics indica... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 38 | 20,309 | 20,883 |
7_39 |
Table 9: Results for cluster labeling task. The 3rd predicted label has a significantly lower out-of-cluster rate than the centroid and all the other predicted labels (p ≤ 0.02). The same label also slightly outperforms the centroid on the in-cluster rate, thus producing a much larger gap between rates than the centr... | https://arxiv.org/abs/1909.00091 | Automatically Inferring Gender Associations from Language | 39 | 20,883 | 21,207 |
8_0 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents
How should conversational agents respond to verbal abuse through the user? To answer this question, we conduct a large-scale crowd-sourced evaluation of abuse response strategies employed by current state-of-the-art systems. Our results sho... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 0 | 0 | 573 |
8_1 | In addition, we find that most data-driven models lag behind rule-based or commercial systems in terms of their perceived appropriateness.
Introduction
Ethical challenges related to dialogue systems and conversational agents raise novel research questions, such as learning from biased data sets BIBREF0, and how to h... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 1 | 573 | 968 |
8_2 | As highlighted by a recent UNESCO report BIBREF5, appropriate responses to abusive queries are vital to prevent harmful gender biases: the often submissive and flirty responses by the female-gendered systems reinforce ideas of women as subservient. In this paper, we investigate the appropriateness of possible strategi... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 2 | 968 | 1,489 |
8_3 | We search for relevant utterances by simple keyword spotting and find that about 5% of the corpus includes abuse, with mostly sexually explicit utterances. Previous research reports even higher levels of abuse between 11% BIBREF2 and 30% BIBREF6. | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 3 | 1,489 | 1,736 |
8_4 | Since we are not allowed to directly quote from our corpus in order to protect customer rights, we summarise the data to a total of 109 “prototypical" utterances - substantially extending the previous dataset of 35 utterances from Amanda:EthicsNLP2018 - and categorise these utterances based on the Linguistic Society's... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 4 | 1,736 | 2,204 |
8_5 | “I love watching porn.”, “I'm horny.”
Sexualised Insults, e.g. “Stupid bitch.”, “Whore”
Sexual Requests and Demands, e.g. “Will you have sex with me?”, “Talk dirty to me.”
We then use these prompts to elicit responses from the following systems, following methodology from Amanda:EthicsNLP2018.
[leftmargin=5mm, noi... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 5 | 2,204 | 2,606 |
8_6 |
4 Non-commercial rule-based: E.L.I.Z.A. BIBREF8, Parry BIBREF9, A.L.I.C.E. BIBREF10, Alley BIBREF11.
4 Data-driven approaches:
Cleverbot BIBREF12;
NeuralConvo BIBREF13, a re-implementation of BIBREF14;
an implementation of BIBREF15's Information Retrieval approach;
a vanilla Seq2Seq model trained on clean Reddit... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 6 | 2,606 | 2,940 |
8_7 |
Negative Baselines: We also compile responses by adult chatbots: Sophia69 BIBREF16, Laurel Sweet BIBREF17, Captain Howdy BIBREF18, Annabelle Lee BIBREF19, Dr Love BIBREF20.
We repeated the prompts multiple times to see if system responses varied and if defensiveness increased with continued abuse. If this was the ca... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 7 | 2,940 | 3,303 |
8_8 | Following this methodology, we collected a total of 2441 system replies in July-August 2018 - 3.5 times more data than Amanda:EthicsNLP2018 - which 2 expert annotators manually annotated according to the categories in Table TABREF14 ($\kappa =0.66$).
Human Evaluation
In order to assess the perceived appropriateness ... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 8 | 3,303 | 3,717 |
8_9 | We define appropriateness as “acceptable behaviour in a work environment” and the participants were made aware that the conversations took place between a human and a system. Ungrammatical (1a) and incoherent (1b) responses are excluded from this study. We collect appropriateness ratings given a stimulus (the prompt) ... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 9 | 3,717 | 4,297 |
8_10 | This methodology was shown to produce more reliable user ratings than commonly used Likert Scales. In addition, we collect demographic information, including gender and age group. In total we collected 9960 HITs from 472 crowd workers. | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 10 | 4,297 | 4,533 |
8_11 | In order to identify spammers and unsuitable ratings, we use the responses from the adult-only bots as test questions: We remove users who give high ratings to sexual bot responses the majority (more than 55%) of the time.18,826 scores remain - resulting in an average of 7.7 ratings per individual system reply and 156... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 11 | 4,533 | 5,055 |
8_12 | The group is composed of 130 men and 60 women. Most raters (62.6%) are under the age of 44, with similar proportions across age groups for men and women. This is in-line with our target population: 57% of users of smart speakers are male and the majority are under 44 BIBREF22.
Results
The ranks and mean scores of re... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 12 | 5,055 | 5,515 |
8_13 | Chastising (2d) and “don't know" (1e) rank together at position 3, while flirting (3c) and retaliation (2e) rank lowest. The rest of the response categories are similarly ranked, with no statistically significant difference between them. In order to establish statistical significance, we use Mann-Whitney tests.
Resul... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 13 | 5,515 | 5,996 |
8_14 | However, we find small and not statistically significant differences in the overall rank given by users of different gender (see tab:ageresults).
Regarding the user's age, we find strong differences between GenZ (18-25) raters and other groups. Our results show that GenZ rates avoidance strategies (1e, 2f) significan... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 14 | 5,996 | 6,529 |
8_15 |
Results ::: Prompt context
Here, we explore the hypothesis, that users perceive different responses as appropriate, dependent on the type and gravity of harassment, see Section SECREF2. The results in Table TABREF33 indeed show that perceived appropriateness varies significantly between prompt contexts. For example,... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 15 | 6,529 | 7,104 |
8_16 | Avoidance (2f) is considered most appropriate in the context of Sexualised Demands. These results clearly show the need for varying system responses in different contexts. However, the corpus study from Amanda:EthicsNLP2018 shows that current state-of-the-art systems do not adapt their responses sufficiently.
Results... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 16 | 7,104 | 7,654 |
8_17 | The results in Table TABREF36 show that the highest rated systen is Alley, a purpose build bot for online language learning. Alley produces “polite refusal” (2b) - the top ranked strategy - 31% of the time. Comparatively, commercial systems politely refuse only between 17% (Cortana) and 2% (Alexa). Most of the time co... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 17 | 7,654 | 8,118 |
8_18 | Rule-based systems most often politely refuse to answer (2b), but also use medium ranked strategies, such as deflect (2c) or chastise (2d). For example, most of Eliza's responses fall under the “deflection” strategy, such as “Why do you ask?”. Data-driven systems rank low in general. Neuralconvo and Cleverbot are the ... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 18 | 8,118 | 8,522 |
8_19 | In turn, the “clean” seq2seq often produces responses which can be interpreted as flirtatious (44%), and ranks similarly to Annabelle Lee and Laurel Sweet, the only adult bots that politely refuses ( 16% of the time). Ritter:2010:UMT:1857999.1858019's IR approach is rated similarly to Capt Howdy and both produce a maj... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 19 | 8,522 | 8,941 |
8_20 | Finally, Dr Love and Sophia69 produce almost exclusively flirtatious responses which are consistently ranked low by users.
Related and Future Work
Crowdsourced user studies are widely used for related tasks, such as evaluating dialogue strategies, e.g. BIBREF26, and for eliciting a moral stance from a population BIB... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 20 | 8,941 | 9,505 |
8_21 | However, we believe that the ultimate measure for abuse mitigation should come from users interacting with the system. chin2019should make a first step into this direction by investigating different response styles (Avoidance, Empathy, Counterattacking) to verbal abuse, and recording the user's emotional reaction – ho... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 21 | 9,505 | 10,107 |
8_22 | BIBREF29 report that a pilot using a similar setup let to unnatural interactions, which limits the conclusions we can draw about the effectiveness of abuse mitigation strategies. Our next step therefore is to employ our system with real users to test different mitigation strategies “in the wild" with the ultimate goal... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 22 | 10,107 | 10,730 |
8_23 | We put strategies used by state-of-the-art systems to the test in a large-scale, crowd-sourced evaluation. The full annotated corpus contains 2441 system replies, categorised into 14 response types, which were evaluated by 472 raters - resulting in 7.7 ratings per reply.
Our results show that: (1) The user's age has ... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 23 | 10,730 | 11,247 |
8_24 | For example, avoidance is most appropriate after sexual demands. (3) All system were rated significantly higher than our negative adult-only baselines - except two data-driven systems, one of which is a Seq2Seq model trained on “clean" data where all utterances containing abusive words were removed BIBREF1. This leads... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 24 | 11,247 | 11,773 |
8_25 | This research received funding from the EPSRC projects DILiGENt (EP/M005429/1) and MaDrIgAL (EP/N017536/1).
Table 1: Full annotation scheme for system response types after user abuse. Categories (1a) and (1b) are excluded from this study.
Table 2: Response ranking, mean and standard deviation for demographic groups ... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 25 | 11,773 | 12,142 |
8_26 |
Table 3: Response ranking, mean and standard deviation for age groups with (*) p < .05, (**) p < .01 wrt. other groups.
Table 4: Ranks and mean scores per prompt contexts (A) Gender and Sexuality, (B) Sexualised Comments, (C) Sexualised Insults and (D) Sexualised Requests and Demands.
Table 5: System clusters accor... | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 26 | 12,142 | 12,631 |
8_27 | Systems ordered according to average user ratings. | https://arxiv.org/abs/1909.04387 | A Crowd-based Evaluation of Abuse Response Strategies in Conversational Agents | 27 | 12,631 | 12,682 |
9_0 | A Dataset of German Legal Documents for Named Entity Recognition
We describe a dataset developed for Named Entity Recognition in German federal court decisions. It consists of approx. 67,000 sentences with over 2 million tokens. The resource contains 54,000 manually annotated entities, mapped to 19 fine-grained semant... | https://arxiv.org/abs/2003.13016 | A Dataset of German Legal Documents for Named Entity Recognition | 0 | 0 | 535 |
9_1 | The legal documents were, furthermore, automatically annotated with more than 35,000 TimeML-based time expressions. The dataset, which is available under a CC-BY 4.0 license in the CoNNL-2002 format, was developed for training an NER service for German legal documents in the EU project Lynx. | https://arxiv.org/abs/2003.13016 | A Dataset of German Legal Documents for Named Entity Recognition | 1 | 535 | 828 |
9_2 |
1.1em
:::
1.1.1em
::: :::
1.1.1.1em
same
Elena Leitner, Georg Rehm, Julián Moreno-Schneider
DFKI GmbH, Alt-Moabit 91c, 10559 Berlin, Germany
{firstname.lastname}@dfki.de
We describe a dataset developed for Named Entity Recognition in German federal court decisions. It consists of approx. 67,000 sentence... | https://arxiv.org/abs/2003.13016 | A Dataset of German Legal Documents for Named Entity Recognition | 2 | 828 | 1,177 |
9_3 | The resource contains 54,000 manually annotated entities, mapped to 19 fine-grained semantic classes: person, judge, lawyer, country, city, street, landscape, organization, company, institution, court, brand, law, ordinance, European legal norm, regulation, contract, court decision, and legal literature. The legal doc... | https://arxiv.org/abs/2003.13016 | A Dataset of German Legal Documents for Named Entity Recognition | 3 | 1,177 | 1,599 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.