1 Motivation

Hate Speech (HS hereafter) has become a widespread phenomenon on social media platforms like Twitter, and automated detection systems are thus required to deal with it. In spite of no universally accepted definition of HS, these messages may express threats, harassment, intimidation or “disparage a person or a group on the basis of some characteristic such as race, color, ethnicity, gender, sexual orientation, nationality, religion, or other characteristic” [26]. HS may have different topical focuses: misogyny, sexism, racism, xenophobia, etc. Which can be referred to as hate speech topics. For each HS topic, hateful content is directed towards specific targets that represent the community (individuals or groups) receiving the hatred.Footnote 1 HS is thus, by definition, target-oriented, and it involves different ways of linguistically expressing hateful content such as references to racial or sexist stereotypes, the use of negative and positive emotions, swearing terms, etc., all of which have to be considered if one is to train effective automated HS detection systems.

Indeed, such systems would be invaluable for a variety of applications, from automated content classification and moderation, to (potentially malicious) community detection and analysis on social media [9].

To that end, various datasets of human-annotated tweets have been proposed, most often using binary generic (e.g., HS/not HS), or multi-label specific schemas (e.g., racism/sexism/neither). Unfortunately, due (in great parts) to the lack of clear consensus on these HS annotation schemas [21], gathering enough data to train models that generalize these concepts effectively is difficult. Various approaches have been proposed to palliate these issues: for example, transfer learning has been successfully used in a variety of NLP settings, in particular thanks to the Transformer architecture [33], which allows to leverage large quantities of unannotated text, by fine-tuning pre-trained models such as BERT [7] on tasks for which annotated data is more sparse, such as HS detection [1, 17, 24, 25].

A complementary type of approach is Multi-Task Learning (MTL) [5, 18, 23], in which one can leverage different tasks and datasets by jointly training a single architecture on multiple objectives at once, sharing all (or parts) of its parameters between them. [32] were the first to showcase how MTL might be used to generalize HS detection models across a variety of datasets, and later on, [16].

Recently, [4] experimented with transferring specific manifestations of hate across HS topics on a varied set of such datasets, showing that MTL could be used to jointly predict both the hatefulness and the topical focus of specific HS instances.

These studies, however, usually consider generic and specific HS datasets as independent (train on one set and test on another) without accounting for common properties shared between both different manifestations of hate, as well as different levels or granularity of annotation. We take here a different perspective and investigate, to our knowledge for the first time, HS detection in a Multi-Level scenario, by answering the following question: Could instances of generic HS be refined with specific labels, using a model jointly trained on these two levels of annotations? To this end, we propose:

  1. 1.

    An easily extensible multitask and multilevel setup designed for HS topic refinement of generic HS instances, based on the T5 architecture [29], which can be used to generate new specific HS labels (see Fig. 1).

  2. 2.

    A qualitative and error analyses of the refined labels produced by this approach, applied to two popular generic HS datasets from the literature.

Fig. 1.
figure 1

Illustration of our topic refinement approach based on the T5 architecture

Full size image

2 Datasets

As our main objective is investigating the problem of transferring knowledge from different datasets, with different annotation granularity and different topical focuses, we leverage six manually annotated HS corpora from previous studies. We selected these datasets as they are freely available to the research community. Among them, two are generic (Davidson [6] and Founta [14]Footnote 2), and four are specific about four different HS topics: misogyny (the Automatic Misogyny Identification (AMI) dataset collection from both IberEval [11] and Evalita [10]), misogyny and xenophobia (the HatEval dataset [2]), and racism and sexism (the Waseem dataset [34]). Each of these HS topics targets either gender (sexism and misogyny) and/or ethnicity, religion or race (xenophobia and racism). In Table 1 we summarize the corpora used in this study.

For the purpose of our experiments, we performed some simplifying split and merge operations on their classes, and their associated labels. For all datasets, we considered the respective’negative’ (i.e., not HS) classes to be equivalent, and used the unified negative-class label “nothing”. In addition, as we are using both generic and specific HS datasets, we merged positives instances from generic datasets in a unified generic class labelled “HS”. The Offensive and Abusive instances were removed from these datasets, as these concepts often co-exist with HS, but without a clear distinction [21, 27].

For the specific HS corpora, we made the simplification of merging the classes related to sexism and misogyny into the single unified label “HS-sexist”. Similarly, we merged racism and xenophobia into the unified label “HS-racist”. These labels are designed with T5’s text-to-text nature in mind (cf. next section): the generic HS label overlaps part of the specific ones, thus a “misprediction” (or more accurately, a partial prediction in this multi-level scenario setup) at training time should only incur a partial error signal (e.g. predicting only “hate speech” in the specific HS task, the correct label being “hate speech - racist”, incurs less error than predicting “nothing”) (see Table 1).

As noted by a number of previous works [12, 13, 20, 21], these types of merging of classes/labels may not be desirable, as each dataset has its own annotation schema. However, as the goal of this work is to explore the viability of HS topic refinement with currently available datasets, we chose to use this simplified annotation schema, and thus consider this added source of label noise to be part of the experimental setting. Addressing these issues, by expanding or reworking this set of labels will likely be explored in future work.

Table 1. General overview of the datasets used in this study.
Full size table

3 Experiments and Cross-Dataset Evaluation

3.1 Models

We rely primarily on a T5 (Text-to-Text Transfer Transformer) architecture [29]. We also experiment with a RoBERTa [19] model, which we use here in an MTL architecture, as a point of comparison for evaluating the performances of these two models across datasets, outside of label refinement (see Sect. 4).

T5 proposes a way to unify text generation and classification tasks in NLP, by reframing all of them as text-to-text problems. This allows the model to both better leverage its pre-training on large quantities of unsupervised text data, but also greatly simplifies MTL setups. Indeed, instead of requiring additional per-task label-space projection layers, the same fine-tuned weights can be used to perform each desired task, which can be indicated to the model by prepending input instances with some task-specific prefix text. MTL with RoBERTa, on the other hand, is traditionally performed by constructing some kind of projection layer (or layers) for each task in the training set, each with their separate target label-space.

We also experimented with BERT-like models which are domain-adapted for HS and toxic language detection, such as fBERT [30], HateBERT [3], or ToxDectRoBERTa [36], but they yielded similar cross-dataset performances, and so to conserve space, we do not present these results.

3.2 Experiments and Results

For the T5 model, we initially experimented with different prefixes and task labels configurations, but settled on “generic HS:” and “specific HS:”, for the generic and specific HS datasets, respectively. In this setup, the model is fine-tuned without task or dataset specific information added, but rather, only the level of HS classification available and/or requested (is HS present or not? vs. which specific topic of HS?). We refer to this particular configuration using unified prefixes as T5-Refine.

To ascertain how well this configuration is able to learn both of these tasks, we perform a comparative evaluation of performance across datasets alongside other configurations, similar but not intended for topic refinement. As such, we also trained our models with MTL architectures as follows.

RoBERTa-MTL: This is a RoBERTa-base classifier, in the “classic” MTL configuration with one dedicated classification layer per task/dataset (a simple linear projection of the [CLS] token; see [7] or [19] for more details), on the same set of multilevel datasets. (output labels: HS/nothing for Davidson & Founta; HS-sexist/ nothing for Evalita & IberEval; HS-racist/HS-sexist/nothing for Waseem & HatEval);

T5-MTL: This is a fine-tuned T5-base model with task-specific prefixes (the names of the corresponding datasets) (output labels: HS/HS-racist/HS-sexist/nothing for all datasets), used here as an intermediate point of comparison between the previous two models (i.e., RoBERTa-MTL and T5-Refine).

Table 2. Comparative evaluation of our models across generic vs. specific HS datasets.
Full size table

We trained T5-Refine on all the training datasets combined (with generic/specific HS task prefixes) while RoBERTa-MTL and T5-MTL models were trained in a multi-task fashion (one head/task prefix per dataset) on the train set of each dataset. Experiments were performed with the AllenNLP [15] and Huggingface Transformers library [35]. Models were trained for a maximum of 12 epochs, with early stopping (patience 4 on validation loss), a batch size of 6, and gradient accumulation of 12. For T5 (RoBERTa) we use the AdaFactor (AdamW) optimizer with a learning rate =1e-3 (1e-5), determined by manual hyperparameter fine-tuning.

Table 2 presents the aggregated averaged results in terms of F-score (F1), precision (P), and recall (R) for the three models when tested on: all generic HS test sets (Davidson and Founta), all specific HS (Waseem, HatEval, Evalita, and IberEval) test sets, and all 6 combined test sets.

Table 3 present a more detailed view of these results, in terms of macro F1-scores only (for conciseness): for clarity, the multi-topic datasets (HatEval and Waseem) have been split into single-topic subsets (HatEval sexist/Waseem sexist and HatEval racist/Waseem racist). Then, for each dataset, “HS” and “not HS” correspond to each respective (sub)set’s relevant binarized HS positive and negative classes (HS[-sexist/-racist]/nothing), alongside the Macro Averaged F1-scores. As can be observed, our HS topic refinement model, T5-Refine, despite training under the most difficult configuration (unified label-space and topic-level merged task prefixes), does not showcase significantly degraded cross-dataset performance, compared to the more task dedicated models.

Table 3. Detailed evaluation results per-dataset (F1-scores).
Full size table

4 Hate Speech Topic Refinement

Using the trained T5-Refine model, we can thus request it to produce specific HS labels for instances of generic HS datasets, here, Davidson and Founta, by simply switching to the specific task prefix at inference time. Table 4 presents a few illustrative examples, of what we consider to be successfully refined labels (examples #1–4), as well as errors (examples #5–9).

To judge the quality of these newly produced labels, we sample 600 instances (200 from each of: [gold = HS | predicted = nothing]; [gold = <any> | predicted = HS - sexist]; [gold = <any> | predicted = HS - racist], where <any> stands for all the possible gold labels) for each of the two generic HS datasets, and compare the predicted labels with the dataset’s gold labels, but also with our own human re-annotationFootnote 3 of those same instances. For both datasets, after manually re-annotating with specific HS labels, the final label was assigned according to a majority vote (at least two annotators always ended up agreeing, so no adjudication was necessary).Footnote 4 For Founta, the re-annotations process shows that in \(\sim \)19% of the cases the instances gold-labelled as “HS” belong to a type of abusive language different from the ones investigated in this paper (e.g., offensive language, reporting/denunciation of hate speech, homophobia, islamohobia, etc.), which were re-annotated as out-of-scope. We obtain similar findings for Davidson, though at a larger scale (\(\sim \)57%). After discarding the instances re-annotated as out-of-scope, we obtained a “soft” agreement (coercing HS - racist and HS - sexist labels as equivalent to the generic HS gold label) with the gold labels of 25% for Founta, and 70% for Davidson. In contrast, the refined HS labels exactly match the human re-annotations in 52% of the in-scope instances for Davidson, and in 44% for Founta. While not perfect, overall, the annotators agree almost twice as often with the model-refined labels than with the gold labels for Founta. For Davidson this agreement instead decreases by 18%.

Qualitatively, we believe the main cause of mis-refinement stems from the significant number of merely offensive or abusive instances having been misannotated as hateful in model’s training data, when they should be distinct according to datasets’ annotation schemes (see last example of Table 4), which is a known problem in HS detection [12, 28]. For example, in Davidson, all the instances containing the substring “b*tch” are gold-labelled as HS, regardless of context of use. After re-annotating, 19% were found to be actually HS - sexist, and 78% out-of-scope (more than 70% offensive). Similarly, the substring “f*g” was gold-labelled as HS, with 88% re-annotated as out-of-scope (mostly offensive, with less than 18% found to be homophobic). This is likely the cause of a number of false positive refined labels, which we argue should not be annotated/refined as HS: for example, reporting of HS, either correctly (#3) or incorrectly refined (#5–6), or offensive language (#8).

Table 4. Examples of refined labels obtained from our approach.
Full size table

Due to our limited unified specific HS labels, the model also struggles with instances containing neither sexist or racist HS (example #7), or those containing multiple simultaneous HS topics (#9): in both cases, a potential solution could be to add training datasets which are annotated for more varied and/or multiple targets per instance, such as [8] for example. Despite those issues, the model was still successful at producing a number of coherent refined labels (examples #1–2), or even “corrected” negative labels for some instances (examples #3–4).

5 Conclusion and Perspectives

In this paper, we show that multilevel and multitask learning for the purpose of topic refinement in HS appears to be a viable way to palliate the relative lack of specific HS annotated data. We experimented with a T5 architecture which presents a number of advantages for future improvements: namely, it is significantly easier to extend after-the-fact, as new tasks and datasets may be further fine-tuned on, without having to modify the model’s architecture to accommodate new labels or levels of annotation. This may enable taking into account other topics of HS, such as homophobia, ableism, etc., which may be present in smaller quantities in generic HS datasets, through the use of Few-Shot learning, for example.