English. This work describes the first experiments conducted with a computational lexicon of Italian in a context of query expansion for full-text search. An application, composed of a graphical user interface and backend services to access the lexicon and the database containing the corpus to be queried, was developed. The text was morphologically analysed to improve the precision of the search process. Some examples of queries are given to show the potential of a text search approach supported by a complex and stratified lexical resource.
Italiano. Il presente lavoro illustra i
primi esperimenti condotti con un
lessico computazionale dell’italiano in
un contesto di query expansion per la
ricerca full-text. È stata sviluppata una
applicazione composta da una
interfaccia grafica utente e un backend
di servizi che permette l’accesso sia al
lessico che al database contenente il
corpus da interrogare. Il testo è stato
analizzato morfologicamente al fine di
migliorare la precisione del processo di
ricerca. Alcuni esempi di query sono
forniti al fine di mostrare le potenzialità
di un approccio di ricerca sul testo
supportato da una risorsa lessicale
complessa e stratificata.
The need of techniques going beyond the mere
“search by keyword” in the querying of textual
resources dates back to the dawn of
computational linguistics. Seminal works in the
Copyright ©️ 2021 for this paper by its authors. Use
permitted under Creative Commons License
Attribution 4.0 International (CC BY 4.0).
60s on the development of the very first
question answering (QA) systems already
included linguistic resources as support
datasets. To bring some “old school” examples,
the “General Inquirer” QA system
The support of linguistic resources has
proved its potential in the field of information
retrieval (IR) too, as highlighted in many of Bill
Woods’ works, culminating in the introduction
of his conceptual indexing technique and the
conceptual taxonomy resource
The use of linguistic resources in QA of the
earliest period of computational linguistics can
be considered as the precursor of “query
expansion” (QE), the technique that Manning
and Raghavanat describe as the most used
“local method” in IR to tackle those situations
in which “the same concept may be referred to
using different words”
Though QE may be obtained in different ways (among which query reformulations based on query log mining) we are here interested in those applications that make use of lexical resources.
Most of the works, published from the 90s to
nowadays (proving that QE is still being
investigated), exploit WordNet
The research work here illustrated places itself in the context of full-text search carried out using a lexical resource-driven QE technique. However, the focus of this research, differently from that of the cited works, is not on the specific QE technique and the relative evaluation, but on the resource we chose to exploit, introduced in the next section, in place of WordNet and on the frontend and backend technologies implemented to query the text, as described in details in Section 3. The advantages derived from the adoption of a rich and highly structured computational lexicon will also be remarked through some query examples shown in Section 4. The developed application can be freely accessed and used to query the corpus1. 2
This work stems from the activities conducted by the Institute of Computational Linguistics of CNR (ILC-CNR) in the context of the Talmud Translation Project2. The need of providing a way to query the Italian translation of the Talmud3 on a linguistic basis was the initial spark that led to the idea of experimenting the use of a computational lexicon for Italian. As a matter of fact, this resource (described below) represents a “linguistic mine” which has never
2https://www.talmud.it/ 3The corpus here queried is limited to eight tractates of the babylonian Talmud: Rosh Hashanah, Berakhot, Ta'anit, Kiddushin, Chagigah, Beitza, Sukkah, and Megillah been exploited for tasks of full-text search or information retrieval. 2.1
“PAROLE-SIMPLE-CLIPS” (PSC) is a
computational lexicon of Italian, developed
from 1996 to 2003 by ILC-CNR
In these features lies the motivation of this work, since the available linguistic information may be combined in ways that go well beyond what resources such as WordNet allow to do in the context of text search support. Even considering semantics alone, the information in PSC is detailed with fine-grained features that are not described in WordNet’s network of synsets: PSC encodes the meaning of each lexical sense as an array of information, including “templates” (see below), semantic traits, semantic roles, and argumental structures.
In this work, we document the first steps in the use of PSC for QE. At this stage we used: i) the Morphological Units, classified according to their POS, which represent the lemmas of the computational lexicon; ii) the Phonological Units that represent the inflected forms of the lemmas; iii) the Semantic Units (SemUs), that describe the senses expressed by the words. Furthermore, we considered the following morphological and semantic information: i) morphological traits (e.g. gender, number); ii) relations between SemUs (at the moment limited to synonymy and hyponymy); iii) the association between SemUs and “templates”, representing sets of senses, labeled according to one of the types represented in the Simple Ontology (Lenci et. al., 2001). The other parts
it.ilc.cnr.it/repository/xmlui/handle/20.500.11752/I LC-88. of linguistic information will be the subject of future works, according to an incremental approach. 3
The whole search process involves a series of steps that can be summarized as follows (see Fig. 1 for a schematic functional architecture of the application): i) the user inserts a first set of data to formulate the desired query in the Graphical User Interface; ii) the interface requests, via Web API, the lexicon backend services which return the linguistic data matching the initial query; iii) the user completes the query taking into account the linguistic data and starts the search; iv) the interface executes the query expansion and requests, via Web API, the text backend services which collect, tag, and return the matching textual portions of the Talmud; v) the interface shows the results to the user.
First of all, to make the lexicon efficiently queryable, it needed to be transformed from relational data into linked data (Section 3.1). At
6We remark that the conversion of PSC Simple is not the focus of this work, but it was necessary for the same time, a list of services to query both PSC and the database storing the Italian translation of the Talmud needed to be developed in order to answer to the interface requests (Section 3.2). The interface itself was designed on the basis of the available linguistic information exposed from PSC and developed accordingly (Section 3.3). Finally, to improve the precision of the search process, the queried corpus was also POS-tagged (Section 3.4). 3.1
The first phase of our work was to consider the relational database of PSC as the data source for the generation of a first Linked Data (LD) conversion. Two main reasons led to the need for a conversion of PSC: i) to ease the reuse of the lexicon itself, in virtue of the intrinsic nature of LD, ii) the possibility of performing automated reasoning on data if appropriately modeled taking into account ontological principles, for example to compute inferred closures, infer new knowledge on the basis of class taxonomies, property hierarchies, and so on. Accordingly to the LD principles, we first had to look for existing vocabularies for the modeling of lexicons.
In the context of the Semantic Web, the de
facto standard for representing lexical
information is the lemon model (Cimiano et al.,
2016). Its core module, called OntoLex, allows
to represent grammatical, basic morphological
and semantic information by means of three
main classes: Lexical Entry, Form (lemma and
inflected forms), and Lexical Sense. Lemon
relies on external vocabularies to define
semantic relations between senses: in this
conversion we modelled PSC’s synonymy and
hyponymy with LexInfo ontology5. Currently,
the converted resource includes 72006 lexical
entries (48735 nouns, 6522 verbs, and 11830
adjectives), 469726 inflected forms, and 57130
senses. Explicit lexico-semantic relations
include 1803 meronyms, 4060 synonyms, and
44487 hyponyms. This initial conversion of
PSC as Linked Data was purely functional to
the linguistic querying of the Italian translation
of the Babylonian Talmud6. Therefore, it was
decided to convert a selected number of
linguistic data to be exploited for the process of
query expansion. At the time of writing this
performing linguistic searches experiments on the
Italian translation of the Talmud.
proposal, a complete conversion of PSC as
LOD (Linked Open Data) is in progress. This
complete conversion will also take full
advantage of the already available works on the
resource as documented in
Once the computational lexicon was converted,
the implementation of the querying system
continued with the creation of the backend
services needed to access both the lexicon and
the database storing the text to be queried.
Regarding the lexicon, a GraphDB7 repository,
containing all the converted data, was set up.
The access to the repository was implemented
with a set of REST services that can be invoked
from any web client8. The services have been
based on the already available backend of
LexO, a collaborative web tool for the creation
and editing of lemon lexical resources
The GUI (Fig. 2) set up to query the corpus was developed using Angular9, one of the most widespread frameworks for frontend Web development, which provides high levels of portability and scalability. In this first version of the search system, the interface was conceived as a sort of “hub” of the whole architecture: from the one side to interact with the user and from the other side to invoke the services exposed by GraphDB and the Talmud database. The interface is divided into two sections. In the left-hand column, the available tractates of the Talmud that can be queried are represented as a tree allowing the user to specify the search context at different levels of granularity. The right-hand section contains the search parameters, where the user can choose 7Ontotext GraphDB is a highly efficient and robust graph database with RDF/OWL and SPARQL support (https://graphdb.ontotext.com/documentation/free/f ree/graphdb-free.html) between three types of search using the available tabs: Keyword, Form/Lemma, or Semantic Traits.
The first one is the classic keyword-based search. The second type, via the Form/Lemma tab, allows to search for a specific word form or the set of inflected forms of a given lemma by specifying some morphological constraints. By entering a word in the text field, the GUI invokes the lexicon backend services to retrieve the lemmas corresponding to the indicated parameters and displays them with their different senses. Users can then proceed with the search or they can select one or more lemmas and apply to them morphological constraints by clicking on the three dots icon on their right. The selection of at least one of the senses enables the semantic extension search feature: a drop-down menu allows users to look for all the other senses in the lexicon appearing as hypernyms, hyponyms, or synonyms at a specified distance. The forms obtained with this extension are subject to the propagation of the morphological constraints applied to the lexical entry to which they are linked, whether explicit (entered from the interface) or implicit (in the case of a search by form). Finally, the “semantic traits” tab provides two template trees on which multiple selections are possible: the first click selects a template with all its descendants, the second deselects the descendants, and the third deselects the node itself. When the selection changes, the lexicon is queried to obtain the list of senses linked to the chosen templates. Users can then select the desired senses which will be used to retrieve the forms of the relative lemmas to be used in the QE.
All the entered data are used to compose the expanded query, which will be constituted by all the inflected forms provided by the lexicon and matching the indicated morphological constraints, semantic extension, or templates.
The results coming from the backend services accessing the Talmud database are then displayed in a table on the right-hand side, upon which a panel lists the forms retrieved from the lexicon and used for the QE. 8The source code of the REST services is available at https://github.com/andreabellandi/LexO-backend 9https://angular.io/ 3.4
For the purpose of reducing the lexical ambiguity in cases where a searched word could match with homographs, the corpus was automatically analyzed and annotated with morphological information.
In particular, we parsed all the sentences of
the eight tractates of the babylonian Talmud
with Stanford's Stanza tools
In this last section, we show a concrete application of the approach by introducing some query examples. Each query can also be tested by the reader by accessing the available application.
The first two examples show the search for words with specific morphological traits and the application of semantic extension. In these cases, the “Form/Lemma” type of search is selected. In the first example, the word “insegnamento” (teaching) is inserted as a lemma. The system finds it in the lexicon and 10https://universaldependencies.org/treebanks/it_isd t/index.html shows it as a noun with one single sense. The user then adds a morphological constraint by setting the “number” trait to “plural”. Finally, the user extends the search to direct hyponyms (distance = 1) and submits the query.
This is a simple case of propagation of the morphological traits through semantics. The lexicon contains the two following key information: i) the fact that the sense of “insegnamento” has three hyponyms: erudizione” (erudition), “istruzione” (instruction), and “catechesi” (catechesis); ii) all the inflected forms and the relative morphological traits of the searched word and its three hyponyms. On the basis of these data, the system composes the final query, which allows to search for all the plural forms of the four lemmas as nouns. As a result, 103 textual segments are retrieved, containing the words “insegnamenti” (97 matches) and “istruzioni” (6 matches) (Fig. 2).
The second example involves the verb “permettere” (to permit/allow), searched as a lemma, with morphological constraints on the finite mood (“indicative”, “subjunctive”, “imperative”, “conditional”). In addition, the user selects just one of the two available senses of the verb (the one with the definition “dare a qlcu la possibilità' di fare qlco” - to give sb the chance to do smth -) and then extends the search to its synonyms. In this case, the lexicon proposes two synonyms of the selected sense: the (single) senses of words “concedere” and “consentire”. The resulting expanded query retrieves from the database a total of 405 matches, containing 334 strings of “permettere” (for 131 available forms of the lexicon), 44 strings of “concedere” (for 45 available forms) and 27 strings of “concedere” (for 41 forms).
The last type of search is structured as a more explorative querying of the corpus. In the semantic traits tab, the user can choose one or more between noun/verb or adjectival templates (group of senses), to look for all words relative to a specific semantic field, such as objects, weather verbs, metalanguage, etc.
In this example, the user selects the template “Air animal”, which appears as a “leaf” of the sub-tree under the parent-node “Entity”. Once the template is chosen, the system retrieves from the lexicon all the relative senses and shows them in a window. It is then possible to select all the available 165 senses or just some of them. Finally, the user can run the search: the system composes the expanded query and retrieves 226 textual segments of the Talmud containing words (both as lemmas and inflected forms) with senses referring to the semantic field of “Air animal”: “uccello” (bird), “mosca” (fly), “cavallette” (grasshoppers), and so on.
Among future developments, a feature for a “grouped” selection of multiple templates will be added, that will allow to search for textual segments containing co-occurrences of words referring to the specified templates. To bring an example, the grouped selection of templates “Color” and “Earth animal” will retrieve segments containing multiword expressions such as “vacca rossa” (red cow), “gatta nera” (black she-cat), “oche bianche” (white gooses), etc. 5
As shown in this paper, the availability of a rich and structured linguistic resource (as the computational lexicon we have taken into account) seems to provide an edge over the standard query expansion techniques for fulltext search based on WordNet. Now that a very first portion of the resource has been made available (though with a preliminary conversion) and the web application has been implemented, the road is cleared for the next steps.
The first critical issue that will need to be faced involves the limitedness of the resource, covering most - but not all - the lemmas, forms, and senses of standard contemporary Italian and that lacks many domain-related terms or senses. To fill this gap the resource will have to be updated and enriched with more entries.
At the same time, as anticipated, a more indepth and rigorous conversion of PSC will have to be carried out, a process that will probably take a lot of time and research effort and that for the sake of this first experiment would have been premature and unnecessary. As soon as the whole conversion will be ready, the rest of the information encoded in the lexicon will be made available and integrated in the search process.
Though the benefits of the availability of a computational lexicon wrt WordNet (or a similar resource) may seem obvious in a context of QE for full-text search, an empirical evaluation would be desirable. However, the set up of a benchmark conceived for this purpose appears anything but easy, mainly due to the lack of comparable works or evaluation campaigns focussing on the role of linguistic resources as support.
In conclusion, we believe these first experiments carried out by querying the talmudic text appear promising, especially considering that only a small part of the lexicon has been used. In addition, the support in the disambiguation provided by the POS tagging of the text suggests that an hybridization of a resource-driven QE technique with a deeper stochastic annotation of the corpus to be queried may constitute an interesting experimental field to be investigated.
This work was conducted in the context of the TALMUD project and the scientific cooperation between S.c.a r.l. PTTB and ILCCNR.