Fighting hate speech through automatic counter-narrative generation is gaining interest because of the increasing capabilities of Large Language Models. However, counter-narrative generation is a challenging task that can bene t from insightful analyses of text. In this work, we present an approach to improve the generation of counter-narratives by providing Large Language Models with high-quality examples. In addition, we show that enhancing the original hate speech with an argumentative analysis, identifying justi cations and conclusions, together with collectives and the properties associated to them, seems to produce some improvements, specially with with smaller training datasets, helping to orient the generation towards a particular response strategy. The dataset of counter-narratives with argumentative information is made publicly available. Warning: This work contains o ensive and hateful text that may be distressing. It does not represent the views of the authors.
In social media platforms, hate speech is ampli ed beyond human scale, spreading faster and increasing their reach, with negative impacts in societies, like polarization or an increase in violent episodes against targeted communities or individuals. It is because of these known consequences that many legal systems typify it as a crime, at least in some of its forms.
The predominant strategy adopted so far to counter hate speech in social media is to recognize,
block and delete these messages and/or the users that generated it. This strategy has two main
disadvantages. The rst one is that blocking and deleting may prevent a hate message from
spreading, but does not counter its consequences on those who were already reached by it.
The second one is that there is no place for subtleties or shades while de ning hate speech: it
must be done as a binary classi cation because the consequence of that classi cation is binary.
This can generate accusations of overblocking or censorship, and not just because of errors
in automated systems, which have been shown to be highly biased [
An alternative to blocking that has been gaining attention in the last years, is to "oppose
hate content with counter-narratives (i.e. informed textual responses)" [
However, the huge volume of online hate messages makes the manual generation of counternarratives an impossible task. In this scenario, automating the generation of counter-narratives is an appealing avenue, but the task poses a great challenge due to the complex linguistic and communicative patterns involved in argumentation.
Traditional machine learning approaches have typically produced less than satisfactory
results for argumentation mining and generation. However, the recent availability of Large
Language Models (LLMs) provides a promising approach to address the task of counter-narrative
generation. Indeed, LLMs seem capable of generating satisfactory text for many tasks. Thorburn
and Kruger [
However, as Hinton and Wagemans [
In this paper we present an initial exploration of the impact of argumentative information in improving the quality of arguments generated by LLMs, more concretely, in improving the quality of automatically generated counter-narratives against hate speech. We compare di erent scenarios: LLMs without any speci c adaptation to the task or domain, with ne-tuning using a dataset of counter-narratives, in a few-shot approach, and providing additional information about some of the argumentative aspects of the hate speech.
To assess the quality of the counter-narratives generated in the di erent scenarios, we carry
out a preliminary evaluation with human judges, who achieved moderate agreement between
each other. Based on those judgements, we can say that argumentative information by itself
does not produce an improvement in the counter-narratives, but high-quality, speci cally
targeted ne-tuning seems to have a positive impact. Argumentative information does produce
improvements in scenarios with very small training data and very speci c ne-tuning, which
seems promising to produce highly tailored counter-narratives, as in Gupta et al. [
The rest of the paper is organized as follows. In the next section, we review relevant work related to automated counter-narrative generation and argumentative analysis of hate speech. Then in Section 3 we describe our dataset of counter-narratives, with which we carry out the 1No Hate Speech Movement Campaign: http://www. nohatespeechmovement.org/ comparison of scenarios described in Section 4, where we also describe extensively our approach to the evaluation of generated counter-narratives, based on human judgements, and the prompts used to obtain the counter-narratives. Results analyzed in Section 5 show how ne-tuned LLMs and argumentative information provide better results, which we illustrate with some examples.
Automated counter-narrative generation has been recently tackled by leveraging the rapid advances in neural natural language generation. As with most natural language generation tasks in recent years, the basic machine learning approach has been to train or ne-tune a generative neural network with examples speci c to the target task.
The CONAN dataset [
Similar approaches were carried out by Qian et al. [
In follow-up work to the seminal CONAN work, Tekiro lu et al.[
Ashida and Komachi [
However, none of the aforementioned datasets or approaches to counter-narrative generation includes or integrates any additional annotated information apart from the hate message, possibly its context, and its response. That is why we consider an alternative approach that aims to reach generalization not by the sheer number of examples, but by providing a richer analysis of such examples that guides the model in nding adequate generalizations. We believe that information about the argumentative structure of hate speech, may be used as constraints for automatic counter-narrative generation.
Chung et al. [
To our knowledge, ours is the only corpus where tweets of hate speech have been annotated
with argumentative information: ASOHMO [
Property that is assigned to such Collective.
In this work, we present counter-narratives manually associated to the hate tweets in ASOHMO and present an initial exploration of the impact of di erent kinds of information (counter-narratives, counter-narratives by subtype and information about argumentative components) in improving automatic generation of counter-narratives.
Here we present CONEAS (Counter-Narratives Exploiting Argumentative Structure), a dataset
of counter-narratives de ned according to the argumentative information labeled on tweets
from ASOHMO [
All counter-narratives, regardless of their type, also follow the guidelines of the Get The Trolls Out project2: don’t be aggressive or abusive, don’t spread hate yourself, try to de-escalate the conversation, respond thinking on a wider audience than the person posting the original tweet and try to build a narrative. Annotators were suggested to try to write at least one counter-narrative of each type but only if they came naturally, otherwise they could leave it blank.
The instructions to generate each type of counter-narrative are as follows:
lation between the justi cation and the conclusion. HATE TWEET: user must deport all illegal migrants india already reeling under constant threat of muslim radicals curb population Justification: india already reeling under constant threat of muslim radicals curb population (fact) Conclusion: must deport all illegal migrants (policy) Collective: illegal migrants Property: muslim radicals COUNTER NARRATIVE A (Negate relation between justification and conclusion)
Deporting illegal migrants will not mitigate the problems with muslim radicals.
COUNTER NARRATIVE B (Negate relation between collective and property)
Illegal migrants are not necessarily muslim radicals.
COUNTER NARRATIVE C (Negate justification based on type)
It is not true that India is reeling under threat of muslim radicals.
Deporting illegal migrants without consideration to their circumstances is an inhumane move.
the property, action or consequence that is being assigned to the targeted group and the targeted group itself.
must be put into question or sources must be asked to prove that fact. If it is of type “value”, it must be highlighted that the premise is actually an opinion, possibly relativizing it as a xenophobous opinion. If it is a “policy”, a counter policy must be provided. Free Counter-Narrative (type D) All counter-narratives that the annotator comes up with and do not fall within any of the other three types.
An example of each type of counter-narrative can be seen in Figure 1. Our dataset3 consists of a total of 1722 counter-narratives for 725 argumentative tweets in English and 355 counternarratives for 144 tweets in Spanish (an average of 2.38 and 2.47 per tweet respectively). Table 1 shows the percentage of tweets that has a counter-narrative of each type.
We designed a series of experiments to assess the impact of high-quality examples and argumentative information in the automatic generation of counter-narratives via prompting LLMs. We want to explore the following approaches: Fine-tuned vs Few-shot Use a LLM that has been trained for general purposes to generate counter-narratives by prompting the LLM with some examples of the desired input-output, as shown in the left column of Figure 3, or take a general LLM and ne-tune it with the examples of hate tweets associated to manually generated counter-narratives.
combinations of argumentative information provided within the input of the model: Collective and Property; Justi cation, Conclusion and Pivot; and all types.
With specific kinds of counter-narratives We pretrained two models for each type of counternarrative using only that type: one without extra information and another adding argumentative information relevant for the correspondent type (Justi cation and Conclusion for type A, Collective and Property for type B and Justi cation for type C).
Small or Big size of the same kind of LLM We want to compare performance of a larger
model with higher hardware requirements against a smaller one, ne-tuned, cheaper to
run but requiring a speci c annotated dataset. After testing behavior of similar
alternatives (Bloom, GPT-J and GPT2), we chose Flan-T5 [
Few-shot experiments were conducted for Flan-T5 Base (small) and XL (larger) models. ne-tuning was only conducted on Flan-T5 Base due to computational resource constraints.
We conducted some manual evaluation of prospective to nd optimal parameters for generation, and we found that using Beam Search with 5 beams yielded the best results, so this is the con guration we used throughout the paper.
To ne-tune FLAN-T5 with our dataset of counter-narratives, we randomly split our dataset in training, development and test partitions, assuring that all counter-narratives for the same hate tweet are contained into the same partition. Details can be seen on Table 1.
Train Dev Test
English Spanish #Tweets #CNs % corpus A B C #Tweets #CNs % corpus 509 1201 69.8% 496 238 467 105 257 72.4% 71 173 10.0% 67 38 68 12 27 7.6% 145 348 20.2% 138 74 136 27 71 20%
Proportion of tweets with counter-narrative 96% 47% 90%
For the few-shot experiments, the prompt has an instruction followed by two random examples taken from the test partition of the dataset. For each example, the hate tweet and its corresponding counter-narrative are enclosed in special tokens de ning the start and end.
Evaluation of counter-narratives is not straightforward. So far, no automatic technique has been
found satisfactory for this speci c purpose. Automatic metrics proposed for other NLP tasks,
like BLEU [
Faced with the lack of appropriate automatic metrics adequate for the task, many authors
have conducted manual evaluations for automatically generated counter-narratives. Manual
evaluations typically distinguish di erent aspects of the adequacy of a given text as a
counternarrative for another. Chung et al. [
Based on these previous works, we have put together a rst version of criteria to manually evaluate4 the adequacy of counter-narratives, considering four di erent aspects: • O ensiveness: if the tweet is o ensive to either the target group, the author of the tweet or any other group or person. Possible values are: O ensive; Possibly O ensive/Not clear; Not o ensive. • Stance: if the tweet supports or counters the speci c message of the hate tweet. Possible values are: Supports the original message; Not clear/Changes subject wrt original tweet; Counters the original message. Stance incorporates a certain notion of suitableness, since it assigns value "Changes the subject" if the counter-narrative is not responding speci cally to the standpoint of the original tweet. • Informativeness: Evaluates the complexity and speci city of the generated text. Only counter-narratives with a "Counters" Stance are evaluated. Possible values are: 1. Generic statement: replies that don’t incorporate any information mentioned on the tweet and could counter many di erent hate messages (e.g "I don’t think so" or "That is not true"). 2. Speci c but not argumentative: the reply is a simple statement, possibly composed of a single sentence without providing justi cation for the stance but referring to some speci c aspect of the original tweet. Usually they comply with a formula composed of a pre x (like "I don’t think that" or "Do you have proof that") and a verbatim copy of some part of the hate tweet. 4Results of the evaluation can be found on https://shorturl.at/aetFZ 3. Speci c and Argumentative: counter-narratives with some degree of elaboration of the information contained on the hate message. We identi ed three common patterns that we associate with this value:
A - replies that take more than one element from the original message and stablish some relation between them (e.g. "I don’t see the relation between {element from the original message} and {other element from the original message}").
B - A simple statement declaring stance over a single element from the original tweet but adding a second coordinated statement with personal appreciations about it (e.g. "I don’t think we should {some policy mentioned on the tweet}. It is a bad idea").
C - An argumentative reply based on information not mentioned explicitly on
the original tweet, but necessarily inferred, showing a comprehensive
understanding of the meaning of the hate message (e.g. a reply to a tweet concluding
with #BuildTheWall saying "Building a wall would cost the taxpayers more" or
"Building a wall won’t give you more control over illegal tra cking").
• Felicity: This category is related to Chung et al. [
Aggregating the results for these four categories, we de ne two extra concepts: Good and Excellent counter-narratives. Good counter-narratives will be those with optimal values on O ensiveness, Stance and Felicity. Excellent counter-narratives will be those that also have the optimal value for Informativeness. We believe Informativeness is the most valuable of the four categories, that is why it is determinant in characterizing Excellent counter-narratives. The Good indicator shows that productions are not harmful or totally random.
We are planning to improve the kind of information that is currently captured in the Informativeness category in a second version of the evaluation criteria.
To properly evaluate the quality of the generated counter-narratives with the presented method, we conducted a preliminary manual evaluation. We evaluated three random subsets of 20 hate tweets in English and 10 in Spanish. One contains only tweets associated with counternarratives of both types A and C on our dataset, and was used to evaluate models ne-tuned only with these kinds of counter-narratives. Another contains only tweets associated with counter-narratives of type B and was also used to evaluate models ne-tuned only with this type of counter-narratives. The last subset contains tweets with counter-narrative pairs of all types, and was used for all the rest of the experiments.
We generated one counter-narrative for each tweet in the corresponding evaluation subset for each combination of features to be assessed: few-shot, ne-tuned, with di erent kinds of O ensiveness Stance Informativeness Felicity argumentative information, with di erent sizes of LLM. For the larger version of FLAN-T5 we only applied the few-shot approach, and, after assessing no improvement on the smaller version, we aborted the rest of experiments with this version of the LLM to reduce the carbon footprint of our experiments. The results for the 18 experiments can be seen in Table 3.
Then, three annotators labeled each tweet according to the four categories described above. The nal value for each category was obtained by calculating the value with more votes (at least two annotators agreed on the value). In total, each annotator labeled 540 hate tweet/counternarrative pairs. Of all these, there were 10 cases where each of the three annotators labeled a di erent value. In these cases, we adopted a conservative criterion and assigned the worst of the three possible values.
Table 2 shows the agreement scores between the three annotators, calculated using Cohen’s
Kappa [
We are currently working on a second version of the evaluation criteria, with more insightful categories, expanding on Informativeness and trying to capture argument acceptability, relevance and persuasiveness. We will check whether this improved criteria improve inter-annotator agreement. If so, we will engage a higher number of judges and aim to obtain a more reliable assessment of the quality of automatically generated counter-narratives.
Results of the manual evaluation of di erent strategies for counter-narrative generation for English can be seen in Table 3. A summary of this table can be seen in Figure 2, which displays the aggregated proportion of Good and Excellent counter-narratives for each strategy.
We can clearly see that the larger versions of the model (XL) produce counter-narratives that
are less satisfactory in general, and that argumentative information only decreases the quality
of the generated text. Fine-tuned models produce better counter-narratives in general, even
if smaller. A very valuable conclusion that can be obtained from these results is that a small
number of high quality examples produce a much bigger improvement in performance than
5The interpretation of the ranges of values of the kappa coe cient is according to Landis and Koch [
If we focus on Informativeness (third dimension of evaluation in Table 3, we can see that the approaches that produce most informative counter-narratives are ne-tuned (lower half of the Table), without a detriment in any of the other dimensions of evaluation. Interestingly, when ne-tuned only with counter-narratives of a single type, providing argumentative information consistently improves the informativeness of the counter-narratives, even if only slightly. We have to take into account that such approaches use a much smaller number of counter-narratives, as can be seen in Table 1. Even in the case of type B counter-narratives, with extremely few examples to ne-tune, argumentative information produces an improvement in informativeness.
When we make a qualitative analysis of the generated counter-narratives, we can see that providing argumentative information about the hate tweet does yield counter-narratives that are more speci c and informative, as can be seen in Figure3. Models counting with this information frequently use it by negating the relation between Collective and Property or between Justi cation and Conclusion.
Results obtained for counter-narratives for Spanish hate tweets were much worse, as could be expected given the much smaller number of examples for ne-tuning and that base LLMs
Tweet with argumentative information: street interview whit italians "send all migrants back to where they came from they block streets to pray " - free speech time - https://t co/d5dqr8pg3r @user | Justification: street interview whit italians "send all migrants back to where they came from they block streets to pray " (fact) | Conclusion: "send all migrants back to where they came from they block streets to pray " (policy) | Pivot: migrants - they - they
Counter-narrative: I don’t think it’s a good idea to send all migrants back to where they came from.
Tweet without argumentative information: street interview whit italians "send all migrants back to where they came from they block streets to pray " - free speech time - https://t co/d5dqr8pg3r @user
Counter-narrative: I don’t think it’s the right thing to do. perform worse for tasks in Spanish in general. Indeed, values for Informativeness and Felicity almost never reach more than 10% positive, and Stance and O ensiveness are almost never beyond 30% positiveness. However, the same tendency as for English could be observed: netuned models perform better than non- ne-tuned models, even if the latter are bigger. Moreover, argumentative information seems to make a bigger impact in improving the generated counternarratives than in the case of English, with increases in the range of 30%-50% in the reduction of negative scores for O ensiveness and Stance, although a decrease in Felicity. Given these encouraging results with such few examples, we will be increasing the number of examples with argumentative information in future work.
We have presented an approach to generate counter-narratives against hate speech in social media by prompting large language models with information about some argumentative aspects of the original hate speech. We have carried out a small manual evaluation of the quality of generated counter-narratives. This evaluation is preliminary, with a small number of judgements and moderate to substantial inter-annotator agreement, but we have found promising tendencies.
We have shown that argumentative information by itself does not improve the quality of counter-narratives generated by LLMs, on the contrary, it may even be detrimental, specially in the case of bigger models. However, ne-tuning a smaller model with a small corpus of high-quality examples of pairs hate speech – counter-narrative yields some improvement in performance. This nding has a signi cant impact both because smaller language models are more accessible to low-budget scenarios, and because of their smaller carbon footprint.
We have also shown that some kinds of argumentative information do have some positive impact in generating more speci c, more informative counter-narratives. In particular, we have found that the types of counter-narrative that negate the relation between the Justi cation and the Conclusion and that negate the Justi cation have an improvement in performance if argumentative information about the Justi cation and the Conclusion is provided.
Moreover, we have also found that argumentative information makes a positive impact in scenarios with very few tweets, as shown by our experiments for Spanish. Although the quality of the counter-narratives generated for Spanish is much lower than for English, the fact that argumentative information has a positive impact is encouraging, and we will continue to annotate examples for Spanish to improve the generation of counter-narratives.
We will also explore other aspects of the quality of counter-narratives, with a more insightful, more extensive human evaluation. We will also explore the interaction between argumentative information and other aspects, like vocabulary, level of formality, and culture.
Finally, the evaluation of counter-narratives is still far from being solved. We are currently considering di erent avenues to improve it, as it is a crucial step to advance the eld. We are working on obtaining a higher number of judgements, but also on more insightful guidelines that re ect more valuable aspects of counter-narratives, more related to argument acceptability.
This work was funded in part by Secretaría de Investigación Cientí ca y Tecnológica FCEN–UBA (RESCS-2020-345-E-UBA-REC), CONICET under the PIP (grant 11220200101408CO), Agencia Nacional de Promoción Cientí ca y Tecnológica, Argentina under grants PICT-2018-0475 (PRH2014-0007), PICT-2020- SERIEA-01481, and the NAACL Regional Americas Fund (2022). This work used computational resources from CCAD – UNC (https://ccad.unc.edu.ar/), which are part of SNCAD – MinCyT, Argentina. We specially want to thank two anonymous reviewers that contributed to improve this work with their thoughtful and constructive comments.