In the GEESE challenge, we present a pipeline to evaluate generated explanations for the task of Recognizing Textual Entailment (RTE) in Italian. The challenge focuses on evaluating the impact of generated explanations on the predictive performance of language models. Using a dataset enriched with human-written explanations, we employ two large language models (LLMs) to generate and utilize explanations for semantic relationships between sentence pairs. Our methodology assesses the quality of generated explanations by measuring changes in prediction accuracy when explanations are provided. Through reproducible experimentation, we establish benchmarks against various baseline approaches, demonstrating the potential of explanation injection to enhance model interpretability and performance.
provide human-understandable explanations has been a key research objective of Machine Learning (ML) and
an unprecedented acceleration in most areas, such as
ming languages [
derstand their reasoning is becoming central. Many recent eforts have been devoted to explaining such models [13], and the importance of interpretability and explainability in AI has become ever more urgent [14, 15, 16].
a consistent body of research. Cambria et al. [17], for
instance, provides a comprehensive survey of approaches
and Sonntag [18] examines the benefits of explanations
for NLP models; Paranjape et al. [19] focuses on
templatecomputer vision [
Large Language Models (LLMs) such as OpenAI’s Chat- based explanations, Lampinen et al. [20] and Ye and Durrett [21] demonstrate the benefits of in-context explana
These models are currently able to produce natural- tions for large models in challenging reasoning tasks. sounding and coherent language, often indistinguishable from natural language [11, 12]. While these results open up new avenues for future applications and research, they also raise ethical issues considering the ubiquitous role of machines in our lives, and in sensitive fields like education, health, justice, and private life. In fact, the scarce transparency of neural architectures makes it hard to interpret their functioning (the so-called ”black-box” problem). In addition, many of the currently available
trained on is not known to either researchers or the general public. Finally, these models have achieved such sizes that their results are dificult to replicate, making them a kind of ”black box in a black box”. nEvelop-O LGOBE
Explanation generation quality has traditionally been
evaluated through automated ovelap metrics like BLEU
[22], ROUGE [23], or BERT-Score [
Recently, human simulatability scores have been
proposed as an alternative method to understand the quality
of explanations from the perspective of the “utility to
an end-user” [
ceur-ws.org based on downstream performance: their methodology • assess the quality of the generated explanations involves training a student model on explanations gener- by taking the delta between prediction accuated by a teacher, using automatic explanation generation racy with and without explanation as a proxy of techniques and training the student for the end task. explanations’ quality.
However, current LLMs may also benefit from
explanation injection even if they are not explicitly trained to Step 1: Generate Explanation: A first LLM ( 1) is
do so, and some works suggest using the explanation to prompted to produce explanations = { 1, 2, … } for
augment the input to condition predictions of future data a specific semantic relation holding between a given
points on both the input and the explanation [
second LLM ( 2) is then provided with the generated
explanations to evaluate if the generated explanations
improve the task of predicting the correct relations. In
practice, this is achieved by appending the explanation
as a “hint” to the prompt, and asking the model to make
To contribute to this line of research, we propose GEESE: a prediction thereof. This process aims to discover how
Generating and Evaluating Explanations for Seman- efectively 2 leverages the explanations from 1 to
pertic Entailment at CALAMITA [
Consider a pair of sentences < 1, 2 >, like the ones in the following example: (1) Il cielo è grigio oggi.
(2) Faresti bene a prendere l’ombrello.2
calamita-geese 2(1) The sky is grey today. (2) You better take your umbrella with you.
Consider a semantic relation holding between 1 and 2 (e.g., 1 entails 2, 1 does not entail 2, 1 contradicts 2).
Let be the set of possible explanations for . GEESE’s explanatory task consists in: The Recognizing Textual Entailment (RTE) task emerged in 2005 [32] as the problem of determining if two sentences stand in an entailment or not-entailment relationship. A common definition of “semantic entailment” (also • generating an explanation ∈ for the semantic referred to as presupposition in some studies) is that “A relationship for each < 1, 2 > in the dataset; sentence S presupposes a proposition p if p must be true • predict the relation with and without the gener- in order for S to have a truth-value (to be true or false)” ated explanation ; [33]. A text t is said to entail another text (hypothesis, h) if h is true in every circumstance (possible world) in which t is true. RTE, however, suggests a more empirical definition, allowing for cases in which the truth of Step 3: Evaluate Explanation Efectiveness Explanation efectiveness is evaluated by analyzing how providing diferent explanations generated in Step 1 afects the model 2 prediction in Step 2. In practice, this is done by calculating the accuracy of the predictions of 2 given the explanations and comparing them to the selected baselines (see Section 4).
3.1. Origin of data 3.2. Detailed data statistics
in Italian (text_t and text_h in the dataset) divided into an 800-example validation and an 800-example test split. Each example is annotated with an entailment label (label): "YES" (entailed), "NO" (contradicted), or "UNKNOWN" (neutral).
format4. A snippet of the data is displayed in Table 1.
Instruct as 1 with a custom implementation in HuggingFace, and Llama-3-8B as 2, using the LLM-EvaluationHarness library [38] in a zero-shot setting5.
We provide baselines for the following settings: the hypothesis is highly plausible, for most practical pur- tions. In our implementation, this is done through regposes, rather than certain. According to [34], this “shal- ular expressions by substituting (“anonimize”) the label low” definition better accounts for the types of uncertain strings ("YES", "NO", "UNKNOWN") and all words startinferences that are typically expected from text-based ing with entail.*, contradict.*, neutr.*, impl*, applications. contradd.* (verbs and nouns directly stating the kind
Recognizing Textual Entailment was formalized of relationship) with ”XXX”. through a series of successful challenges and workshops We therefore also provide the following “anonymized” that began in 2005 [32] and lasted until 2012. Starting additional explanations for each example, which we use from the RTE-3 edition, the task was extended from two in our prompts: labels to a three-label classification, splitting the notentailment label into two classes, contradiction and neu- • anon_whyexp: the anonimized explanation gentrality. Given the interest in the task, an Italian version erated by llama3 as 1; of the RTE-3 dataset was developed to explore language • anon_human: the anonimized human-written excomprehension and textual entailment [35]. planation (from e-RTE-3-it).
The dataset used in the challenge is the e-RTE-3-it dataset [36], which is an emended version enriched with 3.4. Data format human-written explanations of the RTE-3-it dataset [35]. 3.3. Annotation details
ten in Italian by native speakers. For each text-hypothesis pair, annotators provided a natural language explanation justifying the given label (explanation) for the entailment relation (“why does 1 stand in an relation with 2?”)3.
All annotations underwent quality control, involving two expert linguists who manually checked the explanations for grammaticality, fluency, and logical validity.
This process ensured high quality of the final e-RTE-3-it 4.1. Example of prompts for zero shots explanations, informativeness, as well as minimal label leakage (see infra). All experiments have been carried out in a zero-shot
Label leakage [37] refers to the fact that the explana- setting using the following prompts6. tion may be directly suggesting the label without genuinely being informative. While the manual check of (M1 - Generation): Your task all original human explanations ensured minimal label is to clarify the entailment leakage, to prevent this we automatically replace di- relationship between a pair rect references to the label and to the task with place- of sentences by explaining holders in the human-written and generated explana- why a classifier predicted
2. dummy: The hypothesis itself (text_t) provided as a ”non-informative” explanation, controlling for input length and providing a second baseline. 3. human: Human-written explanations (from e
RTE-3-it) anonimized (anon_human) provided as additional input; 4. llama-3: The explanation generated using
LLama-3-8B-Instruct as 1 (anon_whyexp). 3Additionally, the annotator provided a confidence score (1-5) reflecting their certainty about the provided explanation (which we don’t use in the task), an optional alternative label, if they felt the initial label was inaccurate, along with explanations and confidence scores. We don’t consider these annotations in the task, and only use the original label as our gold relationship and the human-written explanation for the original label as a strong baseline. 4https://huggingface.co/datasets/azaninello/explained-full-llama-3 5Generation parameters are: stop sequences: ”</s>”, ”<|eot_id|>”, max. gen. tokens: 128; do sample: false; temperature: 0 6All prompts are provided in English to ensure comparability across tasks and language models. In fact, using English as a “metalanguage” allows us to establish a comparable set-up to study LLMs’ capability regardless of their specific training language. id text_t text_h label text_label explanation anon_whyexp anon_human 393 Gli Eschimesi, i cacciatori di foche sparsi nell’Artico, prevedono di sollecitare un giudizio della Commissione Interamericana dei Diritti dell’Uomo poiché gli Stati Uniti, contribuendo in modo sostanziale al riscaldamento globale, stanno minacciando la loro esistenza.
La caccia alle foche mette a rischio d’estinzione le specie.
UNKNOWN neutrality Nella frase non vengono citate le conseguenze della caccia sulle foche perciò non si può evincere il rischio per loro della caccia.
The XXX relationship between the two sentences is XXX because the relationship is not clearly established. Sentence 1 mentions the Inuit people, the Commission on Human Rights, and the United States, which are not directly related to the topic of the second sentence, which is about the threat of extinction of species due to hunting. The connection between the two sentences is not explicitly stated, and it is not clear how the information presented in Sentence 1 supports or XXX the information in Sentence 2. Therefore, the XXX label is XXX because the relationship between the two sentences is not clear. For example, if the XXX label was XXX, it would mean that Sentence 1 logically implies Sentence 2, but the information in Sentence 1 does not provide enough evidence to support this claim. Similarly, if the XXX label was XXX, it would mean that Sentence 1 logically XXX Sentence 2, but again, the information in Sentence 1 does not provide enough evidence to support this claim. Therefore, the XXX label is XXX because the relationship between the two sentences is not clear.
Nella frase non vengono citate le conseguenze della caccia sulle foche perciò non si può evincere il rischio per loro della caccia.
a specific entailment label. accuracy at 78.12%. In comparison, using human-written Sentence 1: text_t Sentence 2: explanations shows slightly lower accuracy compared text_h Entailment label: label. to machine-generated, but higher scores compared to exp_type7 baselines, suggesting that explanations do enhance the (M2 - Prediction): Your task is models’ understanding of semantic relationships. to predict the entailment label Generated explanations, proving more efective than between two sentences, selecting human-crafted ones, suggest that the quality and type one label among YES (entailment), of explanations provided can influence predictive perforNO (contradiction), or UNKNOWN mance, but also highlight the need for further research (neutrality). Sentence 1: into optimizing explanation generation methods for imtext_t Sentence 2: text_h Hint: proved outcomes in NLP tasks. In fact, note that generanon_explanation. Entailment ated explanations may be positively influenced by factors label:8 other than informativeness alone, such as the lengths of the explanations themselves, or may still be indirectly suggesting the right relationship despite the anonymiza
For example, as reported by one of the anonymous reBaseline results, reported in Table 2, demonstrate the im- viewers, see “anon_whyexp” explanation in Table 1: “In pact of incorporating explanations on the performance of other words, Sentence 2 provides enough information language models in the Recognizing Textual Entailment to infer the truth of Sentence 1”. The generated explatasks. The accuracy scores indicate that models utilizing nation clearly (but not directly) hints at an ”entail” label, explanations generated by Llama-3 achieve the highest potentially compromising the intended anonymity. The fairness of the comparison between human- and machinegenerated explanation is an aspect that deserves further investigation. 7Variables are indicated in color. In our experiments exp_type = “Explain how the two sentences are connected.” and the variables are read from each example. 8Variables are indicated in color. In our experiments, anon_explanation can take the following values: “Not given.” (no-exp), text_h (dummy), anon_human (human), anon_whyexp (llama-3).
Tasks geese_dummy geese_noexp geese_llama3 geese_human n-shot 0 0 0 0
acc acc acc acc
0.5850 0.5437 0.7812 0.7575
significance of efective explanation generation in enhancing the capabilities of language models in RTE tasks. Preliminary results show that models provided with explanations, whether human-written or generated by LLMs, exhibit improved predictive accuracy compared to those lacking such inputs. This supports the hypothesis that explanations can facilitate a deeper understanding of semantic relationships, thus aiding model inference.
The GEESE challenge establishes a framework for generating and evaluating explanations in the domain of semantic entailment. By demonstrating the utility of explanation injection, we contribute to the ongoing discourse on interpretability in AI, advocating for a balanced approach that enhances model transparency while maintaining robustness. Our findings encourage further exploration into the interplay between explanations and model performance, paving the way for more interpretable and user-friendly AI systems. As language models continue to evolve, integrating efective explanation mechanisms will be crucial for ensuring their responsible deployment in sensitive applications.
The study also highlights limitations, including potential biases in the generated explanations and the challenge of ensuring that explanations remain informative without directly revealing the answer. Future research could explore diverse explanation types and their varying impacts across diferent contexts and languages.
We would like to draw the readers’ attention on the following. Firstly, the potential for bias in both the training data and the generated explanations can perpetuate stereotypes or misinformation, leading to harmful consequences, particularly in sensitive domains such as healthcare or legal applications. There is also the risk that users may place undue trust in machine-generated explanations, mistakenly believing them to be infallible. Finally, the collection and use of data for training these models must adhere to strict privacy standards to ensure that individuals’ rights are respected. Addressing these ethical challenges is essential to foster trust and ensure that AI technologies are developed and used responsibly.
FAIR - Future AI Research (PE00000013), under the NRRP MUR program funded by the NextGenerationEU and the ANTIDOTE project (CHIST-ERA grant of the Call XAI 2019 of the ANR with the grant number Project-ANR-21CHR4-0002) D. Cummings, M. Plappert, F. Chantzis, E. Barnes, arXiv:https://academic.oup.com/idpl/articleA. Herbert-Voss, W. H. Guss, A. Nichol, A. Paino, pdf/7/4/233/22923065/ipx022.pdf. N. Tezak, J. Tang, I. Babuschkin, S. Balaji, S. Jain, [16] L. Edwards, M. Veale, Slave to the algorithm: Why W. Saunders, C. Hesse, A. N. Carr, J. Leike, a right to an explanation is probably not the remedy J. Achiam, V. Misra, E. Morikawa, A. Radford, you are looking for, Duke L. & Tech. Rev. 16 (2017) M. Knight, M. Brundage, M. Murati, K. Mayer, 18.
P. Welinder, B. McGrew, D. Amodei, S. McCan- [17] E. Cambria, L. Malandri, F. Mercorio, M. Mezdlish, I. Sutskever, W. Zaremba, Evaluating zanzanica, N. Nobani, A survey on xai and large language models trained on code, CoRR natural language explanations, Information abs/2107.03374 (2021). URL: https://arxiv.org/abs/ Processing Management 60 (2023) 103111. 2107.03374. arXiv:2107.03374. URL: https://www.sciencedirect.com/science/ [8] OpenAI, Gpt-4 technical report, 2023. article/pii/S0306457322002126. doi:https: arXiv:2303.08774. //doi.org/10.1016/j.ipm.2022.103111. [9] G. Team, Gemini: A family of highly capable mul- [18] M. Hartmann, D. Sonntag, A survey on improving timodal models, 2024. URL: https://arxiv.org/abs/ NLP models with human explanations, in: Proceed2312.11805. arXiv:2312.11805. ings of the First Workshop on Learning with Nat[10] H. Touvron, L. Martin, K. Stone, P. Albert, A. Alma- ural Language Supervision, Association for Comhairi, Y. Babaei, N. Bashlykov, S. Batra, P. Bhar- putational Linguistics, Dublin, Ireland, 2022, pp. gava, S. Bhosale, D. Bikel, L. Blecher, C. C. Ferrer, 40–47. URL: https://aclanthology.org/2022.lnls-1.5. M. Chen, G. Cucurull, D. Esiobu, J. Fernandes, J. Fu, doi:10.18653/v1/2022.lnls-1.5.
W. Fu, B. Fuller, C. Gao, V. Goswami, N. Goyal, [19] B. Paranjape, J. Michael, M. Ghazvininejad, H. HaA. Hartshorn, S. Hosseini, R. Hou, H. Inan, M. Kar- jishirzi, L. Zettlemoyer, Prompting contrastive das, V. Kerkez, M. Khabsa, I. Kloumann, A. Ko- explanations for commonsense reasoning tasks, renev, P. S. Koura, M.-A. Lachaux, T. Lavril, J. Lee, in: Findings of the Association for CompuD. Liskovich, Y. Lu, Y. Mao, X. Martinet, T. Mihaylov, tational Linguistics: ACL-IJCNLP 2021, AssoP. Mishra, I. Molybog, Y. Nie, A. Poulton, J. Reizen- ciation for Computational Linguistics, Online, stein, R. Rungta, K. Saladi, A. Schelten, R. Silva, E. M. 2021, pp. 4179–4192. URL: https://aclanthology. Smith, R. Subramanian, X. E. Tan, B. Tang, R. Tay- org/2021.findings-acl.366. doi:10.18653/v1/2021. lor, A. Williams, J. X. Kuan, P. Xu, Z. Yan, I. Zarov, findings-acl.366.
Y. Zhang, A. Fan, M. Kambadur, S. Narang, A. Ro- [20] A. Lampinen, I. Dasgupta, S. Chan, K. Mathdriguez, R. Stojnic, S. Edunov, T. Scialom, Llama 2: ewson, M. Tessler, A. Creswell, J. McClelland, Open foundation and fine-tuned chat models, 2023. J. Wang, F. Hill, Can language models learn arXiv:2307.09288. from explanations in context?, in: Y. Goldberg, [11] T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Ka- Z. Kozareva, Y. Zhang (Eds.), Findings of the Associplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sas- ation for Computational Linguistics: EMNLP 2022, try, A. Askell, et al., Language models are few-shot Association for Computational Linguistics, Abu learners, Advances in neural information process- Dhabi, United Arab Emirates, 2022, pp. 537–563. ing systems 33 (2020) 1877–1901. URL: https://aclanthology.org/2022.findings-emnlp. [12] T. Labruna, S. Brenna, A. Zaninello, B. Magnini, Un- 38. doi:10.18653/v1/2022.findings-emnlp.38. raveling chatgpt: A critical analysis of ai-generated [21] X. Ye, G. Durrett, The unreliability of explanations goal-oriented dialogues and annotations, 2023. in few-shot prompting for textual reasoning, in: arXiv:2305.14556. S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, [13] R. Guidotti, A. Monreale, S. Ruggieri, F. Turini, F. Gi- K. Cho, A. Oh (Eds.), Advances in Neural Informaannotti, D. Pedreschi, A survey of methods for tion Processing Systems, volume 35, Curran Assoexplaining black box models, ACM computing sur- ciates, Inc., 2022, pp. 30378–30392. URL: https:// veys (CSUR) 51 (2018) 1–42. proceedings.neurips.cc/paper_files/paper/2022/file/ [14] A. Vassiliades, N. Bassiliades, T. Patkos, Argumenta- c402501846f9fe03e2cac015b3f0e6b1-Paper-Conference. tion and explainable artificial intelligence: a survey, pdf.
The Knowledge Engineering Review 36 (2021) e5. [22] K. Papineni, S. Roukos, T. Ward, W.-J. Zhu, Bleu: a doi:10.1017/S0269888921000011. method for automatic evaluation of machine trans[15] A. D. Selbst, J. Powles, Meaningful information lation, in: Proceedings of the 40th annual meeting and the right to explanation, International Data of the Association for Computational Linguistics, Privacy Law 7 (2017) 233–242. URL: https://doi.org/ 2002, pp. 311–318. 10.1093/idpl/ipx022. doi:10.1093/idpl/ipx022. [23] L. C. ROUGE, A package for automatic evaluation
D. Livshits, D. Esiobu, D. Choudhary, D. Mahajan, D. Garcia-Olano, D. Perino, D. Hupkes, E. Lakomkin, E. AlBadawy, E. Lobanova, E. Dinan, E. M. Smith, F. Radenovic, F. Zhang, G. Synnaeve, G. Lee, G. L. Anderson, G. Nail, G. Mialon, G. Pang, G. Cucurell, H. Nguyen, H. Korevaar, H. Xu, H. Touvron, I. Zarov, I. A. Ibarra, I. Kloumann, I. Misra, I. Evtimov, J. Copet, J. Lee, J. Gefert, J. Vranes, J. Park, J. Mahadeokar, J. Shah, J. van der Linde, J. Billock, J. Hong, J. Lee, J. Fu, J. Chi, J. Huang, J. Liu, J. Wang, J. Yu, J. Bitton, J. Spisak, J. Park, J. Rocca, J. Johnstun, J. Saxe, J. Jia, K. V. Alwala, K. Upasani, K. Plawiak, K. Li, K. Heafield, K. Stone, K. El-Arini, K. Iyer, K. Malik, K. Chiu, K. Bhalla, L. Rantala-Yeary, L. van der Maaten, L. Chen, L. Tan, L. Jenkins, L. Martin, L. Madaan, L. Malo, L. Blecher, L. Landzaat, L. de Oliveira, M. Muzzi, M. Pasupuleti, M. Singh, M. Paluri, M. Kardas, M. Oldham, M. Rita, M. Pavlova, M. Kambadur, M. Lewis, M. Si, M. K. Singh, M. Hassan, N. Goyal, N. Torabi, N. Bashlykov, N. Bogoychev, N. Chatterji, O. Duchenne, O. Çelebi, P. Alrassy, P. Zhang, P. Li, P. Vasic, P. Weng, P. Bhargava, P. Dubal, P. Krishnan, P. S. Koura, P. Xu, Q. He, Q. Dong, R. Srinivasan, R. Ganapathy, R. Calderer, R. S. Cabral, R. Stojnic, R. Raileanu, R. Girdhar, R. Patel, R. Sauvestre, R. Polidoro, R. Sumbaly, R. Taylor, R. Silva, R. Hou, R. Wang, S. Hosseini, S. Chennabasappa, S. Singh, S. Bell, S. S. Kim, S. Edunov, S. Nie, S. Narang, S. Raparthy, S. Shen, S. Wan, S. Bhosale, S. Zhang, S. Vandenhende, S. Batra, S. Whitman, S. Sootla, S. Collot, S. Gururangan, S. Borodinsky, T. Herman, T. Fowler, T. Sheasha, T. Georgiou, T. Scialom, T. Speckbacher, T. Mihaylov, T. Xiao, U. Karn, V. Goswami, V. Gupta, V. Ramanathan, V. Kerkez, V. Gonguet, V. Do, V. Vogeti, V. Petrovic, W. Chu, W. Xiong, W. Fu, W. Meers, X. Martinet, X. Wang, X. E. Tan, X. Xie, X. Jia, X. Wang, Y. Goldschlag, Y. Gaur, Y. Babaei, Y. Wen, Y. Song, Y. Zhang, Y. Li, Y. Mao, Z. D. Coudert, Z. Yan, Z. Chen, Z. Papakipos, A. Singh, A. Grattafiori, A. Jain, A. Kelsey, A. Shajnfeld, A. Gangidi, A. Victoria, A. Goldstand, A. Menon, A. Sharma, A. Boesenberg, A. Vaughan, A. Baevski, A. Feinstein, A. Kallet, A. Sangani, A. Yunus, A. Lupu, A. Alvarado, A. Caples, A. Gu, A. Ho, A. Poulton, A. Ryan, A. Ramchandani, A. Franco, A. Saraf, A. Chowdhury, A. Gabriel, A. Bharambe, A. Eisenman, A. Yazdan, B. James, B. Maurer, B. Leonhardi, B. Huang, B. Loyd, B. D. Paola, B. Paranjape, B. Liu, B. Wu, B. Ni, B. Hancock, B. Wasti, B. Spence, B. Stojkovic, B. Gamido, B. Montalvo, C. Parker, C. Burton, C. Mejia, C. Wang, C. Kim, C. Zhou, C. Hu, C.-H. Chu, C. Cai, C. Tindal, C. Feichtenhofer,