1 Introduction

Advertising aggregation into web pages is a common strategy for monetization. Too many web sites deliver free content and get income using advertising payment. This kind of monetization allows them to operate and to keep information production. Sometimes advertising causes displeasure for diverse reasons, for instance, when the amount of publicity is excessive and others when the commercial message is discordant for the information read.

Advertising aggregation is an important issue that must be analyzed from different points of view. On one side, commercial information is necessary for guaranteeing the economic survival of web sites. On the other hand, users read information about their interest, i.e., they browse and consume web content according to a specific motivation. We consider that methods that preserve the thematic sense between web content and an advertising message can reduce the reader’s displeasure when publicity is shown.

In this work, we consider the thematic sense as a set of topics which are correlated. For instance, a football match, which is transmitted by TV, regularly is related to the consuming of snacks and beer and, perhaps, within a friend’s meeting. They are different facts, a football match, snacks, and beer drinking and friends meeting. However, they are correlated. A formal approach that is useful to find out correlations into data is the technique of Latent Semantic Analysis (LSA [⁴]).

Here we can state the proposal for this work. We think that a gentle introduction to the publicity can be reached if the sense of the web content concerning the advertising message can be assured. For this reason, we introduce a formal method, to measure the level of correlation among both messages, expressed within text fragments.

Our proposal is motivated, too, by the behavior of new styles of inRead advertising, which dynamically presents commercial information when a user browses a web page. In this way, web page fragments could be semantically compared with the advertising message, and then, the more similar web page fragment could be determined for the publicity location.

Now an important issue must be faced, which web page fragments should be considered for the analysis?

We could treat the text source as plain documents. However, web text is formatted using (X)HTML, i.e., the information is organized in formatting nodes which frequently provide certain implicit unity (all the information in a node is related) imposed by the web designer.

In this way, the main path for this work is going to be established in the context of approaches that exploit the text fragments within a web page and written in natural language. For instance„ those devoted to web filtering such as [¹, ¹⁰, ¹⁷]. Regardless of this focus, formulas here presented for calculus of semantic similarity among text fragments can be applied in a seamless way to any pair of text fragments, and thereby we could analyze sentences of paragraphs instead of web page text fragments.

We consider that methods preserving the unity of (X)HTML nodes are more acceptable. Some works following this approach are [¹, ⁵, ¹⁴, ¹⁶]. For instance, in [¹], a method for information extraction from web pages considering the distance between (X)HTML nodes as measurement of analysis is introduced. However, standard tests of similarity, in the setting of natural language processing, as a basis for producing text measurements, are not employed. We believe that it is necessary to test classical techniques of natural language processing to establish an adequate comparison framework.

In this work, we present a formal method for automatic measuring of semantic similarity among an advertising text and a web page node, based on techniques of similarity employed in natural language processing and inspired on the notion of latent semantic analysis. Our formal method requires only once the calculations of LSA, and then it returns the more relevant web page node for advertising placement.

One of the main contributions of this paper is the definition of one formula to determine the semantic similarity of one text excerpt w.r.t. an advertising text. The formula is not affected by the size of text fragments.

The rest of the paper is organized as follows. In Section 2, Theoretical foundations of our work. In Section 3, we introduce a formal technique for relevance calculation based on semantic similarity and tree-like structure of web pages. Then, in Section 5, we describe a prototype and a set of experiments. Finally, we present related work and conclusions in Section 6.

2 Theoretical Foundations

In this section standard theoretical basis of our work are introduced.

Salton et al. originally introduced the vector space model for automatic indexing in [¹⁵] and it is considered a standard representation technique in information retrieval setting where stored entities (documents) are compared with each other. Given a text document d, a dictionary of terms is a set whose elements are the different words in the document d. V→(d) denotes the vector associated to document d, whose components are the weights for each element in the dictionary.

In a vectorial representation, typographic symbols such as "," or "-" are ignored. The well-known stop words are treated in the same way. For web pages, formatting labels are removed.

3 Calculating the More Relevant Web Page Node for Advertising

Now, we describe the main contribution of this work, i.e., a formal technique for calculating the more relevant web page node for advertising.

The goal of the technique is to find out the web page node (of a specific URL) that presents the best semantic similarity w.r.t. a given advertising text. An advertising text is a fragment of text in natural language, which is regularly inserted in a web page to be exposed.

4 Extraction of Text Fragments from a Web Page

To define a strategy to extract text fragments from a web page, we could suggest many criteria. However, there is the standard scheme DOM, which organizes the content of a web page in hierarchical nodes.

Example 8 Let us consider the following web page:

<html>

   <head>

      <title>Semantics within web pages</title>

   </head>

   <body>

      Semantics is related with meaning.

      <div>

         Semantics can be extracted in two ways:

         <span>

            a) Driven by metadata

         </span>

         <span>

            b) By using mathematical techniques.

               One of them is:

               <strong>

                  latent semantic analysis.

               </strong>

         </span>

      </div>

   </body>

</html>

DOM model is an adequate scheme. It stands for Document Object Model, which is a W3C standard platform—and language-neutral interface that allows programs and scripts to dynamically access and update the content, structure, and style of documents [⁹].

Its corresponding DOM representation is visualized in a graphical mode in Figure 1. We can see a web page as a tree-like data structure where each node is an (X)HTML element, i.e., a (X)HTML tag with its contained text and its attributes, furthermore, its children are the embedded (X)HTML labels.

Fig. 1 The DOM tree of Example 8, visualized in a graphical way 

In a DOM data structure, if a node contains nested (X)HTML labels, there is a relation of embedding between the node and its children. For instance in Example 8, div node embeds two span node children.

Definition 9 (web page tree) A web page tree (V,ℰ) is a directed, acyclic graph whose vertices V are (X)HTML elements and ℰ⊆V×V is a set of ordered pairs (v→v′)∈ℰ called edges, where v embeds v′, and there is a root(V,ℰ) node which is not embedded.

Given a web page tree P=(V,ℰ) and two nodes v1,vn∈V, if there is a sequence v1,v2,…,vn of nodes in P where (vi,vi+1)∈ℰ for 1≤i≤n−1, then we say that there is a path from v1 to vn in P . Given u,v∈V we say that the node v is reachable from u if there is a path from u to v. We use →∗ to denote the reflexive and transitive closure of →.

Now we define candidate branch of a web page tree in order to compute semantic similarity tests w.r.t. the web user query.

Definition 10 (candidate branch) Let P=(V,ℰ) be a web page tree, then B=(V′,ℰ′) is a candidate branch where (V,ℰ)V′={v∈V|v≠root(P)} ∪ {w|w is reachable from v} andℰ′= {(u1,u2)|(u1,u2)∈ℰ, and u1,u2∈V′}

Hence, each node (except the root) of a web page tree and its children is a candidate branch.

Definition 11 (web text) Let B=(V,ℰ) be a candidate branch, then text(B) is its web text.

Web text is necessary to get text excerpts for applying semantic similarity calculations.

Example 12 Let us consider the web page tree of Example 8, then there are 7 candidate branches, i.e., those which are compound by the following (X)HTML elements:

B1={head, title}

B2={title}

B3={body, div, span1, span2, strong}

B4={div, span1, span2, strong}

B5={span1}

B6{span2, strong}

B7={strong}

Their contained texts were not drawn. Hence, the set of candidate branches is

{B1,B2,B3,B4,B5,B6,B7}, and with respect to the function text, for instance, text(B₆) returns the text excerpt "By using mathematical techniques. One of them is: latent semantic analysis." and text(B₇) returns the string: "latent semantic analysis.", i.e., we extract the text without (X)HTML labels.

Definition 13 (semantically relevant nodes)

Given L= {<B,s>|B a candidate branch, A an advertising text and s=semsim(text(B),A)} be the ordered set of semantically relevant nodes induced by the total order ≽ where <Bx,sx>≽<By,sy> if sx≥sy.

Example 14 Here the shape of the semantically relevant nodes is basically an ordered list composed by candidate branches and their corresponding semantic similarity w.r.t. an ad text, for instance: L={<{span2,…},0.53>,<{div,…},0.5>,<{body,…},0.42>,<{span1},0.28>,<{strong},0.28>,<{head,…},0.0>,…}, would show the branches (without text drawing) and their semantic similarity.

Once semantically relevant nodes of a web page were computed, we are able to offer the most significant text excerpts as a result of an ad text.

Next, we present a definition of relevant web page node for advertising placement. For this reason, we require some auxiliary functions: highest(L) returns the web page node with greater semantic similarity found in a set of semantically relevant nodes. In the above example, the function would return <{span2,…},0.53>. Formally,

Definition 15 (relevant web page node) Given a web page url and an advertising text A, let node=highest(L) be the more relevant web page node for advertising placement, where L are the semantically relevant nodes of url with respect to A.

In Figure 2 we present an algorithm to determine the more relevant web page node for advertising location from a given web page.

Fig. 2 An algorithm for computing the more semantically related web page node for advertising location 

The first step of the algorithm prepares a collection from contextual documents. Then, it constructs the matrix M by indexing text from the collection. After, it computes singular value decomposition, reconstructs M′, and calculates the mutual similarity matrix ℳ, once this was done, it has all elements for computing semantic similarity between nodes and advertising text. There, we introduce a function: semantically_relevant_nodes(w,A) which computes the list of semantically relevant nodes from a particular web page and by considering A as text advertising.

The last part of the algorithm extracts the highest semantically related web page node w.r.t. the advertising text and returns it. In other words, the algorithm determines which web page node has more semantic similarity w.r.t. an advertising text.

5 Prototype and Experiments

In this section, some aspects of the practical approach for web page node relevance calculation are presented, a relevant node is the more semantically related w.r.t. a text advertising. The computational tool is shown in Figure 3.

Fig. 3 A tool for web page node relevance calculation 

Once the target web page is defined, the web page is downloaded, for this, a parser is employed. Jericho [⁶] allows us to retrieve and to filter web pages and their text. Inclusive, it is possible to walk through their tree structures, visiting each node, and recovering their text. DOM node tree is visited to take one node, each time, for analysis. Then, DOM node and advertising text are augmented, and finally, a value of semantic similarity is computed. The node that presents the highest semantic similarity is chosen as the more appropriate place for particular commercial information.

Figure 4 is devoted to the natural language processing phase explanation. The first step refers to text preparation. Naturally, the tool requires a set of documents whose content must be in the same context as text advertising. It is crucial because LSA will calculate correlations between terms, which is useful for measuring semantically related texts, and to be exploited in our relevance analysis.

Fig. 4 Semantic NLP process 

Then, the standard actions of natural language processing are shown: text cleaning, stop words deleting, and so on. For that phase, a proper library has been developed. Each document is treated, next, matrix M is formed by indexing text from the collection of documents. SVD calculus is performed by using a library of the third part, M′ is reconstructed, and the mutual similarity matrix is computed. Roughly speaking, this process performs latent semantic analysis.

6 Related Work and Conclusions

To the best of our knowledge, there are many research works focused on measuring the success of advertising campaigns [², ³]. However, there are not for measuring the best place for publicity location on a web page. It may be that this concern is more attractive for companies and, the solutions they reach are implemented instead of being published. However, from an academic point of view, it is interesting to measure the meaning of advertising w.r.t. the meaning of the web content which a user reads. In this work, we proposed a method for the numerical comparison. Other future work will be devoted to analyzing the success of this criterion for publicity location.

In summary, we have introduced a formal technique for semantic similarity between web page nodes and advertising messages, which presents several advantages: always returns a value, is independent of the size of text fragments, privileges (numerically) the existence of common words in text node and ad text, outperforms results of cosine similarity, only once calculation of LSA is required to produce any number of comparisons from an ad text.

The exposed formula of semantic similarity is based on that presented in [¹³], some formalizations are included here. And finally, the technique is described in a technological style (fully procedural), which is convenient to replicate the tool.

We have developed a pair of tools, one for web page (Jericho [⁶] based) parsing and other for text fragment ranking (Figure 5).

Fig. 5 A prototype for automatic relatedness calculation among advertising text and web page nodes 

For future work, we are testing deep methods based on word embedding, to produce a parametric framework for web filtering and web indexing.

The potential applications of the technique are the following: analysis of similarity among content and advertising, filtering of web pages (since we rank DOM nodes with semantic similarity), the transformation of web pages, determining of hot sections in a web page, production of industrial tools, and other more.