Explainability is a crucial open challenge in Natural Language Processing (NLP) research on abusive language [1] as increasing models' complexity [2], models' intrinsic bias [3], and international regulations [4] call for a shift in perspective from performance-based models to more transparent models. Moreover, recent studies have shown the benefits of explanations for users [5,6] and content moderators [7] on social media platforms. The former can benefit from receiving an explanation for why a certain post has been flagged or removed whereas the latter are shown to annotate toxic posts faster and solve doubtful annotations thanks to explanations.

Several efforts have moved towards explainable abusive language detection in the past years, like the development of datasets containing rationales (i.e., the tokens in the text that suggest why the text is hateful) [8] or implied statements (i.e., description of the implied meaning of the text) [9,10], and shared tasks on explainable hate speech detection [11,12], inter alia. With Large Language Models (LLMs) like FLAN-T5 [13] showing remarkable performance across tasks and human-like text generation [14,15,16], recent studies have explored LLMs for explainable hate speech detection, wherein classification predictions are described through natural language explanations [17,18]. For instance, [19] used chain-of-thought prompting [20] of LLMs to generate explanations for implicit hate speech detection.

However, most of these studies rely on empirical metrics like BLEU [21] to evaluate the generated explanations automatically. Consequently, the human perception and implications of these explanations remain understudied, as well as the extent to which empirical metrics approximate human judgements. [22] recruited crowdworkers to evaluate the level of hatefulness in tweets and the quality of explanations generated by GPT-3. Instead, we conduct an expert survey investigating four LLMs and five learning strategies across multi-class abusive language detection tasks to answer the following questions: RQ1: How well do LLM-generated explanations for abusive language detection match human expectations? RQ2: How well do empirical metrics align with human judgements? RQ3: What makes LLM-generated explanations good, according to experts?

To answer these research questions, we design a beforeand-after study, surveying participants about their prior expectations about LLM-generated explanations and then showing them examples generated by several LLMs with diverse learning strategies1 , followed by further interviews. To ensure robustness of our results, we recruited experts in the field, i.e., AI researchers, as described below.

For our experiments, we use the HateXplain [8] and the Implicit Hate Corpus [9] as they encompass different levels of offensiveness (i.e., hate speech, offensive, neutral), expressiveness (i.e., explicit hate, implicit hate, neutral), multiple targeted groups, and explanations for the hateful label (Table 1). These datasets contain unstructured explanations of the words that constitute abuse (in Hat-eXplain) and the user's intent (in Implicit Hate). In view of previous research arguing the need for structured explanations in hateful content moderation [1], we use the following template to create structured explanations, that we will use as ground-truth: "Explanation: it contains the following hateful words (implied statement):" for abusive content in HateXplain (Implicit Hate Corpus), and "The text does not contain abusive content." for neutral content.

We extensively investigate four popular LLMs across five learning strategies on their ability to detect multi-class offensiveness and expressiveness of abusive language and to generate explanations for the classification.

We use different open-source LLMs (Table 2): the base versions of FLAN-Alpaca [23,24], FLAN-T5 [13], mT0 [25], and the 7B foundational model Llama 2 [26], which is an updated version of LlaMA [27].

Summary of models used.

Learning strategies. As different prompting strategies might yield different results, we test five distinct learning strategies using the established Stanford Alpaca template2 (cf. Appendix A for prompt details):

(1) zero-shot learning (zsl): we pass "Classify the input text as list_of_labels, and provide an explanation" in the instruction field of the template. The list_of_labels changes according to the dataset used;

(2) few-shot learning (fsl): we pass three additional examples to the aforementioned template, which are randomly sampled with equal probability among the labels to account for class imbalance in the datasets. We experimented with different numbers of examples (i.e., passing one, three or five examples), and chose three as it was the best strategy;

(3) knowledge-guided zero-shot learning (kg): instead of passing additional examples in the prompts, we add external knowledge retrieved by means of an entity linker 3 , which first detects entities mentioned in the input text, and then retrieves the relevant information from the external knowledge base. We use Wikidata [28] for encyclopedic knowledge, KnowledJe [29] for hate speech temporal linguistic knowledge and ConceptNet [30] for commonsense knowledge. We modify the prompt template with an additional field called 'context' to account for this external knowledge;

(4) instruction fine-tuning (ft): we use the same prompts used in (1) to instruction fine-tune Llama-2;

(5) knowledge-guided instruction fine-tuning (kg_ft): we use the knowledge-guided prompts developed in (3) to instruction fine-tune Llama-2.

Empirical eval metrics. We evaluate how closely the LLM-generated explanations match the ground-truth across eight empirical similarity metrics due to the challenge of simultaneously assessing a wide set of criteria [31,32,33]. Following established NLG research [34,35], we choose BERTScore [36] and METEOR [37] for semantic similarity. For syntactic similarity, we select BLEU [21], GBLEU [38], ROUGE [39], ChrF [40] with its derivates ChrF+ and ChrF++ [41,42]. Additionally, we present an expert evaluation following our survey.

To evaluate how well LLMs align with human expectations and judgements in explanation generation, we design a before-and-after study as follows.

Before treatment. We ask for participant's background information, e.g., gender identity, native language and how they would rate the usefulness and trustworthiness of a language model for explanation generation. Specifically, we ask "How useful would you rate a system that provides you a textual explanation for its classification with respect to receiving only its classification?" and "How trustworthy would you rate a system that provides you a textual explanation for its classification with respect to receiving only its classification?" on a 1-5 Likert scale.

Treatment. As for the treatment, we show participants a sample of 70 texts from the datasets, paired with up to four different explanations. Specifically, given a text and ground-truth explanation, participants are asked if the text is correctly explained. If yes, they are asked to rate three different LLM-generated explanations with respect to the ground-truth on a 1-3 scale. These explanations are randomly sampled among the four LLMs and five learning strategies discussed in Section 2.2.

After treatment. Finally, we ask participants' opinion on the usefulness and trustworthiness of explanation generation, having seen the LLM-generated explanations. In addition, we ask general opinions related to what type of errors they observed most frequently, and what a good explanation would look like.

The full list of questions is in the Appendix B. The institutional ethical board of the first author's university approved our study design. We distributed the survey through channels that allow us to target individuals working in AI who are familiar with the field of language models and/or AI Ethics, including NLP reading groups and AI Ethics interest groups. To ensure the reliability of our before-and-after study, participants were given 1 hour to complete as many answers as they could. We collected answers from 15 participants, of which 33% (67%) identify as female (male), and 33% (67%) are (non) English native-speakers. The average level of participants' expertise in abusive language research is 2.47 out of 5 (self-described) 4 , and their continents

Our 15 participants reach a fair agreement, with Krippendorff's alpha [43] equal to 38.43%.

Fig. 1 shows changes in the relative frequencies of participant scores in the usefulness and trustworthiness of explanations before and after treatment. Participants' responses before treatment have expectations of textual explanations for classifications of being "highly useful" (above 50%; highest possible score) in terms of usefulness, and "moderately trustworthy" or "neutral" (above 40%; second and third best possible score) in terms of trustworthiness. However, scores for after treatment show participants changing their usefulness scores towards "moderately unuseful" (40-50%; second worst possible score) and their trustworthiness scores to "highly untrustworthy" (above 30%; worst possible score). Agreement differs in each category: usefulness is much more consensual, whereas trustworthiness is judged with higher variance. In general, LLM-generated explanations do not meet human expectations in terms of usefulness and trustworthiness. Specifically, exposing participants to these explanations leads to an average percentage decrease of 47.78% and 64.32% in the perception of the usefulness and trustworthiness of explanations, respectively. Fig. 2 shows the scores of all empirical metrics and expert evaluation for all models on explanation generation. Overall, similarity metrics tend to be highly volatile with respect to each other. For instance, FLAN-Alpaca prompted with zero-shot learning (i.e., 'alpaca_zsl' in the figure) generates explanations that are more than 70% semantically similar to the ground-truth explanations according to BERTScore while being less than 20% semantically similar according to METEOR. Similarly for syntax: BLEU and GBLEU similarity scores are less than 3% whereas ROUGE and chrF/+/++ are in the range 9%-21%. Moreover, we observe that BERTScore has a tendency to over-score explanations compared to human evaluation scores. Contrarily, METEOR, BLEU, GBLEU, ROUGE and chrF/+/++ have a tendency to under-score explanations. Instruction fine-tuning helped all metrics to approximate expert evaluations better, especially when tuned on knowledge-guided prompts. We use the Spearman's rank correlation coefficient to compare the correlation between human scores and those provided by all the other metrics. In detail, we rank the models for each type of metric, and then we compute the Spearman correlation between the rank obtained by human scores and those obtained by other metrics. Table 3 reports all the correlation scores. We observe that BERTScore is the most correlated with humans in both tasks. Also, Figure 1: Relative frequencies of Likert scores before and after treatment on usefulness and trustworthiness of LLMs for explanation generation in abusive language detection. chrF/+/++ metrics are highly correlated with humans while all the other metrics based on syntactic matches are slightly correlated with humans. Results show that semantic metrics are more similar to how humans evaluate the quality of the explanation generated by LLMs. Only one metric (ROUGE) shows a different behaviour between the two tasks.

Since 38.55% of the ground-truth explanations were not rated as good explanations by participants, we further investigated what are the most common errors and what makes an explanation good. Table 4 returns the most common error categories reported by participants. Most of them are related to logical fallacies (e.g., contradictory statements, hallucination), especially in the context of sarcasm and self-deprecating humour, rather than linguistic errors (e.g., grammar, misspellings). It is worth noticing that 13.33% of the participants reported that LLM-generated explanations contain cultural bias (e.g., stereotypes), with the implication of potentially perpetuating harms against the targeted victims of abusive language. As for desiderata, 73.33% of participants would like to receive textual explanations that are coherent with human reasoning and understanding, i.e., that are relevant and exhaustive to the text they refer to while being logically and linguistically correct. A remaining 20% thinks that a good explanation must be coherent with model reasoning instead. In other words, participants are much more concerned about how the explanation looks like rather than its reflection of the inner mechanism of the model reasoning. To quote a participant's perspective, "I would want the explanation to be helpful to me and guide my own reasoning".

In this paper, we conducted a before-and-after study to understand human expectations and judgements of LLMgenerated explanations for multi-class abusive language detection tasks. Contrarily to previous research [22], we investigated multiple LLMs and learning techniques, and we surveyed AI experts who are familiar with abusive language research instead of crowdworkers. We found that human expectations in terms of usefulness and trustworthiness of LLM-generated explanations are not met: after seeing these explanations, the usefulness and trustworthiness ratings decrease by 47.78% and 64.32%, respectively. Secondly, our results show that empirical metrics commonly used to evaluate textual explanations are highly volatile with respect to each other, even when they measure the same type of similarity (i.e., semantic vs. syntactic), and therefore pointing at the need of more reliable metrics for the empirical evaluation of textual explanations. In general, BERTScore and METEOR metrics exhibit the strongest correlation with human judgements. Lastly, our study provides evidence of the desiderata for LLM-generated explanations, suggesting that explanations should be coherent with human reasoning rather than model reasoning. Participants value the most textual explanations that are relevant and exhaustive to the text they refer to, while being logically and linguistically correct. Justifications for this preference lie on the fact that abusive language detection heavily relies on additional context and knowledge about slang and slurs, for which receiving an explanation is helpful to participants' understanding of the text. Future work should investigate whether this preference holds for other domains as well. In light of our findings, we conclude with three recommendations to use LLMs responsibly for explainable abusive language detection: (1) be aware of the cultural bias these models might exhibit when generating free-text explanations, which can further harm targeted groups;

(2) if possible, instruction fine-tune LLMs for explanation generation of abusive language detection. This not only could ensure the generation of structured explanations as advised by previous research [1] but it also returns the highest evaluation scores, both empirically and expert-wise, when using knowledge-guided prompts;

(3) opt for a combination of empirical metrics to evaluate textual explanations when no human evaluation is possible, since no particular empirical metric seems to generalise across different learning techniques, models and datasets, making the ground-truth lie somewhere in between BERTScore (upper bound) and BLEU (lower bound).