Providing high quality explanations for AI predictions based on machine learning is a challenging and complex task. To work well it requires, among other factors: selecting a proper level of generality/specificity of the explanation; considering assumptions about the familiarity of the explanation beneficiary with the AI task under consideration; referring to specific elements that have contributed to the decision; making use of additional knowledge (e.g. expert evidence) which might not be part of the prediction process; and providing evidence supporting negative hypothesis. Finally, the system needs to formulate the explanation in a clearly interpretable, and possibly convincing, way. Given these considerations, ANTIDOTE fosters an integrated vision of explainable AI, where low-level characteristics of the deep learning process are combined with higher level schemes proper of the human argumentation capacity. ANTIDOTE will exploit cross-disciplinary competences in deep learning and argumentation to support a broader and innovative view of explainable AI, where the need for high-quality explanations for clinical cases deliberation is critical. As a first result of the project, we publish the Antidote CasiMedicos dataset to facilitate research on explainable AI in general, and argumentation in the medical domain in particular.
under consideration, (iii) referring to specific elements that have contributed to the decision, (iv) making use of ANTIDOTE1 is a European CHIST-ERA project where additional knowledge (e.g. metadata) which might not be each partner is funded by their national Science Agencies. part of the prediction process, (v) selecting appropriate As the Spanish partner in the Consortium is the HiTZ examples and, (vi) providing evidence supporting negaCenter - Ixa, from the University of the Basque Coun- tive hypotheses. Finally, the system needs to formulate try UPV/EHU, the project was funded by the Proyectos the explanation in a clearly interpretable, and possibly de Colaboración Internacional (PCI 2020) program of the convincing, way.
Spanish Ministry of Science and Innovation. The other Taking into account these considerations, ANTIDOTE European partners are the following: Université Côte fosters an integrated vision of Explainable AI (XAI), d’Azur (UCA) from France and coordinators of the in- where the low-level characteristics of the deep learning ternational consortium, Fondazione Bruno Kessler (FBK) process are combined with higher level schemes proper from Italy, KU Leuven/Computer Science, in Belgium and of human argumentation. Following this, the ANTIDOTE Universidade Nova de Lisboa (NOVA) in Portugal. integrated vision is supported by three considerations.
The aim of ANTIDOTE is to exploit cross-disciplinary First, in neural architectures the correlation between incompetences in three areas, namely, deep learning, ar- ternal states of the network (e.g., weights assumed by gumentation and interactivity, to support a broader and single nodes) and the justification of the network classifiinnovative view of explainable AI. cation outcome is not well studied. Second, high quality
Providing high quality explanations for AI predictions explanations are crucially based on argumentation mechbased on machine learning is a challenging and complex anisms (e.g., provide supporting examples and rejected task. To work well it requires, among other aspects: (i) alternatives). Finally, in real settings, providing explanaselecting a proper level of generality/specificity of the tions is inherently an interactive process involving the explanation, (ii) considering assumptions about the fa- system and the user. miliarity of the explanation beneficiary with the AI task Thus, ANTIDOTE will exploit cross-disciplinary competences in three areas, namely, deep learning, argumentation and interactivity, to support a broader and innovative view of explainable AI. There are several research challenges that ANTIDOTE will address to advance the state-of-the-art in explainable AI.
The first challenge is to take advantage of the huge SEPLN-PD 2023: Annual Conference of the Spanish Association for Natural Language Processing 2023: Projects and System Demonstrations $ rodrigo.agerri@ehu.eus (R. Agerri)
© 2023 Copyright for this paper by its authors. Use permitted under Creative CPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g CCoEmmUoRns LWiceonsrekAstthribouptionP4r.0oIncteerenadtiionnagl s(CC(CBYE4U.0)R.-WS.org)
1https://univ-cotedazur.eu/antidote body of past research on argumentation to complement state-of-the-art approaches on explainability. In addition, the recent resurgence of AI highlights the idea that lowlevel system behavior not only needs to be interpretable (e.g., showing those elements that most contributed to the system decision), but that also needs to be joined by high level human argumentation schemes. The second challenge is to automatically learn explanatory argumentation schemas in Natural Language (NL) and to efectively combine evidence-based decision making with high level explanations. The third challenge for ANTIDOTE is that a task-specific prediction model and a general argumentation model need to be combined to produce explanatory argumentations.
While neural networks for medical diagnosis have become exceedingly accurate in many areas, their ability to explain how they achieve their outcome remains problematic. Herein lies the main novelty of the ANTIDOTE project: it focuses on elaborating argumentative explanations to diagnosis predictions in order to assist student clinicians to learn making informed decisions.
The explanatory argumentative scenario envisaged by ANTIDOTE will involve a student clinician, who will need to hypothesize about the clinical case of a patient and will have to provide argumentative explanations about them. The focus of the experimental setting is set on the capacity of the ANTIDOTE Explanatory AI system to provide correct predictions and consistent arguments, without forgetting also the linguistic quality of the dialogues (e.g., naturalness of the utterances, etc.).
In our scenario depicted in Figure 1, the clinician queries the ANTIDOTE XAI for explanations (arguments) on its diagnosis of the clinical case. The ANTIDOTE XAI provides hypotheses (diferential diagnosis) about the clinical case, as well as arguments to support its prediction and arguments discarding alternative predictions.
The student clinician has the possibility to take the initiative to ask additional questions and clarifications. The goal of the explanatory argumentation in a diferential diagnosis is to validate the correctness of the diagnosis and the ANTIDOTE XAI capacity to argue in favour of the correct hypothesis and to counter-argue against alternative hypotheses.
Argumentation mining is a research area that moves between natural language processing, argumentation theory and information retrieval. The aim of argumentation mining is to automatically detect the argumentation of a document and its structure. This implies the detection of all the arguments involved in the argumentation process, their individual or local structure (rhetorical or argumentative relationships between their propositions), and the interactions between them, namely, the global argumentation structure.
Argumentation mining in Natural Language Processing has been applied to various domains such as persuasive essays, legal documents, political debates and social media data [1]. For instance, Stab and Gurevych [2] built an annotated dataset of persuasive essays with corresponding argument components and relations. Using this corpus, Eger et al. [3] developed an end-to-end neural method for argument structure identification. Furthermore, Nguyen and Litman [4] also applied an endto-end method to parse argument structure and used the argument structure features to improve automated persuasive essay scoring. Other approaches studied contextdependent claim detection by collecting annotations for Wikipedia articles [5]. Using this corpus, the task of automatically identifying the corresponding pieces of evidence given a claim has also been investigated [6].
Argumentation generation remains a research area in which there is still a long way to go. Recent work has made progress towards this goal through the automated generation of argumentative text [7, 8, 9, 10]. Thus, Alshomary et al. [11] proposed a Bayesian argument generation system to generate arguments given the corresponding argumentation strategies. Sato et al. [9] presented a sentence-retrieval-based end-to-end argument generation system that can participate in English debating games.
Adadi and Berranda [18] presented an extensive literature review, collecting and analyzing 381 diferent scientific papers between 2004 and 2018. They arranged all of the scientific work in the field of explainable AI along four main axes and stressed the need for more formalism to be introduced in the field of XAI and for more interaction between humans and machines.
In a more recent study [19] introduced a diferent type of arrangement that initially distinguishes transparent and post-hoc methods and subsequently created sub-categories.
Taking into account argumentation principles, ANTIDOTE will explain machine decisions based on four modes of explanations to be auditable by humans: (i) analytic statements in NL that describe the elements and context that support a choice, (ii) visualizations that highlight portions of the raw data that support a choice, (iii) cases that invoke specific examples, and (iv) rejections of alternative choices that argue against less preferred answers based on analytics, cases, and data.
program/explainable-artificial-intelligence
argument generation to select the main talking points to generate a counter-argument [12]. In this line of research, Hidey and McKeown [13] proposed a neural model that edited the original claim semantically to produce a claim with an opposing stance. They also incorporated external knowledge into the encoder-decoder architecture show- 3. Methology and Work Plan ing that their model generated arguments that were more likely to be on topic. The main scientific challenge for the project is the com
Finally, an autonomous debating system (Project De- bination of three models depicted in Figure 1: (1) The bater) able to engage in competitive debates with humans Prediction Model has to predict appropriate International was developed. The system consisted of a pipeline of four Classification of Diseases (ICD) codes given a clinical main modules: argument mining, an argument knowl- case; (2) the Argumentative Model selects proper arguedge base, argument rebuttal, and debate construction ments (i.e., entity and relations) to support or attack a [14]. given topic. It may use both information included in the clinical cases used by the prediction model and additional 2.2. Explainable AI sources of knowledge; (3) the Interaction Model provides argumentative explanations about a certain prediction.
Explainable artificial intelligence (XAI) aims to address An integrated approach is proposed to both predict the the needs of users wanting to understand how a pro- outcome of a clinical course of action and justify a medigram’s artificial intelligence works and how to evaluate cal diagnosis by a language model. A starting point will the results obtained. Otherwise, there is no basis for real be using current large language models [20] to generate confidence in the work of the AI system, as illustrated appropriate explanations guided by the activated view on by Figure 22 The transparency ofered by explainable a textual snippet that contributed to the decision, namely, AI is therefore essential for the acceptance of artificial the argument for the decision. intelligence.
There has been a surge of interest in explainable artifi- 3.1. Work Plan cial intelligence (XAI) in recent years. This has produced a myriad of algorithmic and mathematical methods to The Work Plan is structured in six Work Packages of explain the inner workings of machine learning models which three are focused on the scientific contributions [15]. However, despite their mathematical rigor, these of the project. works sufer from a lack of usability and practical interpretability for real users. Although the concepts of interpretability and explainability are hard to rigorously define, multiple attempts have been made towards that goal [16, 17].
ipants: UCA, UPV/EHU, KU, NOVA. The purpose of WP2 is to define, adapt and integrate the modules, resources, data structures, data formats and module APIs of the ANTIDOTE architecture. This includes designing the experiments, datasets, standard protocols, information flow and main architecture of ANTIDOTE.
outcomes (Leader: KU). Participants: UPV/EHU, FBK, UCA, NOVA. WP3 targets (1) the develop- There are a number of tasks currently being undertaken ment of a multitask learning model to jointly pre- within the project. In this section we provide details of dict and justify a medical diagnosis by a deep the most central ones with respect to the objectives and learning model; (2) surfacing and making explicit motivation provided in the introduction. the underlying aspects (identification of the most relevant/informative terms, identification of re- 4.1. ANTIDOTE Datasets lations among terms) driving neural network decisions during the diagnosis prediction process; In order to carry out the tasks related to the main use-case (3) retrieve external information to support the presented in Figure 1, we need to identify, collect and explanation. annotate the most suitable corpus with which to train WP4: Explanatory arguments in natural language diferent models. In this regard, we have identified two (Leader: UCA). Participants: UPV/EHU, FBK, possible data sources that will help us meet our objectives UCA, NOVA. WP4 relies on the textual arguments and that will constitute an important contribution of the that form the basis for the decisions generated in ANTIDOTE project: SAEI and CasiMedicos. WP3. WP4 targets (1) the definition and analysis of explanatory argumentative patterns to be The SAEI Corpus is a collection of diferential diagused to construct natural language explanatory nosis in Spanish collected by La Sociedad Andaluza de arguments of predictions; (2) the creation of a re- Enfermedades Infecciosas (The Andalusian Society of Insource of annotated natural language explanatory fectious Diseases)3. This society is a non-profit associaarguments; (3) the development of explanatory tion formed almost entirely by physicians specializing arguments in natural language by mining and in Internal Medicine with special dedication to the mancollecting them from trusted textual resources in agement of infectious diseases, whose general purpose the medical domain. is the promotion and development of this medical disciWP5: Evaluation (use cases in healthcare) (Leader: pline (training, care and research). We have selected the UPV/EHU). Participants: KU, UCA, FBK, NOVA. books that are of interest to us in order to carry out our WP5 aims to (1) evaluate the efectiveness and objectives, namely, those that include clinical cases of quality of the prediction and the plausible alterna- infectious diseases for residents that are available, for the tives (2) the quality of the generated explanatory years 2011, 2015, 2016, 2017 and 2020. Among all these arguments regarding the supporting evidence books we have extracted cleaned and pre-processed a found in the clinical case in favor of the prediction total of 244 clinical cases with diferential diagnosis. and the positive or negative evidence found to discard other plausible alternatives, (3) the intrinsic quality of the generated arguments.
ical project run by volunteer medical doctors4. Among all the information created and made publicly available 3.2. Evaluation by this collaborative project, we have identified as an adequate data source the MIR exams commented by volThe generation of arguments will be quantitatively eval- untary medical doctors with the aim of providing answers uated by computing metrics used in text generation to and explanations5 to the MIR exams annually published measure their overlap with ground truth arguments [21]. by the Spanish Ministry of Health. In this data source Moreover, the argumentative model will be evaluated we have extracted and pre-processed 622 commented following the criteria of coherence, simplicity, and gen- questions from the MIR exams held between the years erality [22]: explanations with structural simplicity, co- 2005, 2014, 2016, 2018, 2019, 2020, 2021 and 2022. The herence, or minimality are preferred. With respect to cleaned corpus, named the Antidote Casimedicos dataset, argument mining, standard metrics such as F1 and accu- is publicly available to encourage research on explainable racy will be used. AI in the medical domain in general, and argumentation
Generation of explanatory arguments will be also qual- in particular6. itative evaluated by medical students. Given the objec- Unlike popular Question Anwering (QA) datasets for tives and context of the project, ANTIDOTE will be based English based on medical exams [24], both SAEI and on previous work by Johnson [23], whereby the argu- CasiMedicos include not only the explanations for the ments will be evaluated for their (informal) inferential structure in terms of acceptability, relevance, and sufi- 43hhttttppss::////wwwwww..scaaesiim.oregdicos.com/ ciency of reasons provided, as well as their answerability 5https://www.casimedicos.com/mir-2-0/ to human agents’ doubts and objections. 6https://github.com/ixa-ehu/antidote-casimedicos corrent answer (diagnosis or treatment), but also explana- approaches to Question Answering techniques in the tory arguments written by medical doctors explaining medical domain. why the rest of the possible answers are incorrect.
After pre-processing, these datasets have been translated from Spanish to English with the objective of start- 5. Concluding Remarks ing various annotation tasks at various levels of complexity: (i) linking the explanatory sequences with respect to each possible answer; (ii) labeling of hierarchical argumentative structures; (iii) discourse markers. The resulting corpus will be the first corpus (multilingual or otherwise) with this type of annotations for the medical domain. Whenever ready, the corpus will be distributed under a free license to promote further research and to ensure reproducibility or results.
project, mostly focusing on identifying and generating high-quality argumentative explanations for AI predictions in the medical domain. So far, ongoing work has been focused on dataset collection and annotation and novel experimental work on Question Answering and Crosslingual Argument Mining. This work has leveraged multilingual encoder and decoder large language models [28, 20] for both extractive and generative experimentation.
Still, providing high-quality explanations for AI predictions based on machine learning is a challenging and complex task [24]. To work well, it requires, among other factors, making use of additional knowledge (e.g. medical evidence) which might not be part of the prediction process, and providing evidence supporting negative hypotheses. With these issues in mind, ANTIDOTE aims to address the challenge of providing an integrated vision of explainable AI, where low-level characteristics of the deep learning process are combined with higher level schemes proper of the human argumentation capacity. In order to do so, ANTIDOTE will be focused on a number of deep learning tasks for the medical domain, where the need for high quality explanations for clinical cases deliberation is critical.
on medical exams [24], none of the previously published works contain two features which are unique of both SAEI and CasiMedicos: (i) the presence of explanations for both correct and incorrect answers; (ii) an argumentative structure arguing and counter-arguing about the possible answers. These features make it possible to define new Question Answering tasks, both from an extractive and generative point of view. In extractive QA, the objective would consist of identifying, in a given context, the explanation to the correct answer. In terms of generative QA, it will also allow us to leverage large language models [20] to learn generating the explanatory arguments with respect to both correct and incorrect possible answers. 4.3. Crosslingual Knowledge Transfer We thank the CasiMedicos Proyecto MIR 2.0 for their permission to share their data for research purThe only corpus annotated with argumentative structure poses. ANTIDOTE (PCI2020-120717-2) is a project currently available for the medical domain is the AbstRCT funded by MCIN/AEI/10.13039/501100011033 and by dataset, which consists of English clinical trials [25]. In European Union NextGenerationEU/PRTR. Rodrigo order to investigate the diferent strategies of transfer- Agerri currently holds the RYC-2017-23647 fellowship ring knowledge from English to other languages, espe- (MCIN/AEI/10.13039/501100011033 and by ESF Investcially those applying model- and data-transfer techniques ing in your future). Iker García-Ferrero is supported previously discussed for other application domains [26], by a doctoral grant from the Basque Government ongoing work is focused on adapting such knowledge (PRE_2021_2_0219) and Anar Yeginbergenova acknowltransfer techniques for argumentation in the medical do- edges the PhD contract from the UPV/EHU (PIF 22/159). main. As a result, we are undertaking novel experimental work on argument mining in Spanish for the medical domain [27]. This also involves the generation of the first References Spanish dataset annotated with argumentative structures for the medical domain. Finally, the plan is to apply the [1] M. Dusmanu, E. Cabrio, S. Villata, Argument mindeveloped technique to other languages of interest for ing on twitter: Arguments, facts and sources, in: the ANTIDOTE project (French and Italian). Proceedings of the 2017 Conference on Empirical
In this line of research, and taking as starting point Methods in Natural Language Processing, 2017, pp. the ongoing work mentioned in the previous section, 2317–2322. we plan to investigate also crosslingual and multilingual [2] C. Stab, I. Gurevych, Parsing argumentation struc- Nature 591 (2021) 379–384.
tures in persuasive essays, Computational Linguis- [15] O. Biran, C. Cotton, Explanation and justification in tics 43 (2017) 619–659. machine learning: A survey, in: IJCAI-17 workshop [3] S. Eger, J. Daxenberger, I. Gurevych, Neural end- on explainable AI (XAI), volume 8, 2017, pp. 8–13. to-end learning for computational argumentation [16] Z. C. Lipton, The Mythos of Model Interpretability: mining, arXiv preprint arXiv:1704.06104 (2017). In machine learning, the concept of interpretability [4] H. Nguyen, D. Litman, Argument mining for im- is both important and slippery., Queue 16 (2018) proving the automated scoring of persuasive essays, 31–57. in: Proceedings of the AAAI Conference on Artifi- [17] F. Doshi-Velez, B. Kim, Towards a rigorous science cial Intelligence, volume 32, 2018. of interpretable machine learning, arXiv preprint [5] R. Levy, Y. Bilu, D. Hershcovich, E. Aharoni, arXiv:1702.08608 (2017).
N. Slonim, Context dependent claim detection, [18] A. Adadi, M. Berrada, Peeking inside the black-box: in: Proceedings of COLING 2014, the 25th Inter- a survey on explainable artificial intelligence (XAI), national Conference on Computational Linguistics: IEEE access 6 (2018) 52138–52160.
Technical Papers, 2014, pp. 1489–1500. [19] A. B. Arrieta, N. Díaz-Rodríguez, J. Del Ser, A. Ben[6] R. Rinott, L. Dankin, C. Alzate, M. M. Khapra, netot, S. Tabik, A. Barbado, S. García, S. Gil-López, E. Aharoni, N. Slonim, Show me your evidence-an D. Molina, R. Benjamins, et al., Explainable Arautomatic method for context dependent evidence tificial Intelligence (XAI): Concepts, taxonomies, detection, in: Proceedings of the 2015 conference opportunities and challenges toward responsible on empirical methods in natural language process- AI, Information fusion 58 (2020) 82–115. ing, 2015, pp. 440–450. [20] L. Xue, N. Constant, A. Roberts, M. Kale, R. Al-Rfou, [7] R. Bar-Haim, L. Eden, R. Friedman, Y. Kantor, D. La- A. Siddhant, A. Barua, C. Rafel, mT5: A massively hav, N. Slonim, From arguments to key points: To- multilingual pre-trained text-to-text transformer, wards automatic argument summarization, arXiv in: NAACL, 2021.
preprint arXiv:2005.01619 (2020). [21] T. Zhang, V. Kishore, F. Wu, K. Q. Weinberger, [8] X. Hua, L. Wang, Neural argument generation aug- Y. Artzi, BERTScore: Evaluating Text Generation mented with externally retrieved evidence, arXiv with BERT, in: International Conference on Learnpreprint arXiv:1805.10254 (2018). ing Representations (ICLR), 2020. [9] M. Sato, K. Yanai, T. Miyoshi, T. Yanase, [22] T. Miller, Explanation in artificial intelligence: InM. Iwayama, Q. Sun, Y. Niwa, End-to-end argument sights from the social sciences, Artificial intelligeneration system in debating, in: Proceedings of gence 267 (2019) 1–38.
ACL-IJCNLP 2015 System Demonstrations, 2015, [23] R. H. Johnson, Manifest rationality: A pragmatic pp. 109–114. theory of argument, Routledge, 2012. [10] M. Alshomary, S. Syed, A. Dhar, M. Potthast, [24] K. Singhal, S. Azizi, T. Tu, S. S. Mahdavi, J. Wei, H. Wachsmuth, Argument Undermining: Counter- H. W. Chung, N. Scales, A. Tanwani, H. ColeArgument Generation by Attacking Weak Premises, Lewis, S. Pfohl, P. Payne, M. Seneviratne, P. Gamble, arXiv preprint arXiv:2105.11752 (2021). C. Kelly, N. Scharli, A. Chowdhery, P. Mansfield, [11] M. Alshomary, S. Syed, M. Potthast, H. Wachsmuth, B. A. y Arcas, D. Webster, G. S. Corrado, Y. Matias, Target inference in argument conclusion genera- K. Chou, J. Gottweis, N. Tomasev, Y. Liu, A. Ration, in: Proceedings of the 58th Annual Meeting jkomar, J. Barral, C. Semturs, A. Karthikesalingam, of the Association for Computational Linguistics, V. Natarajan, Large language models encode clinical 2020, pp. 4334–4345. knowledge, in: arXiv 2212.13138, 2022. [12] X. Hua, Z. Hu, L. Wang, Argument generation with [25] T. Mayer, S. Marro, E. Cabrio, S. Villata, Enhancing retrieval, planning, and realization, arXiv preprint Evidence-Based Medicine with Natural Language arXiv:1906.03717 (2019). Argumentative Analysis of Clinical Trials, Artificial [13] C. Hidey, K. McKeown, Fixed that for you: Gen- Intelligence in Medicine (2021) 102098. erating contrastive claims with semantic edits, in: [26] I. García-Ferrero, R. Agerri, G. Rigau, Model and Proceedings of the 2019 Conference of the North data transfer for cross-lingual sequence labelling in American Chapter of the Association for Computa- zero-resource settings, in: Findings of the Associational Linguistics: Human Language Technologies, tion for Computational Linguistics: EMNLP 2022, Volume 1 (Long and Short Papers), 2019, pp. 1756– 2022.
1767. [27] A. Yeginbergenova, R. Agerri, Cross-lingual ar[14] N. Slonim, Y. Bilu, C. Alzate, R. Bar-Haim, B. Bogin, gument mining in the medical domain, in: arXiv F. Bonin, L. Choshen, E. Cohen-Karlik, L. Dankin, 2301.10527, 2023.
L. Edelstein, et al., An autonomous debating system, [28] J. Devlin, M. Chang, K. Lee, K. Toutanova, BERT: pre-training of deep bidirectional transformers for language understanding, in: NAACL-HLT, 2019, pp. 4171–4186.