e-government In this paper, we present ongoing work on the automatic extraction of arguments from textual content, and on the use of interconnected argument structures by recommender systems. Diferently to the majority of existing argument mining methods -which only consider 'premise' and 'claim' as the components of an argument, and 'support' and 'attack' as the possible relations between argument components-, we propose an argumentation model based on a detailed taxonomy of argumentative relations. Moreover, we provide a lexicon of English and Spanish linguistic connectors categorized in our taxonomy. As a proof of concept, we apply a simple, yet efective method that makes use of the built taxonomy and lexicon to extract argument graphs from citizen proposals and debates of an e-participation platform. We then describe how the extracted graphs could be exploited to generate and explain argument-based recommendations. argument-based recommender systems, recommendation explanations, argument mining, natural language processing, ized by textual content, such as books [4], scientific pubtechniques [3], but also to address domains character- both traditional recommendation domains, such as elications [5], news articles [6], and online reviews [7]. are plenty of user reviews--, and less common domains descriptions-, and have distinct levels of linguistic for- transcripts of political speeches, and collections of citizen
The data sources in these domains are heterogeneous in nature and form –ranging from well-defined categories and freely-chosen (social) tags to natural language texts of diferent length, e.g., titles, summaries, and long mality and explicit/implicit structure complexity. These text characteristics, as well as own particularities of natural language (e.g., misspellings, ambiguity, irony) make the content-based recommendation a challenging task.
In this context, many research eforts have been devoted to recommendation approaches aimed to exploit opinions expressed as natural language in unstructured, free-form texts. The opinions can be detailed and focused on a particular item and its aspects, such as those provided in blogs and reviews [8, 7, 9], or can be short nEvelop-O 3rd Edition of Knowledge-aware and Conversational Recommender Systems (KaRS) & 5th Edition of Recommendation in Complex ment structures by recommender systems. Diferently to methods existing in the argument mining field [ 14, 13], which only consider ‘premise’ and ‘claim’ as the components of an argument, and ‘support’ and ‘attack’ (rebuttal) as the possible relations between argument components, The latter works present implementations and evaluawe propose an argumentation model based on a detailed tions of classic content-based, collaborative filtering and taxonomy of argumentative relations. The taxonomy is hybrid recommendation methods that exploit a variety then populated with a lexicon of linguistic connectors for of user-generated content, such as social tags and votes, both English and Spanish, and is preliminary exploited by as well as item (citizen proposal) metadata based on catesimple, yet efective argument extraction and argument- gories, topics, and geographic locations. Diferently to based recommendation methods. As a proof of concept, these works, in this paper, we advocate for recommender we report some results on generated arguments, recom- systems that dig into the semantics underlying the texts mendations, and recommendation explanations for the of the citizens’ proposals and comments. Hence, we aim e-participation domain, in which graphs of arguments to investigate recommendation approaches that exploit exist around citizen proposals and debates. the arguments provided by citizens, in favour or against the created proposals.
In this section, we describe some representative works of the topics addressed in our research. Specifically, we survey recommender systems targeting the e-participation domain (section 2.1) and exploiting argumentative information (section 2.2), and we provide major references on argument mining (section 2.3).
As explained in [15], recommendation solutions are of increasing interest and application for numerous problems, tasks and challenges of (smart) cities. In the paper, the authors survey the academic literature on recommender systems for the principal six dimensions of smart cities, namely economy, environment, mobility, governance, living and people.
With respect to the governance dimension,
recommender systems have been mainly proposed to facilitate
the access to government information and increase
eficiency in municipal management –e.g., by providing
personalized suggestions of electronic government
notifications and services [
In [24], the authors discuss recommender systems for egovernance, diferentiating nine use cases in governmentto-citizen (G2C), government-to-business (G2B), and government-to-government (G2G) e-services. From them, we focus on the G2C case where users are assisted in finding relevant citizen proposals and debates generated in e-participation tools. In this case, among other applications, recommender systems have been used as information filtering mechanisms for e-participatory budgeting (ePB) platforms [25, 26], where citizens propose and debate online a large number (hundreds or even thousands) of ideas, initiatives and projects aimed to address municipal issues.
Surveying the academic literature on argument-based recommender systems, two main groups of researches can be identified. The first group refers to recommendation methods that are based on Defeasible Logic Programming (DeLP) [27]. DeLP is a computational reasoning framework that consists of an argumentation engine operating over a knowledge base expressed in a logic programming language, which accepts encoded facts, and strict and defeasible rules (constraints). In the context of recommender systems, DeLP allows defining as rules user tastes and interests, item features and relations, and contextual conditions [28, 29, 30]. Hence, the engine reasons over a set of defined rules in order to infer potential preferences of users for certain items, that is, to provide lists of item recommendations in an argumentative fashion. Presenting the set of rules applied (satisfied) in such process, DeLP enables the explanation of generated recommendations. It, however, requires building the argument knowledge base, which to date has been done manually [31] or has been limited to simple, automatic transformations of relational databases [32].
The second group is composed of approaches aimed to provide argumentative explanations of recommendations, regardless the filtering algorithm used. In this case, arguments mainly represent relationships between user preferences and item attributes. In [33], the authors propose a framework where diferent types of justification (e.g., ethical, aesthetic) are given for generated recommendations depending on the users’s preferences and according to manually defined rules. In [ 11], the authors address the task of predicting the usefulness of review fragments according to their argumentative content. The estimated usefulness is used to rank the reviews associated to recommended items. Also focusing on user reviews, in [34], the authors propose to identify aspects important for the target user through an attention neural network model, extract and summarize relevant arguments (opinions) about such aspects, and present the arguments as textual explanations of personalized recommendations. A related approach is followed in [35], approaches. On the other hand, a variety of tools are where the authors propose a method that generates ex- available for diferent purposes, such as argumentative planations in an argumentative manner by presenting modeling (e.g., Agora,1 Argunet,2 DebateGraph3 and Raan incremental selection of positive and negative state- tionale Online4), and argument-based text annotation ments that support or contradict recommended items (e.g., Araucaria5 and OVA6). and their aspects, according to opinions expressed in In this paper, we i) preliminary experiment with a user reviews. Lastly, without taking user reviews into simple, yet efective syntactic pattern-based method to account, in [36], the authors exploit Linked Open Data argument extraction (addressing the three main AM tasks to extract descriptive properties about items, and use the explained above), and ii) provide new resources for the extracted properties to feed graph-based explanations of AM community; specifically, a detailed argument relarecommended items. These explanations are generated tion taxonomy that goes beyond the premise-claim and through argumentative, natural language templates. support-attack models, and a lexicon of English and Span
For both groups, to the best of our knowledge, and ish linguistic connectors associated to the taxonomy catdiferently to our proposal, published argument-based egories. recommender systems do not make use of argument mining methods and resources to automatically extract argumentative information from textual content, and exploit 3. Case study such information during the item filtering process.
In this section, we introduce Decide Madrid,7 an eparticipation platform for which we have preliminary 2.3. Argument mining tested our argument mining and argument-based recomEmerged from the confluence of the Computational Lin- mendation methods. guistics (CL) and Natural Language Processing (NLP) Among other citizen participation methods, Decide areas, Argument Mining (AM) [13] is a relatively young Madrid is an online website used by the Madrid City ifeld that dates back to the late 2000s. In [ 37], it was for- Council for its annual participatory budgets. Since mulated with the general aim of automatically extracting September 2015, every year, city residents are allowed structured, argumentative information from text. to freely upload, comment and vote for proposals aimed
This research challenge has been commonly mod- to address city problems and initiatives. A citizen
proeled as a pipeline of three (consecutive) tasks: argu- posal is composed of the following data: title, description,
ment detection [38, 39, 40], argument component identifi- author, date, tags, multimedia elements (i.e., pictures,
cation [
To date, these tasks have been mostly addressed sep- government’s approval. arately through machine learning methods [39, 42], but The large number of proposals, which also occurs in recently, they have been jointly treated as sequence la- e-participatory budgeting processes of other big cities belling tasks of NLP, addressed by specialized neural worldwide, has motivated the investigation of recomnetwork models [44]. In both cases, the desired, final mender systems to assist on the exploration of proposoutcome of the AM process is a tree or graph structure als [24, 26]. Published recommenders have exploited that semantically interconnects the arguments existing content-based (e.g., topics, categories) and collaborative in an input text. (e.g., supports/votes) data of the proposals. However,
Additionally to algorithmic solutions, significant advances have been made on the development of linguistic resources. On the one hand, there are a number of corpora annotated with structured argument information from diferent sources –such as persuasive essays, online debates, and news media items (cf. [13] for a detailed survey)–, which can be used to build and evaluate AM 1http://agora.gatech.edu 2https://sourceforge.net/projects/argunet 3https://debategraph.org 4https://www.rationaleonline.com 5http://staff.computing.dundee.ac.uk/creed/araucaria 6http://ova.arg-tech.org 7https://decide.madrid.es 8https://datos.madrid.es they have not considered the textual content of the pro- 4.2. Argument relation taxonomy and posals’ descriptions and comments. In the ongoing work lexicon presented in this paper, by contrast, we advocate for the use of such content, in particular, its underlying argumentative information.
model that we propose to follow aims to consider a variety of relations that go beyond the support-attack schema.
Surveying the academic literature, we find studies that
4. Argument mining framework have have presented distinct types of relations, and have
compiled sets of linguistic connectors (or indicators)
asIn this section, we present our framework to automati- sociated to such types.
cally identify arguments in textual content, split them For instance, in [
The Toulmin’s model [45] is one of the most popular tion via consequence, contrast, and concession. Lastly, argument models. It structures an argument into six com- in [47], the authors consider a total of 115 lexical indiponents: the claim (i.e., the conclusion of the argument), cators categorized as ‘forward’ (e.g., as a result, because, the ground (i.e., the premise, foundation or basis for the thus), ‘backward’ (e.g., additionally, besides, moreover ), claim), the warrant (i.e., the reasoning that legitimizes ‘thesis’ (e.g., all in all, finally , in conclusion), and ‘rebutthe claim by showing the relevance of the ground), the tal’ (e.g., but, however, though) indicators. Regardless backing (i.e., the support for the warrant), the qualifier these taxonomies, one can find works (e.g., [ 42, 48, 49]) (i.e., the degree of certainty of the claim), and the rebuttal that also provide lists of connectors used as features of (i.e., an exception that may apply to the claim). machine learning models for AM tasks.
In CL in general and in AM in particular, however, the Carefully revising and jointly considering all these majority of existing computational methods and tools references, we have developed a two-level taxonomy of to design, extract and share arguments follow simpler argument relations, and have gathered a relatively large argument models [13]. Specifically, most of them only set of linguistic connectors classified with the taxonomy. consider premises and claims as argumentative units, and The taxonomy and the set of connectors, referred as an support and attack (rebuttal) as argument relations. ‘argument relation lexicon,’ are made accessible online9
Our argument model extends this basic representa- in English and Spanish. tion as follows. First, as done in some works [13], in Table 1 shows the categories and subcategories of the addition to premises and claims, we also consider major proposed taxonomy, with their primary intents (i.e., supclaims as fundamental argument units. They refer to port, attack, qualifier ), and gives some examples of Enthe principal, resultant parts of argumentative chains glish and Spanish connectors of each (sub)category. within a discourse. Hence, other claims (and premises) As it can be seen, our taxonomy includes the following relate or depend on major claims. Second, instead of types of argument (component) relations: narrowing the scope to support and attack relations, we take more fine-grained relation types into account, e.g., • Cause. This relation links an argument that reby distinguishing whether an attack really represents lfects the reason or condition for another arguan opposition or, on the contrary, it suggests an alterna- ment. tive, a comparison or a concession for an argument. The considered argument relation types form a taxonomy, as explained next.
9Developed taxonomy and lexicon, https://github.com/ argrecsys/connectors • Clarification . This relation introduces a conclu- people, organizations, places) of the sentence are also recsion, exemplification , restatement or summary of ognized to enrich the underlying arguments, since they an argument. could be considered to relate the diferent arguments, in • Consequence. This relation evidences an explana- addition to their topics.
tion, goal or result of a previous argument. On the sentence, constituency parsing is finally con• Contrast. This relation links attacking arguments, ducted to extract a parse tree that represents the syndistinguishing between several types of attack: tactic structure (i.e., interconnected phrases) of the sengiving alternatives, doing comparisons, making tence. This structure will be used to recursively group the concessions, and providing oppositions. phrases of the sentence. Within the built phrase groups, • Elaboration. This relation introduces an argument syntactic patterns –e.g., in the premise-connector-claim that provides details about another one. The de- form– will be searched, thus identifying the existing artails can entail addition, precision or similarity guments. issues about the target argument. All these NLP tasks are performed using the Stanford CoreNLP [50] library, both for English and Spanish. 4.3.2. Identifying arguments
sentences that have connectors, using the outputs generated in the previous phase.
Specifically, the grouped phrases (obtained from the constituency parsing process) are traversed from the bottom to the top of the sentence constituency tree, and are matched with predefined syntactic patterns. For the moment, arguments are recognized as matches with any of the following two patterns: [{ _ } + + { formed by three grouped phrases. [{ _ } + [ + { formed by two grouped phrases. _ }] _ }]]
and 384 Spanish connectors. As shown in the table, the English connectors are evenly distributed into the taxonomy categories (with an average of 44.5 connectors per category), except the contrast category, which is the only one with (70) connectors whose primary intent is ‘attack.’
This method is a simple heuristic approach that aims to automatically identify and extract arguments from textual content using basic syntactic patterns. It performs in a simple but efective way the three basic tasks of argument mining, namely: argument detection (from citizen proposals), argument component identification (i.e., claims and premises linked through a connector), and argument relation recognition using the proposed taxonomy and lexicon. For such purpose, the method is divided into two (consecutive) phases, where the output of the first phase serves as input for the second phase.
In particular, the phases are processing natural language and identifying arguments (and their relations).
In the patterns, both the claim and the premise can contain the main verb of the sentence. The verb is first searched within the claim (it is more likely to be found here), and then within the premise. In the future, other more complex syntactic patterns could be considered, since they are easy to integrate into our method. 4.3.1. Processing natural language Once one of the two aforementioned patterns is matched, the sentence is split into claim and premise In this phase, the source text –i.e., a citizen proposal according to and linked with the sentence connector (exdescription– is first split into sentences, where arguments isting in the lexicon). The identified argument structure is will be searched (isolatedly in this stage of our work). finally stored into a JSON data type along with: i) the conOnly those sentences that contain at least one of the nector and its argument relation category, sub-category connectors in the proposed lexicon are then taken into and primary intent, ii) the sentence lists of nouns, verbs account. and named entities, iii) the main verb of the argument,
For a given sentence, part-of-speech (PoS) tags are ex- and iv) the identifier of the citizen proposal where the tracted, identifying the grammatical category (i.e., noun, argument was found. verb, adjective, adverb, etc.) of each word. In this process, Figure 1 shows an example in JSON format of an arthe identified verbs are stored into a list, which will be gument extracted from a citizen proposal on a specific used to establish the main verb (action) of an argument, topic: public transportation. In this example, the premise and the nouns are stored in another list, which will be directly attacks the claim of the argument, in order to used to set the possible topics or aspects the argument support (by contrast) the major claim, extracted from the refers to. All this information could be exploited by a rec- citizen proposal title. ommendation method as well. The named entities (e.g., actually, in [actual] fact, indeed, of course, for that matter for, to, in order to, aimed/aiming to, that/which allows/entails/implies therefore, thus, hence, then, so [that] as a result [of], this/that/such reason, accordingly, in/as a consequence on the other hand, in another case, 21 if not, instead [of], rather than, alternatively [to], otherwise, else while, whereas, compared [to/with], 7 in comparison to/with, as long as
although, [even] though, despite [that], 20 in spite/despite of, regardless [of]
but, however, nonetheless, albeit, 22 nevertheless, in contrast [to/with] 35 21 56 19 14 34 12 79 8 18 44 70 29 20 38 46 133 46 384 si [alguna vez/es así], en caso de/que con/bajo la condición de [que], a no ser que porque, ya que, debido a [que], pues, dado que, basándose en [que], puesto que para concluir, en/como conclusión, en definitiva, atendiendo a/con [todo] lo considerado por ejemplo, como ejemplo [de], tales como, por dar/poner un ejemplo [de] en otras palabras, es decir, esto es, mejor dicho, dicho de otro modo resumiendo, concluyendo, para acabar, por resumir/concluir, en pocas palabras realmente, de hecho, en realidad, por supuesto, en efecto, para el caso para, por, con el fin de, lo que/cual permite/conlleva/implica por [lo] tanto, por consiguiente/ende como resultado, por esta/esa razón, así que, es por ello que, de este/ese modo por otra parte, por otro lado, en otro caso, si no, en vez/lugar de, en cambio/su defecto, alternativamente [a], de otro modo mientras [que], comparado con, en comparación a/con, a la vez de/que aunque, aún/incluso [si/así], a pesar de/del, a pesar de que, pese a [que], pese al pero, sin embargo, no obstante, en contraste a/con, en contra [de/del] Cause
Condition Reason qualifier support Conclusion
support Exemplification support Alternative
support/attack Comparison
support/attack Restatement Summary Explanation Goal Result Concession Opposition Addition Similarity support support support support support support/attack attack support support support also, besides, as well, too, moreover, furthermore, additionally, in addition [to] in particular, particularly, especially, mainly, [more] specifically/precisely similarly/analogously [to], like, likewise, in the same way, correspondingly
también, además/aparte [de], [lo que] es más, 22 asímismo, encima de, adicionalmente [a]
en particular, particularmente, especialmente, 13 principalmente, [más] especificamente/ precisamente similarmente/analogamente [a], como, al 11 igual que, del mismo modo [que], de la misma manera [que]
To conclude, we present some statistics from a preliminary ofline test (with a subset of lexicon connectors) on the automatic identification and extraction of arguments from the citizen proposals available in the Decide Madrid database: • From a reduced list of 10 connectors (belonging to the CAUSE and CONTRAST categories), 1,744 proposals with possible arguments were identiifed out of the 21,744 proposals available. • Arguments were automatically extracted in 1,362 of the 1,744 proposals identified, entailing a coverage of 78.0%.
• Of the 1,379 arguments extracted (some proposals had more than one argument), 1,034 were identiifed with connectors from the CAUSE category and 345 from the CONTRAST category. • An accuracy of 78.8% was achieved in a manual evaluation of 47 arguments about public transportation. }, ”pattern”: ”P1 -> CLAIM + CONNECTOR + PREMISE”
recommendations for the target topic, and the arguments that support or attack these proposals grouped by topics and aspects.
A contribution of our work is the proposal of this new Once the arguments are automatically identified and recommendation paradigm, which is based on the minextracted from a set of citizen proposals, they can be ing of arguments, that is, instead of just recommending exploited as complex inputs of an argument-based rec- proposals that satisfy a user’s information needs ( → ommendation method. As a proof of concept, given a → ), we propose to recommend proposals particular topic –e.g., public transportation–, we consider that have arguments concerning the user’s topics of intera recommender that, via content-based filtering, first re- est ( → / → → ), in trieves and filters proposals about the topic, and then con- order to not only filter relevant information for the user, siders the arguments given in such proposals to rerank but also to assist her on decision making tasks. Moreover, and present recommended proposals (and arguments). the proposed approach allows creating in a direct and pre
More specifically, from the selected proposals and their cise way explanations of the generated argument-based associated arguments, the recommender identifies the recommendations. The following are possible explanadiscussed aspects of the topic of interest (e.g., price, lo- tion templates: cation, quantity) for which there are arguments in favor • “[These] citizen proposals about [this] topic are or against. With these aspects, the recommender builds recommended because they have the following a graph that relates proposals, topics, aspects and ar- supporting (attacking) arguments...” guments, and exploits such graph to find relevant (i.e., highly connected) proposals which are recommended to • “Regarding [these] aspects on [this] topic of inthe user. terest, the following proposals are recommended
These proposals are presented along with their respec- since they have more arguments in favor” tive arguments in the form, claim-connector-premise for We believe that these types of recommendations and each aspect. Figure 2 shows a subset of recommended explanations not only may help improving the efectiveproposals about public transport in the context of the ness of the system, but also may increase its transparency Decide Madrid e-participation platform. The output of and foster the user’s trust. the argument-based recommender is an XML file which is composed of two blocks: the recommended proposals
The ongoing work presented in this paper has resulted in a novel taxonomy of argumentative relations that goes beyond the commonly adopted support-attack schema, and a rich lexicon of argument connectors for both English and Spanish. The use of these resources has been preliminary exemplified through the automatic extraction of arguments from text contents, and the generation of argument-based recommendations in a real eparticipation case study, where graphs of interconnected topics, premises and claims underlay citizen proposals and debates.
We believe that this new paradigm of argument-based
recommendation, which provides transparency in the
form of intuitive, justified explanations, can not only
be applicable to other e-government contexts –such as
parliamentary debate [51] and political discussion in
social networks [
There are, however, many research lines that should be addressed before. First, we have to conduct more sophisticated text processing, e.g., by correcting misspellings, dealing with lexical and syntactic variations, and better identifying named entities. We then have to extend our extraction method with additional syntactic argument patterns, and other argument features diferent to linguistic connectors, as done by other methods in the argument mining field [ 14, 13]. We also have to formally define recommendation methods that exploit argument structures in both the filtering and explanation phases. These methods could be empirically compared with existing argument-based recommenders, e.g., based on DeLP frameworks [27, 28], since they could be built on knowledge bases generated with argument extraction methods. In this context, we should explore and evaluate recommendation explanations in a more natural language form [36]. For such purpose, we expect that user studies will be needed. In these studies, we may also consider evaluating potential benefits of argument-based recommendations, such as transparency, fairness, and accountability [53]. These issues are of special interest in e-participation contexts such as the one addressed in this work, where, among others, controversy topics and minority groups are of high relevance [54].
cities, Information Systems 92 (2020) 101545. H. Alani, What’s going on in my city? Recom[16] M. Eirinaki, S. Dhar, S. Mathur, A. Kaley, A. Patel, mender systems and electronic participatory budA. Joshi, D. Shah, A building permit system for geting, in: Proceedings of the 12th ACM Confersmart cities: A cloud-based framework, Computers, ence on Recommender Systems, 2018, pp. 219–223.
Environment and Urban Systems 70 (2018) 175–188. [27] A. J. García, G. R. Simari, Defeasible logic
program[17] M. Janssen, R. Wagenaar, J. Beerens, Towards ming: An argumentative approach, Theory and
a flexible ICT-architecture for multi-channel e- Practice of Logic Programming 4 (2004) 95–138.
government service provisioning, in: Proceedings [28] C. I. Chesñevar, A. G. Maguitman, ArgueNet: An
of the 36th Annual Hawaii International Confer- argument-based recommender system for solving
ence on System Sciences, IEEE, 2003, pp. 10–pp. web search queries, in: Proceedings of the 2nd
In[
62–76. [34] T. Donkers, B. Loepp, J. Ziegler, Explaining rec[23] K. Dyczkowski, A. Stachowiak, A recommender ommendations by means of user reviews, in: Joint system with uncertainty on the example of political Proceedings of the ACM IUI 2018 Workshops, ACM, elections, in: Proceedings of the 14th International 2018.
Conference on Information Processing and Manage- [35] S. Naveed, T. Donkers, J. Ziegler, Argumentationment of Uncertainty in Knowledge-Based Systems, based explanations in recommender systems: ConSpringer, 2012, pp. 441–449. ceptual framework and empirical results, in: Ad[24] M. E. Cortés-Cediel, I. Cantador, O. Gil, Recom- junct Publication of the 26th Conference on User mender systems for e-governance in smart cities: Modeling, Adaptation and Personalization, ACM, State of the art and research opportunities, in: Pro- 2018, pp. 293–298. ceedings of the 1st International Workshop on Rec- [36] C. Musto, F. Narducci, P. Lops, M. de Gemmis, G. Seommender Systems for Citizens, 2017, pp. 1–6. meraro, Linked open data-based explanations for [25] I. Cantador, A. Bellogín, M. E. Cortés-Cediel, O. Gil, transparent recommender systems, International Personalized recommendations in e-participation: Journal of Human-Computer Studies 121 (2019) Ofline experiments for the ’Decide Madrid’ plat- 93–107. form, in: Proceedings of the International Work- [37] R. Mochales Palau, M.-F. Moens, Argumentation shop on Recommender Systems for Citizens, 2017, mining, Artificial Intelligence and Law 19 (2011) pp. 1–6. 1–22. [26] I. Cantador, M. E. Cortés-Cediel, M. Fernández, [38] B. Hachey, C. Grover, Automatic legal text summari