Cognitively Motivated Query Abstraction Model Based on Associative Root-Pattern Networks

1 Introduction

Humans tend to rely on the multiplicity provided by natural languages in expressing their request for information. A search topic or query can be formulated in different morphological forms and terms. Furthermore, humans are disposed to use a limited number of keywords for requesting information. For example, according to Klink et al. [20], about two to three words are the average term words used in a search query.

However, incomplete and ambiguous query words pervaded with uncertainty for requesting information form a potential source for misinterpreting the conveyed information in a query, particularly in case of Arabic and Arabic script-based computer systems. This problematic issue of natural languages has generally been investigated in cognitive infocommunication science [3], cognitive informatics and information retrieval (IR). Retrieving the most relevant documents within a large collection satisfying the intended human inquiry expressed in a small number of words is still a demanding task in IR and cognitive linguistics. Furthermore, using natural language understanding systems, as tools for inter-cognitive infocommunications between humans and machines, which are capable of simulating query cognition; represents an important step toward supporting the co-evolution process between human and machines.

In this framework, as Arabic is known for its hugely inflectional morphology, and its propensity to polysemy, on the stem, root, particle and pattern levels, the human-computer interaction can be harmfully affected. Incomplete and noisy query terms possibly spread out with real word spelling errors, complicate the inter-cognitive communication by recalling irrelevant and possibly hindering relevant information of becoming visible, depending on the employed indexing strategy, i.e. root-, stem- or lemma-based. Such types of miscommunication can be considered as a type of misunderstanding in an inter-cognitive communication between human and machines.

The following examples might illustrate the particularity of these issues in the case of Arabic:

A major source for the occurrence of these difficulties might be explained by following observations.

Non-linearity of the morphology. Arabic and Semitic languages are dominantly non-linear (i.e. non-concatenative), whereas most indexing, search and query expansion techniques are based on a concatenative word structure such as Latin-based languages. Resolving this aspect within a search process might require considering the following:

1. Different central elements of Arabic morphology such as roots, stems and lemmas. We propose even to consider patterns as part of the indexing process.

2. Morpho-based contextual semantic word forms.

Root-pattern particularity in word perception. The current research concerned with cognitive word identification and recognition for Semitic languages, and in particular for Arabic and Hebrew, is increasingly providing evidence for the tendency of Arabic of being root-morpho-phonetic pattern based during visual and textual word processing and recognition [4, 7, 12]. Employing this principle in exploring a query potential provides us with a cognitively motivated approach for query expansion and resolving possible ambiguity working on a higher level of abstraction.

In this study, we propose a novel approach considering the above-mentioned observations relying on associative bi-directional relations introduced in the APRoPAT (Associative Probabilistic Root PATtern) model in the framework of query abstraction and construction [12]. This model can be regarded as one of the application potentials of the APRoPAT model in the context of human-centred interaction. Furthermore, as this model relies on semantic network representation, it is capable of supporting the core computational problems such as predication, disambiguation and gist extraction [25]. However, it differs from the classical associative word-network in the following sense:

1. This model dominantly operates a higher level of abstraction than simple word-associative networks. In this context, roots are not real-world words; they are rather higher semantic abstractions of basic meanings of possible Arabic words and they are unalterable. Whereas, patterns are variables representing phonetic templates for some possible word forms and do not exist as real words. We can consider them as some kind of template of acoustic patterns representing the tones of possible words, which need roots to be recognised [4, 12, 16].

2. The associative relationship between the abstract concept is always bi-directional to enable backward and forward chains.

The principal idea of the proposed model is based on proposing the most associative root-pattern subnetwork from query terms. Roots, stems, particles and patterns are defined as abstract cognitive variables on the morpho-phonetic level of word cognition. Based on extraction of certain associative relationships within these cognitive variables, a query morpho-phonetic network expressed in terms of degrees of association strength-function is established. Moreover, the notion of query morpho-phonetic vector is used to represent the most associative query vector within the query vector space. The initial motivation relies on the characteristic of human language cognition, particularly in Semitic languages; that is, phonetic patterns are involved in the mental word perception and identification, which can be expressed in terms of the morpho-phonetic characteristic of Arabic in the context of query perception.

2 Proposed Processing Model

As stated earlier, the proposed model is cognitively motivated and human centred, which proceeds from the view that a query, textual, contextual and conceptual knowledge can be extracted from an associative bi-directional semantic network established between basic cognitive variables on the morpho-phonetic and semantic levels of perception. Furthermore, a query can consequently be represented as a vector of keywords going together with a certain phonetic search information. Operating on roots and patterns leverage associative relationships on a higher level of abstraction. An associative root-root relationship implies multiple word-word co-occurrences (see Figure 1). Subsequently, a query represented by basic cognitive variable such as ROOT, STEM, PARTICLE and PATTERN can be expanded into a larger associative bi-directional network representing a local knowledge of the involved query. Roots and pattern are decisive in extracting textual and contextual query terms. Extracted words can be classified as textual (i.e. based on the root literal) and contextual (i.e. based on associative root-pattern relationships).

Figure 1:

An Example of Associative Root-Pattern (Morpho-Phonetic) Network Containing Bi-directional Associative Relationship between Roots and Pattern.

The pattern (/mustaC₁C₂C₃/) might sound phonetically like (/mustas˘fā/, Hospital); however, it is cognitively not perceivable as long it has been not applied to the root (/s˘fw/ healing).

The process of anticipating the intended users’ requests can be understood as a semantic process directed toward building the most associative similar query root-pattern subnetwork. This process implicates instantiating the most reasonable query vector within the morpho-phonetic space of a query. Figure 2 shows the architecture of the involved processes:

1. Morphological data processing;

2. Query subnetwork construction;

3. Morpho-phonetic query vector construction;

4. Query instantiation;

5. Query candidate generation.

Figure 2:

Process of Construction of an Abstract Morpho-Phonetic Query Vector Based on Associative Cognitive Variables Relationships and Their Instantiation.

Some major benefits of this model can be summarised as follows:

1. Pattern polysemy and root homonymy can also be decreased by considering the most associative relative query root-pattern vector.

2. New contextual and textual associative word forms can also be extracted.

Next, we will focus attention^[2] on the most relevant relations in the context of local query root-pattern network construction.

Based on Definition 2, a global associative network on the morpho-phonetic level can be viewed as a network of all bi-directional associative relationships between all involved morpho-phonetic cognitive variables ∈ 𝔐. Furthermore, the strength of association between these binary relations can also be expressed in terms of binary fuzzy relations, whereas the degree of the strength of association is estimated through a corpus considering bi-directional prior probabilities.

To generalise and abstract the meaning of association, we will merely use the CODAS(𝒳, 𝒴) to annotate some function determining the grade of association between cognitive variables. Moreover, ∀ x_i and y_j ∈ 𝒳 and 𝒴, respectively, we use codas(x_i, y_j) to formalise the instantiated (concrete) degree of strength of association between some instances x_i and y_j in the global network. In the proposed model, APRoPAT-based associative relations were adopted as an estimation for CODAS(𝒳, 𝒴).

2.1 Cognitive Variable-Headed Associative Network

As indicated earlier, each instance of a cognitive variable is interconnected by multiple bi-directional associative relationships. However, considering a query as a network of related morpho-phonetic forms allows us to follow the potential semantics of a query by determining the most associative textual and contextual subnetwork for each word form of the query. Based on a global associative network, a local associative subnetwork for each reduced query word can be constructed. Subsequently, query morpho-phonetic vectors can be followed by determining different query strong associative morpho-phonetic paths, where each vector can be considered as a query expansion model for the involved query words.

At this level, a query subnetwork represents an abstract morpho-phonetic model for cognition, where roots and their possible pattern variables are processed bi-directionally to determine concrete word candidates required to perceive a word.

Next, we will focus on definitions of cognitive variable-headed instances as subnetworks, followed by the meaning of the concept of abstract morpho-phonetic vectors and some characteristics of the model.

Definition 3

Definition 3 (cognitive variable-headed associative network).

Let x be instance for some cognitive variable and Y be any cognitive variable ∈ 𝔐, then the ⟨ x⟩-headed associative subnetwork is defined as the sub-graph associated with the x-Y relation for “ x” and ∀ y_i ∈ Y and ∀ codas(x, y_i), where

(9) x−Y≜{((x,yi),codas(x,yi))∣(x,yi)∈X×Y}

The significance of Definition 3 comes to effect when considering a query involving terms in their reduced form, such as roots, patterns, stems or particles. Focusing the search process on certain central query term subnetworks directs the constructing process towards proposing a query local associative subnetwork, where certain plausible query vectors can be predicted as candidates for query expansion or more precisely for completing missing contextual or textual terms in the original query.

Definition 4

Definition 4 (query associative root-pattern-local network).

Let Q be a query, Q={W1,..,Wn}, containing the words W1,…,Wn with the corresponding reduced cognitive instances, w1,w2,..,wn. The related query network is constructed by establishing the most associative ⟨ w_i⟩-headed subnetworks of the involved words.

Figure 3:

A Local Associative Subnetwork for a Query Consisting of One Word Query Term, i.e. Q = ⟨ W⟩.

The query is headed by the instance pattern ⟨Pt_W⟩ and the root instance ⟨R_W⟩. The instances ⟨R₁⟩ and ⟨Pt₃⟩ represent a possible contextual one-term query candidate for the word W. The degree of association strength is depicted by intervals for simplification.

Figure 4:

Local Associative Subnetworks of a Query Consisting of Three Keywords: Q = ⟨W₁, W₂, W₃⟩.

In Figure 3, the instances R₁ and Pt₃ represent an abstract textual and contextual associative expansion of the original query root, i.e. R_w. The cognitive degree of association strength is depicted by the CODAS(X,Y) function on the morpho-phonetic level of abstraction, e.g. codas(R_w, Pt₃) ∈ [0.75, 1]. Applying Pt₃ to root R_w in the form of ⟨⟨ Pt₃⟨R_w⟩⟩ would generate a new word form with the same basic semantic of the query root R_w. Figure 4 represents three local associative subnetworks of a query consisting of three words, i.e. Q = ⟨ W₁, W₂, W₃⟩, whereas Figure 5 shows the resulting associative subnetwork of the same query.

Figure 5:

A Joint Local Associative Subnetwork for a Query Consisting of Three Words: Q = ⟨W₁, W₂, W₃⟩.

3 Model Evaluation

Evaluating a generic model based on subnetwork abstraction and construction is difficult to standardise and formalise. In this model, the main reason for this difficulty relies on the following facts:

1. The major outcomes of this model are query subnetworks, i.e. weighted graphs and templates for abstract phonetic vectors, which are difficult to evaluate with traditional evaluation methods.

2. The associative values between concepts are based on fuzzy binary relations, which need fuzzy assessments.

3. A major goal of this model is to stress on the cognitive aspect of this model and its closeness to human behaviour in the context of topic expansion. Measuring the performance in terms of IR metrics is out scope of this presentation at this stage of research.

To consider these aspects, this approach proposes employing the fuzzy subsethood theorem [21] as a metric to highlight the major aspects of the model:

1. Representing constructed query associative word networks (M) as fuzzy subsets expressing the predicted concepts in relationship to the involved test queries, i.e. the outcome.

2. Human-based associative word networks (H) represented as a fuzzy subset acting as standard and a reference for the assessment, i.e. the test gold standard.

3. Bi-directional fuzzy subsethood degrees between M and H [15], i.e. F-S(H,M) and F-S(M,H) as evaluation metric assessing the similarity between human model and machine model:

   (11) F−S(H,M)≜Grade(M⊂H),where

   (12) F−S(H,M)≜∑i=1nMin{μM(xi),μH(xi)}∑i=1nμH(xi)

and analog

In this context, F-S(H, M) is assessing the grade of the overall fuzzy-based accuracy of the model, where F-S(M,H) is assessing the grade of the subsethood of the human-based associative model in the proposed query model. Here, it is worthwhile to mention that F-S(H, M) is similar to the concept proposed by Binaghi et al. [5] as a fuzzy set-based accuracy assessment of soft classification.

4 Outlook and Conclusion

This paper attempted to introduce an elementary study towards establishing a novel approach concerned with some aspects related to formalising a cognitively motivated model for query abstraction. The proposed approach relies on establishing the most associative abstract phonetic templatic structures from local associative networks on the morpho-phonetic level of abstraction. The strength of association is expressed in terms of associative relations, which can be interpreted as binary fuzzy relations. To ensure the general and global character of this model, the degrees of the strength of the association were estimated based on a global corpus containing around 11.5 billion word forms and 9.3 million associative relationships. Some types of word polysemy and homonymy can also be reduced during constructing the corresponding most associative query subnetwork. A major goal of this presentation is to emphasise the morpho-phonetic and root-pattern dimension and its effects in understanding and predicting the intended information conveyed by some query patterns in a human language. This concept provides us with a novel approach to investigate a query and topics on morpho-phonetic level of abstraction. This aspect is important as this model is dealing with query understanding rather the internal indexing of an IR system. However, integrating such models within certain human-machine interaction such as IR and inter-cognitive communication systems would be an important improvement in query expansion modelling. Expanding and integrating an IR index of into APRoPAT semantic is a subject of our future work. For the evaluation of this approach, the fuzzy subsethood theorem has been proposed as an assessment metric reflecting the fuzzy accuracy and the closeness to human-created associative word-networks. These associative word-networks were created under considering the priming effect before reading query terms and after reading certain related documents and human-based evaluation.