Ana səhifə

The psychology of linguistic form

Yüklə 131.5 Kb.
ölçüsü131.5 Kb.
  1   2   3
Lee Osterhout, Richard A. Wright, and Mark D. Allen
To appear in the Cambridge Encyclopedia of the Language Sciences

Humans can generate and comprehend a stunning variety of conceptual messages, ranging from sophisticated types of mental representations, such as ideas, intentions, and propositions, to more primal messages that satisfy demands of the immediate environment, such as salutations and warnings. In order for these messages to be transmitted and received, however, they must be put into a physical form, such as a sound wave or a visual marking. As noted by the Swiss linguist de Saussure (2002), the relationship between mental concepts and physical manifestations of language is almost always arbitrary. The words cat, sat, and mat are quite similar in terms of how they sound, but are very dissimilar in meaning; one would expect otherwise if the relationship between sound and meaning was principled instead of arbitrary. Although the relationship between linguistic form and meaning is arbitrary, it is also highly systematic. For example, changing a PHONEME in a word predictably also changes its meaning (as in the cat, sat, and mat example).

Human language is perhaps unique in the complexity of its linguistic forms (and, by implication, the system underlying these forms). Human language is compositional; that is, every sentence is made up of smaller linguistic units that have been combined in highly constrained ways. A standard view (Chomsky 1965, Pinker 1999) is that units and rules of combination exist at the levels of sound (phonemes and PHONOLOGY), words (MORPHEMES and MORPHOLOGY), and sentences (words and phrases, and SYNTAX). Collectively, these rules comprise a grammar that defines the permissible linguistic forms in the language. These forms are systematically related to, but distinct from, linguistic meaning (SEMANTICS).

Linguistic theories, however, are based on linguistic description and observation and therefore have an uncertain relation to the psychological underpinnings of human language. Researchers interested in describing the psychologically relevant aspects of linguistic form require their own methods and evidence. Furthermore, psychological theories must not only describe the relevant linguistic forms but also the processes that assemble these forms (during language production) and disassemble them (during language comprehension). Such theories should also explain how these forms are associated with a speaker’s (or hearer’s) semantic and contextual knowledge. Here, we review some of what we have learned about the psychology of linguistic form, as it pertains to sounds, words, and sentences.

Sound units. Since the advent of speech research, one of the most intensively pursued topics in speech science has been the search for the fundamental sound units of language. Many researchers have found evidence for phonological units that are abstract (i.e., generalizations across any number of heard utterances, rather than memories of specific utterances) and componential (constituent elements that operate as part of a combinatorial system). However, there is other evidence for less abstract phonological forms that may be stored as whole words. As a result two competing hypotheses about phonological units have emerged: an abstract componential one vs. a holistic one.

The more widespread view is the componential one. It posits abstract units that typically relate either to abstract versions of the articulatory gestures used to produce the speech sounds (Liberman and Mattingly 1985, Browman and Goldstein 1990), or to ones derived from descriptive units of phonological theory such as the feature (see FEATURE ANALYSIS) an abstract sub-phonemic unit of contrast; the phoneme, an abstract unit of lexical contrast that is either a consonant or a vowel; the phone or allophone, surface variants of the phoneme; the syllable, a timing unit that is made up of a vowel and one or more of its flanking consonants; the prosodic word, the rhythmic structure that relates to patterns of emphasized syllables; or various structures that related to tone, the lexically contrastive use of the voice’s pitch, and intonation, the pitch-based tune that relates to the meaning of a sentence (for reviews see Frazier 1995; Studdert-Kennedy 1980).

In the holistic view, the word is the basic unit while other smaller units are considered to be epiphenomenal (e.g., Goldinger, Pisoni, and Logan, 1991). Instance-specific memory traces of particular spoken words are often referred to as episodes. Proponents of this view point out that while abstract units are convenient for description and relate transparently to segment-based writing systems, such as those based on the alphabet, there is evidence that listeners draw on a variety of highly detailed and instance-specific aspects of a word’s pronunciation in making lexical decisions (for reviews see Goldinger and Azuma 2003; Nygaard and Pisoni 1995).

Some researchers have proposed hybrid models in which there are two layers of representation: the episodic layer in which highly detailed memory traces are stored, and an abstract layer organized into features or phones (Scharenborg, Norris, ten Bosch, and McQueen 2005). The proponents of hybrid models try to capture the instance-specific effects in perception that inspire episodic approaches as well as the highly abstracted lexical contrast effects.

Processes. SPEECH PRODUCTION refers to the process by which the sounds of language are produced. The process necessarily involves both a planning stage, in which the words and other linguistic units that make up an utterance are assembled in some fashion, and an implementation stage in which the various parts of the vocal tract, for example the articulators, execute a motor plan to generate the acoustic signal. See Fowler (1995) for a detailed review of the stages involved in speech production. It is worth noting here that even if abstract phonological units such as features are involved in planning an utterance, at some point the linguistic string must be implemented as a motor plan and a set of highly coordinated movements. This has motivated gestural representations that include movement plans rather than static featural ones (Browman and Goldstein 1990; Fowler 1986, 1996; Saltzman and Munhall 1989; Stetson 1951).

SPEECH PERCEPTION is the process by which human listeners identify and interpret the sounds of language. It too necessarily involves at least two stages: 1) the conversion of the acoustic signal into an electrochemical response at the auditory periphery and 2) the extraction of meaning from the neurophysiological response at the cortical levels. Moore (1989) presents a thorough review of the physiological processes and some of the issues involved in speech perception. A fundamental point of interest here is perceptual constancy in the face of a massively variable signal. Restated as a question: how is it that a human listener is able to perceive speech sounds and understand the meaning of an utterance given the massive variability created by physiological idiosyncrasies and contextual variation? The various answers to this question involve positing some sort of perceptual units, be they individual segments, sub-segmental features, coordinated speech gestures, or higher level units like syllables, morphemes, or words.

It is worth noting here that the transmission of linguistic information does not necessarily rely exclusively on the auditory channel; the visible articulators, the lips and to a lesser degree the tongue and jaw, also transmit information; a listener presented with both auditory and visual stimuli will integrate the two signals in the perceptual process (e.g., Massaro 1987). When the information in the visual signal is unambiguous (as when the lips are the main articulators) the visual signal may even dominate the acoustic one (e.g., McGurk and Macdonald 1976). Moreover, writing systems convey linguistic information albeit in a low-dimensional fashion. Most strikingly, sign languages are fully as powerful as speech-based communication systems and are restricted to the visual domain. Despite the differences between signed and spoken languages in terms of the articulators and their perceptual modalities, they draw on the same sorts of linguistic constituents, at least as far as the higher level units are concerned: syllable, morpheme, word, sentence, prosodic phrase (e.g., Brentari 1998). Some have also proposed decomposing signed languages into smaller units using manual analogs of phonological features despite the obvious differences in the articulators and the transmission media (for a review see Emmory 2002). The parallel of signed and spoken language structure despite the differences in transmission modalities is often interpreted as evidence for abstract phonological units at the level of the mental lexicon (Meier, Cormier, and Quinto-Pozos 2002).

The history of the debate: Early phonological units. The current debate about how to characterize speech sounds has its roots in research that dates back over a century. Prior to the advent of instrumental and experimental methods in the late 19th century, it was commonly accepted that the basic units of speech were discrete segments that were alphabetic in nature and serially ordered. While it was recognized that speech sounds varied systematically depending on the phonetic context, the variants themselves were thought to be static, allophones, of an abstract and lexically contrastive sound unit, that is, a phoneme. Translating into modern terminology, phonological planning involved two stages: 1) determining the contextually determined set of discrete surface variants given a particular lexical string and 2) concatenating the resulting allophones. The physiological implementation of the concatenated string was thought to result in a series of articulatory steady states or postures. The only continuous aspects of sound production were believed to be brief transitional periods created by articulatory transitions from one state to the next. The transitional movements were thought to be wholly predictable and determined by the physiology of a particular speaker’s vocal tract. Translating again into modern terminology, perception (when considered) was thought to be simply the process of translating the allophones back into their underlying phonemes for lexical access. The earliest example of the phoneme-allophone relationship is attributed to Panini c. 500 BC who’s sophisticated system of phonological rules and relationships influenced structuralist linguists of the early 20th century as well as generative linguists of the late 20th century (for a review see Anderson 1985; Kiparsky 1979).

The predominant view at the end of the 19th century was typified by Bell’s (1867) influential descriptive work on English pronunciation. In it, he presented a set of alphabet-inspired symbols whose shapes and orientations were intended to encode both the articulatory steady states and their resulting steady state sounds. A fundamental assumption in the endeavor was that all sounds of human language could be encoded as a sequence of universal articulatory posture complexes whose subcomponents were shared by related sounds. For example, all labial consonants (p, b, m, f, v, w, etc.) shared a letter shape and orientation, while all voiced sounds (b, d, g, v, z, etc.) shared an additional mark to distinguish them from their voiceless counterparts (p, t, k, f, s, etc.). Bell’s formalization of a set of universal and invariant articulatory constituents aligned as an alphabetic string influenced other universal transcription systems such as Sweet’s (1881) Romic alphabet, which laid the foundation for the development of the International Phonetic Alphabet (Passy, 1888). It also foreshadowed the use of articulatory features, such as those proposed by Chomsky and Halle (1968) in modern phonology, in that each speech sound, and therefore each symbol, was made up of a set of universal articulatory components. A second way in which Bell’s work presaged modern research was the connection between perception and production. Implicit in his system of writing was the belief that perception of speech sounds was the process of extracting the articulations that produced them. Later perceptual models would incorporate this relationship in one way or another (Chistovich 1960; Dudley 1940; Fowler 1986, 1996; Joos 1948; Ladefoged and McKinney 1963; Liberman and Mattingly 1985; Stetson 1951).

The history of the debate: Early experimental research. Prior to the introduction of experimental methods into phonetics, the dominant methodologies were introspection about one’s own articulations and careful but subjective observations of others’ speech, and the measurement units were letter-based symbols. Thus, the observer and the observed were inextricably linked while the resolution of the measurement device was coarse. This view was challenged when a handful of phoneticians and psychologists adopted the scientific method and took advantage of newly available instrumentation, such as the kymograph, in the late 1800s. They discovered that there were no segmental boundaries in the speech stream and that the pronunciation of a particular sound varied dramatically from one instance to the next (for a review of early experimental phonetics see Kühnhert and Nolan 1999; and Minifie 1999). In the face of the new instrumental evidence, some scholars, like Sievers (1876), Rousselot (1897), and Scripture (1902) proposed that the speech stream, and the articulations that produced it, were continuous, overlapping, and highly variable rather than being discrete, invariant, and linear. For them the fundamental sound units were the syllable or even the word or morpheme. Rousselot’s research (1897-1901) revealed several articulatory patterns that were confirmed by later work (e.g., Stetson 1951). For example, he observed that when sounds that are transcribed as sequential are generated by independent articulators (such as the lips and tongue tip), they are initiated and produced simultaneously. He also observed that one articulatory gesture may significantly precede the syllable it is contrastive in, thereby presenting an early challenge to the notion of sequential ordering in speech.

Laboratory researchers like Stetson (1905, 1951) proposed that spoken language was a series of motor complexes organized around the syllable. He also first proposed that perception was the process of perceiving the articulatory movements that generate the speech signal. However, outside of the experimental phonetics laboratory, most speech researchers, particularly phonologists like Leonard Bloomfield (1933), continued to use phonological units that remained abstract, invariant, sequential, and letter-like. Three events that occurred in the late 1940s and early 1950s changed this view dramatically. The first of these events was the application to speech research of modern acoustic tools such as the spectrogram (Potter 1945), sophisticated models of vocal tract acoustics (e.g., House and Fairbanks 1953), reliable articulatory instrumentation such as high speed X-ray cineflourography (ex: Delattre and Freeman 1968), and electromyographic studies of muscle activation (Draper, Ladefoged, and Whitteridge 1959). The second was the advent of modern perception research in which researchers discovered complex relationships between speech perception and the acoustic patterns present in the signal (Delattre, Liberman, and Cooper 1955). The third was the development of distinctive feature theory in which phonemes were treated as feature matrices that captured the relationships between sounds (Jakobson 1939; Jakobson, Fant, and Halle 1952).

When researchers began to apply modern acoustic and articulatory tools to the study of speech production, they rediscovered and improved on the earlier observation that the speech signal and the articulations that create it are continuous, dynamic, and overlapping. Stetson (1951) can be seen as responsible for introducing kinematics into research on speech production. His research introduced the notion coproduction, in which articulatory gestures were initiated simultaneously, and gestural masking, in which the closure of one articulatory gesture hides another giving rise to the auditory percept of deletion. Stetson’s work provided the foundation for current language models which incorporate articulatory gestures and their movements as the fundamental phonological units (ex: Browman and Goldstein 1990; Byrd and Saltzman 2003; Saltzman and Munhall 1989).

In the perceptual and acoustic domains, the identification of perceptual cues to consonants and vowels raised a series of questions that remain at the heart of the debate to this day. The coextensive and covarying movements that produce the speech signal result in acoustic information that exhibits a high degree of overlap and covariance with information about adjacent units (e.g., Delattre, Liberman, and Cooper 1955). Any single perceptual cue to a particular speech sound can also be a cue to another speech sound. For example, the onset of a vowel immediately following a consonant provides the listener with cues that identify both the consonant and vowel (Liberman, Delattre, Cooper, and Gerstman 1954). At the same time, multiple cues may identify a single speech sound. For example, the duration of a fricative (e.g., “s”), the fricative’s noise intensity, and the duration of the preceding vowel all give information about whether the fricative is voiced (e.g., “z”) or voiceless (e.g., “s”) (Soli 1982). Finally, the cues to one phone may precede or follow cues to adjacent phones. The many-to-one, the one-to-many, and the non-linear relationships between acoustic cues and their speech sounds poses a serious problem for perceptual models in which features or phones are thought to bear a linear relationship to each other. More recently, researchers studying perceptual learning have discovered that listeners encode speaker-specific details and even utterance-specific details when they are learning new speech sounds (Goldinger and Azuma 2003). The latest set of findings pose a problem for models in which linguistic sounds are stored as abstract units.

In distinctive feature theory, each phoneme is made up of a matrix of binary features that encode both the distinctions and the similarities between one class of sounds and the next in a particular language (Jakobson, Fant and Halle 1952; Chomsky & Halle, 1968). The features are thought to be drawn from a language universal set, and thus allow linguists to observe similarities across languages in the patterning of sounds. Moreover, segmenting the speech signal into units that are hierarchically organized permits a duality of patterning of sound and meaning that is thought to give language its communicative power. That is, smaller units such as phonemes may be combined according to language- specific phonotactic (sound combination) constraints into morphemes and words, and words may be organized according to grammatical constraints into sentences. This means that with a small set of canonical sound units, together with recursion, the talker may produce and the hearer may decode and parse a virtually unbounded number of utterances in the language.

In this section we focus on those representations of form that encode meaning and other abstract linguistic content at the most minimally analyzable units of analysis—namely, words and morphemes. As such, we will give a brief overview of the study of lexical morphology, investigations in morphological processing, and theories about the structure of the mental lexicon.

Lexical form. What is the nature of a representation at the level of lexical form? We will limit our discussion here largely to phonological codes, but recognize that a great many of the theoretical and processing issues we raise apply to orthographic codes as well. It is virtually impossible for the brain to store exact representations for all possible physical manifestations of linguistic tokens that one might encounter or produce. Instead, representations of lexical form are better thought of as somewhat abstract structured groupings of phonemes (or graphemes) which are stored as designated units in long term memory, either as whole words or as individual morpheme constituents and associated with any other sources of conceptual or linguistic content encoded in the lexical entries that these form representations map onto. As structured sequences of phonological segments then, these hypothesized representational units of lexical form must be able to account for essentially all the same meaning-to-form mapping problems and demands that individual phonological segments themselves encounter during on-line performance, due to idiosyncratic variation among speakers and communicative environments. More specifically, representations of morphemes and words at the level of form must be abstract enough to accommodate significant variation in the actual physical energy profiles produced by the motor systems of individual speakers/writers at under various environmental conditions. Likewise, in terms of language production, units of lexical form must be abstract enough to accommodate random variation in the transient shape and status of the mouth of the language producer.

Form and meaning: Independent levels of lexical representation. The description of words and morphemes given above to some degree rests on the assumption that lexical form is represented independently from other forms of cognitive and linguistic information, such as meaning and lexical syntax (e.g., lexical category, nominal class, gender, verbal subcategory, etc.). Many theories of the lexicon have crucially relied on the assumption of separable levels of representation within the lexicon. In some sense, as explained by Allport and Funnell (1981), this assumption follows naturally from the arbitrariness of mapping between meaning and form identified above, and would thus appear to be a relatively non-controversial assumption.

The skeptical scientist, however, is not inclined to simply accept assumptions of this sort at face value without considering alternative possibilities. Imagine, for example, that the various types of lexical information stored in a lexical entry are represented within a single data structure of highly interconnected independent distributed features. This sort of arrangement is easy to imagine within the architecture of a CONNECTIONIST model (McClelland & Rumelhart 1986). Using the lexical entry “cat” as an example, imagine a connectionist system in which all the semantic features associated with “cat,” such as [whiskers], [domestic pet], etc. (which are also shared with all other conceptual lexical entities bearing those features, such as , , etc.) are directly associated with the phonological units that comprise its word form /k/, /ae/, /t/ (which are likewise shared with all other word forms containing these phonemes) by means of individual association links that directly tie individual semantic features with individual phonological units (Rueckl et al. 1997). One important consequence of this hypothetical arrangement is that individual word forms do not exist as free-standing representations. Instead, the entire lexical entry is represented as a vector of weighted links connecting individual phonemes to individual lexical semantic and syntactic features. It logically follows from this model, then, that if all or most of the semantic features of the word “cat,” for example, were destroyed or otherwise made unavailable to the processor, then the set of phonological forms /k/ /ae/ /t/, having nothing to link to, would have no means for mental representation, and would therefore not be available to the language processor. We will present here experimental evidence against this model, which instead, favors models in which a full phonological word (e.g., /kaet/) is represented in a localist fashion, and is accessible to the language processor, even when access to its semantic features is partially or entirely disrupted.

Several of the most prominent theories of morphology and lexical structure within formal linguistics make explicit claims about modularity of meaning and form (Anderson 1992). Jackendoff (1997), for example, presents a theory that has a tripartite structure, in which words have separate identities at three levels of representation -- form, syntax, and meaning -- and that these three levels are sufficient to encode the full array of linguistic information each word encodes. Jackendoff’s model provides further details in which it is proposed that our ability to store, retrieve, and use words correctly, as well as our ability to correctly compose morphemes into complex words, derives from a memorized inventory of mapping functions that pick out the unique representations or feature sets for a word at each level and associate these elements with one another in a given linguistic structure.

While most psycholinguistic models of language processing have not typically addressed the mapping operations assumed by Jackendoff, they do overlap significantly in terms of addressing the psychological reality of his hypothetical tripartite structure in the mental lexicon. Although most experimental treatments of the multi-level nature of the lexicon have been developed within models of language production, as will be seen below, there is an equally compelling body evidence for multi-level processing from studies of language comprehension as well.

The most influential lexical processing models over the last two decades make a distinction between at least two levels: the lemma level, where meaning and syntax are stored, and the lexeme level, where phonological and orthographic descriptions are represented. These terms and the functions associated with them were introduced in the context of a computational production model by Kempen and Huijbers (1983) and receive further refinement with respect to human psycholinguistic performance in the foundational lexical production models of Bock (1982), Garrett (1975), and Levelt (1989), Much compelling evidence for a basic lemma/lexeme distinction has come from analyses of naturally occurring speech errors generated by neurologically unimpaired subjects, including tip-of-the-tongue phenomena (Meyer and Bock 1992), as well as from systematic analyses of performance errors observed in patients with acquired brain lesions. A more common experimental approach, however, is the picture-word interference naming paradigm, in which it has been shown that lemma and lexeme level information can be selectively disrupted during the course of speech production (Schriefers, Meyer, and Levelt 1990).

In terms of lexical comprehension models, perhaps the most straightforward sources of evidence for a meaning/form distinction have come from analyses of the performance of brain-damaged patients. A particularly compelling case for the independence of meaning and form might be demonstrated if an individual with acquired language pathology were to show an intact ability to access word forms in his/her lexicon, yet remains unable to access meaning from those form representations. This is precisely the pattern observed in patients designated as suffering from word meaning deafness. These patients show a highly selective pattern of marked deficit in comprehending word meanings, but with perfect or near perfect access to word forms. A good example is patient WBN as described in Allen (2005), who showed an entirely intact ability to access spoken word form representations. In an auditory lexical decision task, WBN scored 175/182 (96%) correct, which shows he can correctly distinguish real words from non-words (e.g., flag vs. flig), presumably relying on preserved knowledge of stored lexemes to do so. However, on tasks that required WBN to access meaning from spoken words, such as picture to word matching tasks, he performed with only 40-60% accuracy (at chance in many cases).

Lexical structure: Complex words. A particularly important issue in lexical representation and processing concerns the cognitive structure of complex words, that is, words composed of more than one morpheme. One of the biggest debates surrounding this issue stems from the fact that in virtually all languages with complex word structures, lexical information is encoded both in consistent, rule-like structures, as well as idiosyncratic, irregular structures. This issue can be put more concretely in terms of the role of morphological decomposition in single-word comprehension theories within psycholinguistics. Consider the written word wanted, for example. A question for lexical recognition theories is whether the semantic/syntactic properties of this word [WANT, Verb, +Past, …] are extracted and computed in a combinatorial fashion each time wanted is encountered—by accessing the content associated with the stem want- [WANT, Verb] and combining it with the content extracted from the affix -ed [+Past]—or whether instead a single whole-word form wanted is stored at the lexeme level and associated directly with all its semantic/syntactic content. To understand the plausibility that a lexical system could in principle store whole-word representations such as wanted, one must recognize that in many other cases, such as those involving irregularly inflected words, such as taught, the system cannot store a stem and affix at the level of form, as there are no clear morpheme boundaries to distinguish these constituents, but must instead obligatorily store it as a whole-word at the lexeme level.

Many prominent theories have favored the latter, non-decompositional, hypothesis for all words, including irregular words like taught as well as regular compositional words like wanted (Bybee 1988). Other influential processing models propose that complex words are represented as whole-word units at the lexeme level, but that paradigms of inflectionally related words (want, wants, wanted) map onto a common representation at the lemma level (Fowler 1985). In addition to this, another class of models, which has received perhaps the strongest empirical support, posits full morphological decomposition at the lexeme level whenever possible (Allen and Badecker 1999). According to these fully decompositional models, a complex word like wanted is represented and accessed in terms of its decomposed constituents want- and -ed at the level of form, such that the very same stem want- is used during the recognition of want, wants, and wanted. According to these models, then, the recognition routines that are exploited by morphological decomposition at the level of form resemble those in theoretical approaches to sentence processing, in which meaning is derived compositionally by accessing independent units of representation of form and combining the content that these forms access into larger linguistic units, according to algorithms of composition specified by the grammar.

While there is compelling empirical support for decompositional models of morphological processing, researchers are becoming increasingly aware of important factors that might limit decomposition. These factors are regularity, formal and semantic transparency, and productivity.

Regularity refers to the reliability of a particular word formation process. For example, the plural noun kids expresses noun-plurality in a regular, reliable way, while the plural noun children does not.

Formal transparency refers to the degree to which the morpheme constituents of a complex structure are obvious from its surface form. For example, morpheme boundaries are fairly obvious in the transparently inflected word wanted, compared to those of the opaquely (and irregularly) inflected word taught.

Semantic transparency. Although an irregular form like taught is formally opaque, as defined above, it is nonetheless semantically transparent, because its meaning is a straightforward combination of the semantics of the verb teach and the feature [+Past]. In contrast to this, an example of a complex word which is formally transparent, yet semantically opaque is the compound word dumbbell, which is composed of two recognizable morphemes, but the content associated with these two surface morphemes do not combine semantically to form the meaning of the whole word.

Productivity describes the extent to which a word formation process can be used to form new words freely. For example, the suffix -ness is easily used to derive novel nouns from adjectives (e.g., nerdiness, awesomeness, catchiness), while the ability to form novel nouns using the analogous suffix -ity is awkward at best (?nerdity) if not impossible.

Another phenomenon associated with these lexical properties is that they tend to cluster together in classes of morphologically complex word types across a given language, such that there will often exist a set of highly familiar, frequently used forms that are irregular, formally opaque and non-productive, and also a large body of forms that are morphologically regular, formally transparent, and productive. Given the large variety of complex word types found in human languages with respect to these dimensions of combinability, as well as the idiosyncratic nature of the tendency for these dimensions to cluster together from language to language, it would appear that empirical evidence for morphological decomposition must be established on a “case-by-case” basis for each word-formation type within each language. This indeed appears to be the direction that most researchers have taken.

  1   2   3

Verilənlər bazası müəlliflik hüququ ilə müdafiə olunur © 2016
rəhbərliyinə müraciət