Web User Interact Task Recognition Based on Conditional Random ...

Web User Interact Task Recognition Based on Conditional Random Fields Anis Elbahi(B) and Mohamed Nazih Omri Research Unit MARS, Computer Science Department, Faculty of Sciences of Monastir, Monastir, Tunisia [email protected], [email protected]

Abstract. Recognition activity of web users based on their navigational behavior during interaction process is an important topic of Human Computer Interaction. To improve the interaction process and interface usability, many studies have been performed for understanding how users interact with a web interface in order to perform a given activity. In this paper we apply the Conditional Random Fields approach for modeling human navigational behavior based on mouse movements to recognize web user tasks. Experimental results show the efficiency of the proposed model and confirm the superiority of Conditional Random Fields approach with respect to the Hidden Markov Models approach in human activity recognition. Keywords: Conditional random fields · Hidden markov models · User task recognition · Cursor behavior analysis · Human computer interaction · Pattern recognition · Machine learning

1

Introduction

Inferring the activity of web users based on their navigational behavior is an important topic of HCI which are extensively studied during last decade. For years, various techniques have been used in this field, such as eye movements tracking [1], mouse tracking [2], physiological and psychological tracking [3] and click-through analysis [4]. These techniques have proven good efficiency in user behavior understanding and interfaces usability evaluation. Understanding navigational behavior of users can help designers to improve interfaces usability, to provide assistance for users with disabilities and others applications such as e-learning. Obviously mouse pointing device is the most commonly used tool during interaction with computer interfaces. On the one hand, the activity of mouse cursor such as movements, clicks and scrolling can be easily captured and recorded during interaction process. On the other hand, analysis of cursor behavior can provide high quality clues of a spontaneous, precise, direct and unbiased trace of user behavior. Such trace can be considered as a good indicator of the user reasoning strategy during a web activity. Consequently many studies have been performed to understand users cognitive c Springer International Publishing Switzerland 2015  G. Azzopardi and N. Petkov (Eds.): CAIP 2015, Part I, LNCS 9256, pp. 740–751, 2015. DOI: 10.1007/978-3-319-23192-1 62

Web User Interact Task Recognition based on Conditional Random Fields

741

strategy based on their cursor navigational behavior. Many other researches [5], [6] have proposed different models based on possibility theory, on bayesian and semantic networks to recognize the goal of the users. In this paper, we used the Conditional Random Fields (CRF) approach [7] in order to recognize the tasks of web users, based on their navigational behavior using mouse movement data.

2

Navigational Behavior Analysis Based on Mouse Movement Tracking

During web session, users seek to perform various tasks such as logging, ordering a product and sending an email. For each task, users perform basic operations such as keyboard events, moving a cursor, selecting an option, clicking a link and pressing a button. Since early computers, the mouse has been the most commonly used device during human computer interaction process. Using a cursor pointing device during web activities, users draw their navigational behavior. For many years, the cursor activity tracking attracts much attention of researchers who are interested in user navigational behavior. Mouse movement tracking has been evaluated as an alternative to eye tracking for determining attention on the web page. Therefore, various studies have been achieved in this context such as the study of Chen et al. [8] who have found that mouse and eye movements are strongly related and that 75% of mouse saccades move to significant regions of the screen where eye gaze are moved. In the same study, it has been confirmed that mouse data can be used to infer the intent of user. Those findings have led to extend the studies of user behavior analysis using mouse tracking technique. Mouse movements are explored to provide insights into the intention behind a web search query [2] and click prediction . Zheng et al. [9] presented a new approach for user re-authentication using behavioral biometrics provided by mouse dynamics. In a similar way, [10] analyses the interaction of users with a computer system in order to identify users only by analyzing their interaction behavior mainly based on mouse events. Today, mouse movement tracking is a very effective technique, easy to use, freely available and that does not disturb user behavior during the interaction. Mainly based on mouse trajectory, we propose a new model using CRF approach to automatically recognize web user tasks.

3

A User Task as a Sequence of Fixed Areas of Interest

Web interfaces can be presented as a set of significant regions called interface Items Of Interest (IOI) or Areas Of Interest (AOI) which can be manually specified or automatically discovered [11]. Usually, users perform various tasks (activities) using pointing device. During a task, without being aware, users move the cursor across the web interfaces and fix various AOI. The following figure presents an example of a sequence describing fixed AOI during logging into Gmail account task.

742

A. Elbahi and M.N. Omri

Fig. 1. Example of user task defined as a sequence of fixed AOI.

To define a task, we rely on activity theory [12] and the work of Gotz et al. [13] who characterized user behavior at four levels based on the semantic richness of the activity. The following figure presents logging into Gmail account via Google interface task based on Gotz task description.

Fig. 2. Example of user task representation based on Gotz description.

As shown in figure 2, according to different levels the user task can be defined as a temporal sequence of subtasks or a succession of AOI or elementary events. In our study, we focus on mouse trajectory crossing different AOI during a web user task.

Web User Interact Task Recognition based on Conditional Random Fields

743

Thus, according to Gotz task representation, we define a task as a temporal sequence of fixed AOI during a period of time T. TSKi = { AOI1 , AOI2 , . . . , AOIT }. Despite this clear definition of tasks, their automatic recognition is very challenging to solve for many reasons. Firstly, a user task is a result of a complex human navigational behavior because the same user may perform the same task twice differently. Secondly, users doing a same task do not necessarily have the same succession of fixed AOI. Thirdly, same AOI can be similarly relevant to perform different tasks.

4

Utility of Task Recognition

Knowing user tasks can tell us about: • First, second and last performed tasks: the order of tasks execution can provide insights about the user cognitive strategy. • Tasks that consume more or less users time: the task execution time can give clues about the most attractive or ignored tasks. • Most repeated tasks: task frequency can provide evidences about the difficulty of the task. Thus, automatic task identification can improve the general interaction process by giving help in real time to unfamiliar users, helping users with disabilities, improving systems security and interfaces usability. In pattern recognition applications, probabilistic graphical models have been successfully used and various studies prove that Conditional Random Fields (CRF) outperforms Hidden Markov Models (HMM) in sequence labeling tasks [14,15]. Motivated by this assumption, we propose to use a CRF model to automatically recognize a web user task based on mouse movements. Thereafter the recognition rate of our proposed model will be compared with HMM model developed by Elbahi et al.[16]. In next section, we briefly present the basics of HMM and CRF approaches in order to better understand the proposed model.

5

HMM and CRF :A Brief Presentation

For more than two decades, Hidden Markov Models [17] have been widely used in various fields for modeling and labeling stochastic sequences such as speech, object and web users activity recognition[18]. Recently, CRF theory [7], have been proposed to alleviate HMM assumptions and then have been used successfully in various fields such as the information retrieval process[19,20]. To better understand why CRF gained more power and then more popularity with respect to the HMM approach, we must have at least basics of both approaches. First, HMM are generative probabilistic directed graphical models which makes two important assumptions. Assumption 1: Each label yt , depends only on its previous label yt−1 . Assumption 2: Each observation xt depends on the current label yt .

744

A. Elbahi and M.N. Omri

Based on these simplifications, a HMM defined over a set of N hidden states and an alphabet of discrete symbols, can be specified by λ = (A, B, Π) with A = {aij } is the matrix of transition probabilities, B = {bj (k)} is the matrix of observations emission probabilities and Π = {πi } is the initial probability distribution vector over initial states. Learning HMM parameters A, B and Π is done by maximizing the joint probability P (Y, X) in the training data, using Baum-Welch algorithm. P (Y, X) =

T 

P (xt |yt )P (yt |yt−1 )

(1)

t=1

Secondly, CRF are discriminative undirected graphical models. Due to the discriminative nature of CRF, it becomes possible to represent much more knowledge in the model using feature functions. With CRF we try to maximize a conditional probability P (Y |X):  T N  N T    1 P (Y |X) = exp λk fk (yt−1 , yt , X) + +μk gk (yt , X) (2) Z(X) t=1 t=1 k=1

k=1

HMM which is based on (1) and CRF, based on (2), are very similar because λk fk (yt−1 , yt , X) are similar to transition probability P (yt |yt−1 ) and μk gk (yt , X) are analogous to observation probability emission P (xt |yt ). For more details about HMM and CRF approaches, reader can see [21,22].

6

CRF for User Task Modeling

6.1

User Task Modeling

Like presented above, a web interface can be presented as a set of items called AOI which can be pointed by mouse cursor during users tasks.

Fig. 3. Areas Of Interest in Equation Grapher interface.

Figure 3 presents the “Equation Grapher” online simulator1 interface described as a set of 15 areas of interest AOI={A,B,C,D,E,F,G,H,I,J,K,L,M,N,O} judged by an expert as frequently pointed regions during users tasks. 1

Phet available on : http://phet.colorado.edu.

Web User Interact Task Recognition based on Conditional Random Fields

745

Like presented previously, each task can be defined as a finite, temporal, stochastic sequence of AOI set by a user during a period of time. In order to get a sequence of fixed AOI during a task, coordinates of mouse cursor have been recorded at each time slice t. The obtained data (vector of cursor coordinates) will be provided as input for the following algorithm. In output, the vectorization algorithm provides the sequence of AOI related to the mouse movement coordinates during user task.

Algorithm 1. Vectorization algorithm Input: X={}; Coordinates of mouse cursor path recorded for a task; Details of each area of interest AOI; T ← Total duration of the task; Δt ← Time between recordings of two cursor coordinates; begin for t ← 1 to T (with Δt step) do if (mouse cursor coordinates is in AOIk ) then X[t] ← AOIk end end end Output: X = {x1 , x2 , . . . , xT } // the observation sequence

6.2

The Proposed Model

In this section we present the proposed CRF model. • TASKs= {tsk 1, tsk 2,. . . , tsk M} : set of M labels concerning M tasks that can be performed by users. • AOI={aoi 1, aoi 2,. . . , aoi N} : set of N AOI of the web interface that can be pointed by users during tasks. • X={aoi k1 ,. . . , aoi kt , . . . , aoi kT } :the sequence of observations describing AOI fixed by mouse cursor during a task, with 1 ≤ k ≤ N • Y={tsk i1 ,. . . , tsk it , . . . , tsk iT } : the label sequence describing the interaction process. At each time step (t) only one label is used to describe the task of user, with 1 ≤ i ≤ M. Each observation sequence X given to CRF model corresponds to a sequence of fixed AOI during only one task. So, the model is designed to infer only one task for a given observation sequence Y. Thus, each sequence of observations given to model must be entirely labeled using a single tag corresponding to performed task. Graphically, our model can be presented as follows:

746

A. Elbahi and M.N. Omri

Fig. 4. Graphical representation of CRF proposed model.

Features functions are key components of any CRF model and must be well defined before model training. Let F = {f1 , f2 , . . . , fn } be a set of features functions. Each function fj (yt−1 , yt , X, t) looks at a pair of adjacent labels (yt−1 and yt ) and all the observation sequence (X) at each time step (t). For the proposed model, yt refers to the current task label (tsk it ) and xt represents the current observation at time t (aoi kt ). In order to define, to train and to test the model, we used CRF++ tool2 that implements a fast learning Newton’s method LBFGS and a decoding method using Viterbi algorithm. Using CRF++ we can define templates to automatically generate a set of features functions. Next, we present some examples of used features functions generated using CRF++ templates. Template 1 : U00 :%x[0,0] , generate a set of functions like:  1 if yt ∈ TASKs and xt ∈ AOI ; fi (yt−1 , yt , X, t) = 0 otherwise For example:  1 if yt =tsk 1 and xt =aoi 2 ; f1 (yt−1 , yt , X, t) = 0 otherwise f1 return 1 if the current label (yt ) is tsk 1 and the current observation (xt ) is aoi 2 else f1 return 0. Template 2 : U01 :%x[-1,0]/%x[0,0]/%x[1,0] , generate a set of functions like: = f  j (yt−1 , yt , X, t) 1 if yt ∈ TASKs and xt ∈ AOI and xt+1 ∈ AOI and xt−1 ∈ AOI; 0 otherwise For example: =  f2 (yt−1 , yt , X, t) 1 if yt =tsk 2 and xt =aoi 1 and xt+1 =aoi 2 and xt−1 =aoi 4; 0 otherwise f2 return 1 if the current label (yt ) is tsk 2 and the current observation (xt ) is aoi 1 and next observation (xt+1 ) is aoi 2 and previous observation (xt−1 ) is aoi 4 else f2 return 0. After defining the model structure and features functions, we train the model in order to reestimate the features functions parameters vector Θ. For the train|D| ing step, we use a training set D defined by: D = {x(i) , y (i) }i=1 and containing 2

CRF++ available on: http://crfpp.googlecode.com/svn/trunk/doc/index.html.

Web User Interact Task Recognition based on Conditional Random Fields

747

labeled observation sequences. Each one corresponds to a single task and entirely labeled using a single tag.

7

Experiments and Discussion

7.1

Experimental Settings

Once, the structure of the CRF model and the set of features functions have been defined as previously described, we prepared a “training and test” set to validate the proposed model. The prepared data set is based on real manipulations, which consists of three different tasks performed by 51 students of secondary school using the “equation grapher” interface. The three tasks are: Task1 = called “DEG2”: during this task a student is asked to perform a graphical representation of a quadratic equation of the form ax2 + bx + c = 0 (a, b and c =0) and to keep in memory the shape of the drawn curve. Task2 = called “DEG1”: student is asked to perform a graphical representation of a quadratic equation of the form ax2 + bx + c = 0 (a=0 and b, c =0) and to keep in memory the shape of the drawn curve. Task3 = called “INT”, during this task a student is asked to discover (and keep in memory) intersection coordinates of a quadratic equation of the form ax2 + bx + c = 0 (a,b,c =0) and the same quadratic equation ax2 + bx + c = 0 (a,c=0). Like described below, the three tasks are very similar and complex to distinguish based only on cursor trajectory. The properties of used training set are summarized in the following table: Table 1. Characteristics of used training set Characteristics Interface number Areas Of Interest in used interface Sentences (Performed tasks) Task 1 (DEG2) Task 2 (DEG1) Task 3 (INT) Words (AOI pointed during all tasks) Used tags

Number 1 15 51 17 17 17 79304 3

Once, different AOI have been defined as described by figure 3. In order to obtain a sequence of observations, we asked each user to perform only one task (among three mentioned above) and we use OGAMA tool3 for recording mouse cursor coordinates at each time slice (1 centisecond) during a task. The obtained data are given to a vectorization algorithm to produce an observation sequence corresponding to the performed task. Once, the 51 observation sequences are 3

OGAMA available on : http://www.ogama.net/.

748

A. Elbahi and M.N. Omri

prepared and labeled, we estimate the parameters of the CRF model. The used sampling technique is LOOCV (Leave One Out Cross Validation) consisting of taking, at each time, one single task (labeled sequence of observations) from the “training and test” set for testing .The remaining tasks are used for training. Obtained results are shown and discussed in the next section. 7.2

Experimental Results and Discussion

Although, three tasks are quite similar as described above, experimental results, presented in table 2, showed that HMM model recognize 76.47% of users tasks while CRF model make out 88.23%.These results show that CRF proposed model have better ability than the HMM model developed by Elbahi et al. [16]. Table 2. Recognition rate of CRF and HMM models HMM CRF Task Type Samples Errors Recognition rate Samples Errors Recognition rate Task1(DEG2) 17 8 52,94% 17 3 82,35% Task2 (DEG1) 17 0 100% 17 2 88,23% Task3(INT) 17 4 76,47% 17 1 94,11% Total 51 12 76,47% 51 6 88,23%

To explain why CRF outperforms HMM approach in user task recognition, remember that a task is defined as a finite, temporal, stochastic sequence of AOI manipulated during a period of time T. Each task is described by a sequence of observations X={aoi k1 ,. . . , aoi kt , . . . , aoi kT }. Therefore, in order to recognize a given task, it is necessary to take into consideration all focused AOI during a given task. HMM models are founded on the independence assumption which says that variables do not depend on each other and they do not affect each other in any way except as allowed by the Markov property. Consequently, each used AOI at time t depends only on AOI at time t-1, and each label at time t depends only on the current observation at time t. This is not always the case in real applications such as web user tasks. Due to primary advantage of CRF approach which is the relaxation of the independence assumption, CRF models can take into account more complex dependencies between variables. Therefore, all focused AOI during a task can be taken into consideration by CRF models. For this reason, CRF outperforms HMM model in users tasks recognition. As shown in table 2, for task DEG1 HMM performed better than CRF, because when performing task DEG1, users unconsciously spend more time pointing at some AOIs that are more relevant for task DEG2 and task INT. This bad annotation has a bigger impact on the accuracy of discriminative models because during learning step, classes compete to find the best discriminative fit, while in generative model the parameters for each class are learned separately.

Web User Interact Task Recognition based on Conditional Random Fields

749

Table 3 shows the average cursor fixations in each AOI for tasks DEG2, DEG1 and INT. The same table also presents average cursor fixations in each AOI two tasks, the first called TASKX and the second called TASKY. TASKX is INT task and correctly recognized by both models while TASKY is INT task and judged by CRF and HMM models as DEG2 task. Likewise, Table 3 shows that areas A, H, K, L and N are rarely fixed during three tasks, therefore they can be considered as unimportant areas [23] which do not attract the user cursor during interaction. Such ignored areas do not have a great impact on user strategy during tasks. As shown in table3, task DEG2 is too dependent on areas E (28.84%) and B(15.18%) and task DEG1 is too dependent on areas G(21.84%) , F(18.57%), D(14.91%) and C(14.74%) while most used AOI for task INT are O(24.51%) and E(13.64%). These results show that each type of task attracts user attention into well defined regions in the interface. Therefore, during each task mouse movements draw the user behavior in a well defined order. This trajectory can be used to describe the strategy of user during interaction process [24]. Table 3. Mouse fixations rate per AOI during tasks AOIs DEG2 DEG1 INT TASKX TASKY

A 1,81 1,02 0,00 0,00 0,00

B 15,18 2,57 10,60 6,91 18,86

C 6,33 14,74 9,14 16,31 6,93

D 6,47 14,91 10,03 6,01 21,24

E 28,84 3,61 13,64 9,85 29,51

F 11,05 18,57 6,83 4,93 2,84

G 4,57 21,84 7,82 6,33 0,53

H 0,84 2,05 0,82 0,00 2,51

I 7,45 11,25 4,84 2,11 9,30

J 4,96 3,45 1,49 3,33 1,58

K 1,45 0,69 1,40 0,45 1,12

L 1,04 1,39 0,80 0,13 1,12

M 1,04 0,17 7,66 9,60 0,00

N 1,04 0,03 0,43 0,00 0,00

O 7,93 3,71 24,51 34,04 4,47

TASKX which is INT task was correctly recognized by CRF and HMM models, in fact, mouse fixations rate of TASKX show that user focuses on areas E(9.85%), M(9.60%) and O(34.04%) which are more relevant for INT task than DEG1 and DEG2 tasks. During TASKY which is of type INT, but recognized by both models as DEG2, the user usually focuses on relevant AOI for task DEG2 E(29.51%) and B(18.86%) and ignores area M (0.00%) and O(4.47%) considered as important to perform a task INT. Knowing that all users have successfully performed the required tasks, we can see that during TASKY, the user adopts a different strategy to perform an INT task. This explains the failure of two models in recognizing TASKY. CRF and HMM models estimate their parameters based on observation sequences. Each one describes the user strategy during a task based on fixed AOI. If the majority of users adopt a similar way, “strategy of group”, to perform a given task, HMM and CRF models will adjust their configurations based on this strategy of the group. To perform a given task, a human may adopt a strategy which is quite different of the one adopted by the majority of users; this task may be the cause of CRF and HMM failure. In fact a normal realisation of a given task results in a normal use of important AOI which are relevant for this task and ignoring (non-use) of important areas (region, link, button), or overusing of unimportant areas, should be considered as an indicator of different user strategy during task realization.

750

8

A. Elbahi and M.N. Omri

Conclusion

User interaction analysis based on mouse movement tracking can tell us about the users intended task, AOIs that highly attract the user’s attention and ignored AOIs. Also, the analysis of cursor behavior can give insights about the strategy adopted by the majority of users and the particular users strategy during a given task. Furthermore, HMM and CRF models present a good ability in human activity recognition with a clear superiority of CRF compared to HMM. In this work, we used CRF approach in order to recognize tasks performed by users, based on mouse movements during interaction process with web interface. Experimental results show the good performance of the proposed model and confirm the superiority of CRF with respect to HMM in user task recognition mainly based on mouse movements. Also, results show that each task type have a great impact on mouse behavior because the cursor is more attracted by some AOI than others according to each type of task. In spite of the effectiveness of probabilistic models, such as HMM and CRF, in labeling sequences especially in human activity recognition, there are some constraints that overwhelm the power of used approaches to model the human interaction due to the complex nature of human behavior and because during parameter estimation, CRF and HMM will be adjusted based on “the group’s strategy” adopted by the majority of users during a realisation of a given task.

References 1. Bulling, A., Ward, J., Gellersen, H., Tr¨ oster, G., et al.: Eye movement analysis for activity recognition using electrooculography. IEEE Transactions on Pattern Analysis and Machine Intelligence 33(4), 741–753 (2011) 2. Guo, Q., Agichtein, E.: Exploring mouse movements for inferring query intent. In: Proceedings of the 31st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 707–708. ACM (2008) 3. Dufresne, A., Courtemanche, F., Prom Tep, S., Senecal, S.: Physiological measures, eye tracking and task analysis to track user reactions in user generated content. In: 7th Internationcal Conference on Methods and Techniques in Behavioral Research (Measuring Behavior 2010), p. 218 (2010) 4. Huang, J., White, R.W., Dumais, S.: No clicks, no problem: using cursor movements to understand and improve search. In: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pp. 1225–1234. ACM (2011) 5. Omri, M.N.: Fuzzy knowledge representation, learning and optimization with bayesian analysis in fuzzy semantic networks. In: 1999 Proceedings of the 6th International Conference on Neural Information Processing (ICONIP 1999), vol. 1, pp. 412–417. IEEE (1999) 6. Omri, M.N., Tijus, C.A.: Uncertain and approximative knowledge representation in fuzzy semantic networks. In: The Twelfth International Conference On Industrial & Engineering Applications of Artificial Intelligence & Expert Systems IEA/AIE 1999, Cairo, Egypt (1999) 7. Lafferty, J., McCallum, A., Pereira, F.C.N.: Probabilistic models for segmenting and labeling sequence data, Conditional random fields (2001)

Web User Interact Task Recognition based on Conditional Random Fields

751

8. Chen, M.C., Anderson, J.R., Sohn, M.H.: What can a mouse cursor tell us more?: correlation of eye/mouse movements on web browsing. In: CHI 2001 Extended Abstracts on Human Factors in Computing Systems, pp. 281–282. ACM (2001) 9. Zheng, N., Paloski, A., Wang, H.: An efficient user verification system via mouse movements. In: Proceedings of the 18th ACM Conference on Computer and Communications Security, pp. 139–150. ACM (2011) 10. Heimg¨ artner, R.: Identification of the User by Analyzing Human Computer Interaction. In: Jacko, J.A. (ed.) HCI International 2009, Part III. LNCS, vol. 5612, pp. 275–283. Springer, Heidelberg (2009) 11. Mathe, S., Sminchisescu, C.: Action from still image dataset and inverse optimal control to learn task specific visual scanpaths. In: Advances in Neural Information Processing Systems, pp. 1923–1931 (2013) 12. Nardi, B.A.: Activity theory and human-computer interaction. Context and Consciousness: Activity Theory and Human-Computer Interaction, pp. 7–16 (1996) 13. Gotz, D., Zhou, M.X.: Characterizing users’ visual analytic activity for insight provenance. Information Visualization 8(1), 42–55 (2009) 14. Quattoni, A., Collins, M., Darrell, T.: Conditional random fields for object recognition. In: Advances in Neural Information Processing Systems, pp. 1097–1104 (2004) 15. Vail, D.L., Veloso, M.M., Lafferty, J.D.: Conditional random fields for activity recognition. In: Proceedings of the 6th International Joint Conference on Autonomous Agents and Multiagent Systems, p. 235. ACM (2007) 16. Elbahi, A., Mahjoub, M.A., Omri, M.N.: Hidden markov model for inferring user task using mouse movement. In: Fourth International Conference on Information and Communication Technology and Accessibility (ICTA), pp. 1–7. IEEE (2013) 17. Rabiner, L.R.: A tutorial on hidden markov models and selected applications in speech recognition. Proceedings of the IEEE 77(2), 257–286 (1989) 18. Elbahi, A., Omri, M.N., Mahjoub, M.A.: Possibilistic reasoning effects on hidden markov models effectiveness. In: IEEE International Conference on Fuzzy Systems (FUZZ-IEEE). IEEE (in press, 2015) 19. Fkih, F., Omri, M.N.: Complex terminology extraction model from unstructured web text based linguistic and statistical knowledge. International Journal of Information Retrieval Research (IJIRR) 2(3), 1–18 (2012) 20. Boughamoura, R., Omri, M.N., Youssef, H.: Use of conditional random field to extract pertinent information from web resources. In: An International Forum for Exploring e-Business and e-Government Research, Applications, and Technologies (TIGERA) (2008) 21. Rabiner, L.R., Juang, B.-H.: An introduction to hidden markov models. IEEE ASSP Magazine 3(1), 4–16 (1986) 22. Sutton, C., McCallum, A.: An introduction to conditional random fields for relational learning. Introduction to Statistical Relational Learning, pp. 93–128 (2006) 23. Poole, A., Ball, L.J.: Eye tracking in hci and usability research. Encyclopedia of Human Computer Interaction 1, 211–219 (2006) 24. Tan, D., Nijholt, A.: Brain-computer interfaces and human-computer interaction. In: Brain-Computer Interfaces, pp. 3–19. Springer (2010)