1 Introduction

A forward dictionary contains mappings of a word to its meaning. It serves the purpose of looking up an unknown word while reading a text, thus addressing the needs specific to readers. The needs of language producers (speakers and writers) are, however, different. While composing a piece of text, the problem often faced is not to look up the meaning of a word but to get/recall an appropriate word corresponding to a meaningful phrase. Given its organization, a forward dictionary falls short in addressing this problem, thereby introducing the concept of a Reverse Dictionary (may be henceforth referred to as RD). A Reverse Dictionary maps a phrase to semantically appropriate words. Specifically, it addresses the Tip-of-Tongue (TOT) problem [3], meaning the word is on the tongue tip but the person is not able to articulate it. Several works exist in the literature addressing the problem of Reverse Dictionary (could be referred to in [24])and can be grouped into the following categories: Information Retrieval (IR) System based Approach [2, 5, 8, 22, 23], Graph Based Approach [7, 21, 27], Mental Dictionary based Approach [34, 35], Vector Space Model based Semantic Analysis Approach [4, 12], Neural Language Model based Approach [1, 9, 10, 14, 15, 18, 19, 29, 32] .

A few commercial RD applications also exist; namely, Onelook.com [17] and reversedictioanry.org [20], out of which the latter is the most popular and serves as the benchmark for assessing the quality of the research-based RD works. Amongst the Information Retrieval (IR) System-based RD works, Wordster RD [22] is reported to outperform Onelook.com. We have built a specific Reverse Dictionary on the basis of the architecture of the Wordster RD.

In this study, we have build a RD solution in the specific context of the coronavirus pandemic. As this outbreak was recognized as a global health emergency, many texts addressing it started flooding. A study reports that as many as 23,634 unique published articles were indexed on Web of Science and Scopus in just the initial phase of the pandemic.Footnote 1 The text included varied forms, the primary being the news updates. Others include articles on awareness, diagnostic symptoms, safety measures, medical treatments etc. With due course of time, as this pandemic affected almost all the functionalities across the globe, the form of the relevant text became far diverse including modified government regulations, education system guidelines, notices and circulars in various organizations, research articles etc. Correspondingly, the set of language producers for these texts encompassed a variety of people ranging from doctors, journalists, political leaders, researchers, office staff, medical staff, social workers, non-government organizations, patients, education system staff, and the general mass on social platforms.

Owing to its specificity, the glossary of the coronavirus pandemic is unfamiliar to the language producers. Hence, the problem of missing an apt technical term while writing on this subject is inevitable. This problem becomes exaggerated given the variety of language producers including the professionals as well as the non-professionals. Consequently, a Reverse Dictionary on this subject is a requisite, and we address the same through this paper. In the process, we have compiled a database of the COVID-19 glossary from various Internet sources.Footnote 2, Footnote 3, Footnote 4, Footnote 5, Footnote 6, Footnote 7 This could be accounted as another contribution of this paper, as it could assist the interested readers on this subject.

The work carried out in this study is remarkably different from the similar-appearing task of medical terminology mapping [11, 16, 28]. Firstly, the former deals with finding an apt term in semantic coherence with a natural language input description, whereas, the latter deals with the identification and extraction of medical terms in natural language text. Secondly, in terms of objective, the former assists the language producers in general whereas the latter assists the medical professionals in particular. Lastly, in terms of scope, the former is associated with medical (such as Severe Acute Respiratory Infection (SARI) and Acute Respiratory Stress Syndrome (ARDS))terms as well as non-medical terms (such as lockdown and Work From Home(WFH)) related to the COVID-19 pandemic whereas the latter is associated with medical terms only.

The paper is organized as follows. In Sect. 2, we brief the architecture of the COVID-19 Reverse Dictionary, outlining the key algorithms. In Sect. 3, we provide the details of the implementation and results. Finally, the paper is concluded.

2 COVID-19 reverse dictionary: architecture and working

The architecture of the COVID-19 RD is based on the Wordster RD [22] which in turn is based on the Information Retrieval System’s architecture. Given a forward dictionary consisting of pairs of word and its definition, the Wordster RD accepts a user input description in natural language, converts it into a Boolean expression query, extracts definitions based on a semantic similarity measure and then outputs the words corresponding to the definitions in the form of a ranked list. Accordingly, these constitute the primary steps in the COVID-19 RD implementation. The flow of the system is illustrated in Fig. 1. Along the same lines, a Reverse Dictionary with specific application to English idioms is reported in [25]. We employ the primary algorithms of the Wordster RD in our implementation. The modules of the COVID-19 RD is outlined as under:

Fig. 1
figure 1

A high level representation of COVID-19 Reverse Dictionary

Full size image
  1. 1.

    Building the Reverse Map Set: The Reverse Map Set (RMS) applies to the terms in the vocabulary of a given forward dictionary. Specifically, a forward dictionary consists of word-definitions pairs and the set of terms appearing in the word definitions constitute its vocabulary.

    For a term t, RMS(t) consists of the vocabulary words whose definition contains the term t. Mathematically,

    $$\begin{aligned} RMS(t) = \{ W_1, W_2, \ldots , W_n \} \; s.t. \; t \in def(W_i), i = 1,2, \ldots , n \end{aligned}$$
    (1)

    For example, consider the word numismatics meaning the study or collection of coins, banknotes, and medals. The term set of the definition after stemming and stop words removal consist of the terms study, collect, coin, banknote, medal. Accordingly, numismatics appears in the RMS of the these terms.

  2. 2.

    Query Generation: For a given user input description, a query in the form of a Boolean expression is generated. Initially, the query is formulated as an ANDed expression of the terms appearing in the user input description (after it is processed for basic natural language processing tasks).

    Let \(t_1, t_2, \ldots , t_n\) be the terms appearing in the processed user input description U. Correspondingly, the query Q is represented as,

    $$\begin{aligned} Q = t_1 \; \odot \; t_2 \cdots \; \odot \; t_n \end{aligned}$$
    (2)

    such that \(\odot\) correspond to the AND operator. In case the required number of potential candidates are not fetched, the query is expanded using lexical relatedness like synonymy. The lexically related terms are connected to the query terms via the OR operator. Let, the synonyms of the query term \(t_i\) be \(t_{i_{s1}}, t_{i_{s2}}, \ldots , t_{i_{sk}}\). The expanded query \(Q'\) takes the form as shown below:

    $$\begin{aligned} Q' = (t_{1_{s1}} \; \oplus \; t_{1_{s2}} \; \oplus \; \cdots \; \oplus \; t_{1_{sm_1}}) \; \odot \; \cdots (t_{n_{s1}} \; \oplus \; t_{n_{s2}} \; \oplus \; \cdots \; \oplus \; t_{n_{sm_n}}) \end{aligned}$$
    (3)

    such that \(\oplus\) correspond to the OR operation.

  3. 3.

    Candidate Selection: For a given Boolean expression query, the candidate selection module extracts candidate definitions for ranking. This is done by executing the set operations on the RMS of the query terms (refer Eq. 1).

    The set operation corresponding to the \(\odot\) and \(\oplus\) operators in the query expression is . (intersection operation) and \(+\) (denoted by union operation) respectively. Performing these operations on the Reverse Map Set of the query terms result in a set of those words whose definition contain all the query terms (either in original from or connected through lexical relatedness), thus qualify to be relevant to the user input.

    Correspondingly, the definition phrases of these vocabulary words form the set of potential candidates. Assume the following to be a sample query,

    $$\begin{aligned} Q = (t_1 \; \oplus \; t_2 \; \oplus \; t_3) \odot (t_4 \; \oplus \; t_5) \odot (t_6) \end{aligned}$$
    (4)

    Then, the potential candidates correspond to the set of following words:

    $$\begin{aligned}{}[RMS(t_1) + RMS(t_2) + RMS(t_3)] . [RMS(t_4) + RMS(t_5)] . [RMS(t_6)] \end{aligned}$$
    (5)
  4. 4.

    Ranking: The ranking module ranks the given set of potential candidates on the basis of the similarity score of the candidate definition with the query. The similarity score consists of following two components:

    • Term Similarity, \(t_s\): depicts similarity between each pair of terms of the definition (potential candidate) and the query. In the Wordster RD, the employed measure is Wu and Palmer simiarity [33] according to which the similarity \(t_s\) between terms a and b is calculated as under:

      $$\begin{aligned} t_s = \frac{2*D(LCA(a,b))}{(D(a)+E(b))} \end{aligned}$$
      (6)

      such that D(t) is depth of the term t in the WordNet [13] hierarchy and LCA(a, b) is the least common ancestor of the terms a and b.

    • Term Importance, \(t_i\): depicts the importance of a term in a phrase (definition/query). This is based upon the structure of the parse tree of the phrase. The term importance value for the term t appearing in the phrase p is calculated as follows:

$$\begin{aligned} t_i(t,p) = \frac{(D(Z(p))-D(t))}{D(Z(p))} \end{aligned}$$
(7)

such that Z(p) is the parse tree of the phrase p, D(a) gives the depth of the entity a.

The two components, \(t_s\) and \(t_i\) are aggregated as \(S_w\) for the term \(t_D\) of the definition, D, and the term \(t_Qb\) of the query, Q using the following equation:

$$\begin{aligned} S_w(t_D,D,t_Q,Q) = t_i(t_D,D) * t_i(t_Q,Q) * t_s(t_D,t_Q) \end{aligned}$$
(8)

where the terms a and b appear in the candidate definition D and the user input U respectively. The weighted similarity score \(S_w\) for each pair of terms is then used to calculate the overall similarity using [6]. This final score forms the basis of ranking the candidates in the order of decreasing values. The result set consists of the vocabulary words corresponding to the candidate definitions in the ranked list.

3 Implementation details and evaluation results

We have compiled a COVID-19 dictionary dataset consisting of 110 words and their definitions from relevant links on the Internet.Footnote 8, Footnote 9, Footnote 10, Footnote 11, Footnote 12, Footnote 13 While compiling, if more than one definitions for a word is available from multiple web links, it is taken into account. Specifically, of the total of 110 words, 37 words have multiple definitions.

To prepare the test set, we have randomly selected 30 words from the compiled data set. These words are given to a set of 4 potential users of varied backgrounds, medical as well as non-medical. Each user is asked to write a phrasal description for each test word. The word’s actual definition is provided to the user only if it is required by him/her for clarification. The collected descriptions are checked for quality by providing them to the user other than the one who wrote it. If the test word is not correctly guessed based on the test description, rewriting is done. In this way, 120 test descriptions are collected, 4 for each test word.

3.1 Comparison benchmarks

As reported in the previous RD works, we have considered the commercial Onelook.com Reverse Dictionary [17] for comparison. Also, we have taken into consideration the recently reported WantWords Online RD [19]. As mentioned (in the Introduction section), the vocabulary addressed by the COVID RD implementation encompasses terms related to the pandemic. Thus, it includes medical-related terms (such as Severe Acute Respiratory Infection (SARI) etc.) as well as non-medical related terms (such as lockdown etc.).

We concluded that while the considered benchmarks could generate related responses for non-medical test terms definitions, the generated responses for the medical-related test term definitions were highly unrelatable. This in view that these RDs are for general English vocabulary and do not cater to specific applications. The overall accuracy of the test set for both is too low to consider. Onelook.com nevertheless achieved higher accuracy than WantWords. In view of this, we carried out a manual-based evaluation of our implementation as reported in the following section.

3.2 Manual-based evaluation results

For the prepared test set, we have obtained an accuracy of 0.48 based on the top-3 responses generated by the system. Given the number of participants, the reported accuracy value is the average over four runs. This implies that for about half of test user descriptions, the implementation can generate the sought word within the top-3 positions. Tables 1 and 2 lists a few samples of the successful and unsuccessful system responses, respectively.

Table 1 Sample successful system responses
Full size table
Table 2 Sample unsuccessful system responses
Full size table

4 Limitations and future directions

The current implementation takes into account the lexical semantics of the phrases: user input descriptions as well as the dictionary descriptions. The phrases are treated as a bag-of-keywords, hence, the implicit semantics is disregarded. For example, consider the test description at S. No. 5 of Table 2. The terms ‘but’ and ‘not’ in the test description “Person tested positive by private hospital but not by government” conveys the specific intent of the user. If the description is treated as a bunch of keywords, such user intents could not be addressed.

To address this gap, we need to base our RD solution on a paradigm capable of dealing with implicit semantics in natural language. In view of this, we bring into consideration the paradigm of Computational Theory of Perceptions (CTP) [30] proposed by L.A. Zadeh in which the objects of computation are words rather than numbers. Based on this paradigm, we propose to incorporate the concept of Precisiated Natural Language (PNL) [31] in building a Reverse Dictionary solution. A specific instance of the same is reported in our study [26].

5 Conclusion

As the COVID-19 pandemic has affected almost all the functionality across the globe, text in varied forms are found to appear in significant amount. The diversity of text forms encompasses a variety of language producers. Unlike the general English language vocabulary, the COVID-19 glossary is different and unfamiliar to the language producers and hence the problem of finding an appropriate technical term during composition of text/speech is inescapable. In view of this, we implement a Reverse Dictionary on the subject of COVID-19 pandemic. As a part of the implementation, we compile a data set of COVID-19 glossary which provides assistance to the readers as well. The implementation carried out is based on the framework of Wordster Reverse Dictionary and an accuracy of 0.48 is reported based on top-3 responses generated by the implementation. The scope of improvement in the task of RD lies in incorporating paradigms capable of dealing with implicit semantics, particularly Zadeh’s Computational Theory of Perceptions (CTP).