1. Introduction

1As the amount of linked data published on the Web keeps increasing, there is an expanding demand for multilingual tools and user interfaces that simplify the access and browsing of data by end-users, so that information can be explored in an intuitive way. This need is what motivated the development of tools such as Question Answering (QA) systems, whose main aim is to make users be able to explore complex datasets and an ever growing amount of data in an intuitive way, through natural language.

2While the default approach for many NLP tasks has recently been represented by machine learning systems, the use of such approaches (Chakraborty et al. 2019) for QA over RDF data suffers from lack of controllability, making the governance and incremental improvement of the system challenging, not to mention the initial effort of collecting and providing training data for a specific language.

3An alternative is the so-called model-based approach to QA, in which a model is first used to specify how concepts and relations are realized in natural language, and then this specification is employed to interpret questions from users. One such system is the one proposed by  (Benz et al. 2020), which makes use of a lexicon in lemon format  (McCrae, Spohr, and Cimiano 2011) to specify how the vocabulary elements of an ontology or knowledge graph (e.g., entities and relations from a Knowledge Graph) are realized in natural language.

4The previous work on this approach shows how, leveraging on lemon lexica, question answering grammars can be automatically generated, and those can, in turn, be used to interpret questions and then parse them into SPARQL queries. A QA web application developed in previous work  (Elahi et al. 2021) has further shown that such QA systems can scale to large numbers of questions and that the performance of the system is practically real-time from an end-user perspective.

5In this work we describe the extension to the Italian language of the model-based approach described in  (Benz et al. 2020) and the QA system described in  (Elahi et al. 2021). By doing so, we develop a QA system that can answer more than 1.6 million Italian questions over the DBpedia knowledge graph1.

2. Related Work

6Besides the goal of creating QA systems that are robust and have high performance, an important goal is also to develop systems that can be ported to languages other than English. The interest in other languages is, for example, explicitly stated in the Multiple Language Question Answering Track at CLEF 2003 (Magnini et al. 2004), that includes Italian among others.

7One of the earlier attempts in this regard has been the DIOGENE model (Magnini et al. 2002; Tanev et al. 2004), which exploits linguistic templates and keyword recognition to answer questions over document collections. Other efforts have been made in the QALL-ME project (Cabrio et al. 2007, 2008; Ferrández et al. 2011), where a system was created for the tourism domain through an instance-based method, that is by clustering together similar question-answer pairs.

8More recently, the QuASIt model (Pipitone, Tirone, and Pirrone 2016), makes use of the Construction Grammar and an abstraction of cognitive processes to account for the inherent fluidity of language, while exploiting linguistic and domain knowledge (in the form of an ontology) to answer essay and multiple choice questions. Similarly, the authors of (Leoni, Torre, and Vercelli 2020) built a system to answer questions regarding a specific domain using IBM Watson services and online articles as source of information.

9These kind of systems, built to answer questions using textual information, have been largely growing in recent years, especially since the availability of large QA datasets such as the Stanford Question Answering Dataset (SQuAD)2, which allows to train complex deep learning models with millions of parameters (Rajpurkar et al. 2016; Rajpurkar, Jia, and Liang 2018). While the performance shown by these models is impressive, they suffer from major drawbacks: first of all, they need an extremely large dataset to be trained on, making the porting of such a system to another language extremely demanding;3 furthermore, they show a lack of controllability in the sense that it is unclear which new examples are to be added to make a new question answerable. This makes systems opaque and difficult to maintain.

10The MULIB system (Siciliani et al. 2019) tackles the problem of answering questions in Italian over structured data. The system is based on a modified version of the automaton developed for CANaLI (Mazzeo and Zaniolo 2016), but it employs a Word2Vec model (Mikolov et al. 2013) to allow for more flexibility in language use. In contrast to these trained approaches, we present a model that generates (i) a deeper interconnection of semantic and syntactic information through the integration of a lemon lexicon with the DBpedia ontology, and (ii) the focus on Linked Open Data as a source of knowledge.

3. Methodology

11The architecture consists of two components: (i) the grammar generator and (ii) the QA component. The approach to grammar generation for different syntactic frames according to LexInfo (Cimiano et al. 2011) for the English language was described in a previous work (Benz et al. 2020). In this paper we show that, through a simple language adaptation, we are able to adjust the system so that the system also accepts questions in Italian language.

12In a nutshell, the grammar generation approach relies on a mapping between syntactic constructions and classes and properties from a given ontology and/or knowledge graph. This generation process makes use of several frames, each describing the linguistic realizations of specific properties that might appear in questions. Thus, the frames employed in this work are: NounPPFrame, TransitiveFrame, IntransitivePPFrame, AdjectiveAttributive and AdjectiveGradable.

13For example, the (lexicalized) construction for the NounPPFrame ‘the capital of X’, can be regarded as expressing the DBpedia property dbo:capital, with Country as domain and City as range. This would lead to the generation of the following questions:

• What is the capital of X (Country)?

• Which city is the capital of X (Country)?

14Similar grammar generation rules exist for transitive constructions (TransitiveFrame) as well as constructions involving an intransitive verb with a prepositional complement (IntransitivePPFrame) as well as adjective constructions in attributive (AdjectiveAttributive) and predicate form (AdjectiveGradable).

15In the context of this work, we adapted the generation of rules to the Italian language, without extending or modifying the existing types of constructions4.

16In adapting the grammar generation to Italian, we had to accommodate for the following language-specific properties:

• Sentence order, e.g., in sentence starting with interrogative pronouns the subject has to be placed at the end of the sentence, e.g., Dove si trova Vienna? (Where is Vienna?)

• The presence of auxiliary verbs, either avere (have) or essere (be), in compound tenses;

• Interrogative pronoun rules, e.g., chi (who) is invariable and refers only to people;

• The use of interrogative adjectives, e.g., quale (which);

• The use of different prepositions, either simple or articulated, on the basis of range/domain semantics (e.g., toponyms might require different prepositions);

• The presence of a determiner/articulated preposition on the basis of range/domain semantics (e.g., toponyms are preceded by a determiner when the noun refers to a country).

17Consider the lemon lexical entry in Figure 15 for the relational noun ‘capitale della’. The entry states that the canonical written form of the entry is “capitale”. It states that the entry has a NounPPFrame as syntactic behaviour, that is it corresponds to a copulative construction X è la capitale della Y with two arguments, where copulativeArg corresponds to the copula subject X and the prepositional adjunct corresponds to the prepositional object Y.

Figure 1

18We give examples for the different syntactic frames below to illustrate the behaviour of the Italian grammar generation.

NounPPFrame

19Assuming that in the corresponding lemon lexicon we model the connection between the NounPP construction capitale della (capital of) as referring to the property dbo:capital with domain Country and range City, we can generate questions automatically such as:

1. Qual è la capitale della (What is the capital of) (X|Country_NP)?

2. Quale città è la capitale della (Which city is the capital of) (X|Country_NP)?

where X is a placeholder allowing to fill in a particular country, e.g. Germania (Germany), or a noun phrase, e.g., paese dove si parla tedesco (the country where German is spoken).

TransitiveFrame

20Assuming that the lemon lexicon captures the meaning of the construction X ‘scrive’ (write) Y as referring to the property dbp:author, with Song as domain and Person as range, the following questions would then be covered by an automatically generated grammar:

1. Chi ha scritto (Who wrote) (X|Song_NP)?

2. Quale cantante ha scritto (Which singer wrote) (X|Song_NP)?

3. Quale (Which) (X|Song_NP) è stata scritta da (was written by) (Y|Person_NP)?

IntransitivePPFrame

21Assuming that the lemon lexicon captures the meaning of the construction ‘X pubblicare nel Y’ (‘X published in Y’) as representation of the property dbp:published, with Song as its domain and Date as its range, the following questions would be generated:

1. Quando è stata pubblicata (X|Song_NP)? (When was (X|Song_NP) published?),

2. Quale (X|Song_NP) è stata pubblicata nel (Y|date)? (Which (X|Song_NP) was published in (Y|date)?

3. In quale data è stata pubblicata (In which date was) (X|Song_NP)?

AdjectiveAttributive and AdjectiveGradable

22Assuming that the lemon lexicon would capture the meaning of the (gradable) adjective lungo (long) as referring to the ontological property dpb:length, the grammar generation approach would generate the following types of questions:

1. Quanto è lungo il (How long is the) (X|River_NP)?

2. Qual è il fiume più lungo (del mondo, del Kentucky)? (What is the longest river in (the world, Kentucky)?).

23The rules implemented for the generation of Italian questions are shown in further detail in Table 1. In particular, we use the tagset6 from the Penn Treebank Project (Marcus, Marcinkiewicz, and Santorini 1993), with V* defining all possible forms of a given verb, words in brackets defining nouns/verbs/adjectives that realize a specific property, and dbo:range/dbo:domain defining the possible labels that may represent classes (e.g., dbo:Country might be represented by either paese or stato).

4. Results

24We apply our system to the DBpedia dataset and manually created a lemon lexicon comprising of 249 lexical entries7. Table 2 shows the number of grammar rules and questions generated for each syntactic type. Altogether, the approach generates 620 grammar rules and about 1.6 million questions. The web-based demonstration is available online8.

25We used the training set of multilingual QALD-79 to evaluate our approach. QALD-7 contains a total of 214 questions over linked data, covering for more relations than the ones we considered so far. In order to overcome this issue, a total of 109 entries were added to our system (22 NounPPFrame, 41 TransitiveFrame, 41 IntransitiveFrame, 1 AdjectiveAttributiveFrame and 4 AdjectiveGradable).

26The results of the evaluation process (Table 3) show a quite satisfying precision, but a low recall. The main reason behind such results is related to the presence of different types of questions in QALD. Indeed, besides single-triple questions, QALD presents also complex questions referring to more than one triple, e.g., A quale movimento artistico apparteneva il pittore de I tre ballerini? (What was the artistic movement of the author of The Three Dancers?), which are not covered yet by our model. Nevertheless, when taking into account all the questions in QALD-7, our system recognizes 46.98% (101 questions) of the total set of questions.

5. Conclusion and Future Work

27We presented an approach to developing Italian QA systems over linked data that relies on the automatic generation of grammars from corresponding lemon lexica describing how elements of the dataset are realized in natural language. The approach is controllable, since the introduction of a lexical entry increases the question coverage in a fully predictable way. Our proof-of-concept implementation over DBpedia covers 1.6 million questions generated from 249 lemon entries.

28In future work, we intend to further automatize grammar generation by using LexExMachina (Ell, Elahi, and Cimiano 2021), which induces lexicon entries bridging the gap between ontology and natural language from a corpus in an unsupervised manner.