Meaning-Text Theory (MTT) is a linguistic framework for machine translation (MT) launched in Moscow in the late 1960s by A. Zholkovsky and I. Mel'čuk (Mel'čuk Reference Mel'čuk1981). This theory persists in the core question: How is meaning expressed in language? Meaning-Text Model (MTM), the model proposed by MTT, attempts to answer this question by simulating the linguistic activity of native speakers. The activity refers to a procedure that a speaker encodes a purely linguistic meaning into a particular linguistic utterance (text), i.e. a synthesis. In MTT, the procedure is represented as a many-to-many bi-direction ‘{SemR_i} ⇐ language ⇒ {PhonR_j} | i ≠ j, 0 < i, j≦∞’. Obviously, it is impossible to express a ‘meaning’ to its corresponding phonological form directly. Stages of synthesis in MT systems always pass through semantic synthesis, syntactic synthesis, morphology synthesis, and phonology synthesis. However, MTM devotes itself to describe meaning to text correspondence. It further divides the synthetic procedure into seven independent stratificational linear levels of representation such as SemR(epresentation), D(eep)Synt(tactic)R, S(urface)SyntR, DMorph(ological)R, SMorphR, DPhon(ological)R, and SPhonR. In modular, there are six modules that ensure transition between two linguistic representations of adjacent levels; for instance, semantic module ensures transition between SemR and DSyntR, Deep-Syntactic module ensures transition between DSyntR and SSyntR, etc. The semantic module plays a crucial role in MTM, and serves as a starting point and a key topic of the volumes of the book: Semantics: From Meaning to Text (henceforth SMT).

SMT mainly concerns itself with the two levels of representation (SemR and DSyntR) of semantic module and a formal lexicon named Explanatory Combinatorial Dictionary (ECD). It arranges five main parts in three volumes (parts I and II in vol. 1, parts III and IV in vol. 2, and part V in vol. 3). The present volume, i.e. parts III and IV, is divided into five chapters, focusing on DSyntR, the rules and principles for corresponding operations as well as ECD.

Part III (pp. 1–93) provides a detailed expression of DSyntR. The ‘Introduction’ (pp. 3–17) briefly compares SemR and DSyntR, serving as a summarization of vol. 1 and as a connection to this new one. Similar to SemR (Mel'čuk Reference Mel'čuk, Embleton and Embleton1998: 6), DSyntR consists of the following four structures: ‘DSyntS(trcture), DSynt-Com(municative)S, DSynt-Ana(aphoric)S, DSynt-Pro(sodic)S’, with the first one as the central or basic structure, and the rest as peripheral structures.

Being a production-oriented theory, emphasis on the formal linguistic description of a language is one of the tenets of MTT. Recently, both dependency trees and phrase structure trees are used in formal linguistic description, while the former one concentrates on the description of words or the relationship among them. In Natural language processing (NLP), ‘semantic’ refers to the ‘meaning’ of words or idioms, i.e. interpretations of words. The semantic module in MTM is specific for the description of meaning. Therefore, the expression of module in MTM turns to the dependency way. In vol. 1, SemR, the first level of correspondence in the semantic module of MTM, is visually expressed in a dependency network. In the present volume, the second level DSyntS is designed to express a sentence S's meaning completely in a formal manner. It dedicates itself to a particular sentence, especially to its formal structure, its actual physical components, and the arrangement of linear order, by means of linearly unordered dependency tree.

A well-formed dependency tree is composed of nodes and the directed relations among them. Nodes of a dependency tree of DSyntS are deep lexical units, containing full lexical utterances (LUs), standard lexical functions (LFs), and fictitious lexemes, which can be either lexemes or idioms. Although DSyntS is a syntactic description of a sentence, it refers to the concrete meaning of a particular sentence. Thus, the semantically empty words such as governed prepositions, conjunctions, and auxiliary verbs are omitted. What's more, for those inflectional languages, the so-called ‘deep grammemes’ are labeled onto nodes as subscripts, functioning as marks of semantic inflectional categories, e.g. the voice of verbs like ACT(ive) DIR(ect) etc. The grammemes ensure correspondence between the keywords and their arguments or modifiers in the two adjoining structures – DSyntS and SSyntS – from the two-dimensional tree diagrams to their linear ordered sequences.

In the network of SemR, relations are represented by simple numbers ‘1, 2, 3   ’. Such numbers assure an existence of some kind of semantic relationship between every two connecting nodes in the network, but ignore their classification. Differently, relations in the DSyntS dependency trees are labeled as I, II, III, IV, V, VI, II_dir-sp, COORD, QUASI-COORD, APPEND, ATTR, and ATTR_descry, which are further clustered into the following four groups: ‘actantial’ (I, II, III, IV, V, VI, II_dir-sp), ‘coordinative’ (COORD, QUASI-COORD), ‘appenditive’, and ‘attributive’ (ATTR and ATTR_descry). Furthermore, they stand for a family of particular syntactic constructions, each corresponding to the SSyntR in every particular language. For example, the ‘actiantial’ relations connect the semantically predicative LU with all kinds of its actants, which map to the surface syntactic ‘verb-subject’, ‘verb-object’, and ‘verb-component’ constructions. The ‘coordinate relations’ correspond to coordinate constructions. The ‘attributive’ relations correspond to different types of syntactic modifications. The APPEND is then for all kinds of extra-structural constructions like the parenthesis, or interjections of a sentence. Clearly, the ‘semantic relation’ here in MTT is quite different with conventional linguistic semantic relations, i.e. cases.

Generally speaking, in the MT field, a language is considered to be comprising vocabulary and grammar. The nodes and arrows in the DSyntS dependency tree, the labels on the nodes and arrows, and the deep grammemes above constitute the vocabulary of the Deep-Syntactic Language of MTT. Besides, three subsets of grammar rules are needed: the formation rules to ensure the correctness of the DSyntS, the transformation rules for the inner-level paraphrasing, and the interpretation rules for all symbols.

The first three chapters of part IV (pp. 97–258) are dedicated to the grammar for the operation of the ‘SemR ⇔ DsyntR’ correspondence. MTT postulates that the synthesis of a sentence is a many-to-many ‘correspondence’ of semantic and phonetic representations of a sentence. Correspondence rules (‘A⇔B’, the transition between two corresponding LUs in two adjoining levels) and equivalence rules (‘A ≡ B’, the paraphrase between two synonymous LUs in the same level) are then used to interpret this postulate. So to speak, a particular representation in the sublevel of MTM, taking DSyntR and SemR as examples, can be obtained via at least three paths: (1) directly from underlying SemR, like ‘SemR⇔DsyntR’ by transition; (2) indirectly from SemR, with a Sem-paraphrasing from SemR to a SemR’ at first, like ‘SemR ≡     ≡ SemR'⇔DsyntR’; (3) directly paraphrased from a DsyntR’, which is transferred from a SemR, like ‘SemR⇔DSyntR’ ≡     ≡ DsyntR’ (p. 101). The three paths in the semantic module refer to ‘Semantic paraphrasing’ (chapter 8), ‘Deep-syntactic paraphrasing’ (chapter 9), and ‘Semantic transition’ (chapter 10) respectively.

Being a complete formal device for the description of natural languages, MTM has to be stored in some kind of database so as to be automatically recognized and run by machines. And as a semantic-oriented language theory, MTT does not give any primacy to syntax (grammar) as most of the current linguistic theories do. Conversely, it regards ‘the lexicons as the central, pivotal component of a linguistic description; the grammar is no more than a set of useful generalization over the lexicon’ (p. 262). It is based lexically, so the database for storing its constituents is a dictionary, named ECD, containing the E(xplanation) of each L(exeme) in a language and the C(ombinational) data of L's lexical (meaning) and syntactic (text) correspondence. It is considered ‘a basis for a full range of possible practical adaptations’ (p. 259) in MTM.

Section 1 of chapter 11 (pp. 260–275) defines ECD with its basic units, general properties, and specific properties. Considering the particularities of language expressions in different languages (Ls), i.e. particular collocations, the basic units of an ECD include both single word items and set phrases. ECDs are monolingual. However, all information about entries in different ECDs can be exhaustively described in a particularly similar way – the three zones: semantic zone, syntactic zone, and lexical zone (section 2, pp. 276–315). Most of the lexicons in MT systems mainly focus on the basic information of lexical entries, e.g. part-of-speech (POS), lexicographical explanations, etc. Compared with lexicons in most of the MT systems, the three zones make ECD quite different. On the basis of these three zones, the corresponding semantic arguments, syntactic functions, and collocations of a lexical entry can be interpreted in detail.

In an ECD, meaning of an entry is considered to be the first and foremost motivation. The description of an entry starts from the semantic zone, which defines all of its semantic properties and can be transferred into a semantic network. This kind of network confirms a dynamic connection between entries, and distinguishes families of synonymous LUs from languages.

The syntactic zone then links the entry's SemR to its syntactic representations in a rectangular matrix called government pattern (GP). No matter the ordinary dictionaries for language learners or the formal dictionaries for machines, they do not pay enough attention on the syntactic status of the entries. However, GP in ECDs lists all the possible syntactic combinations, positions, and the context of every entry.

The lexical zone majors in presenting the whole entry's semantic derivation and its restricted lexical co-occurrence by LFs. There are about sixty standard simple LFs in the ECDs of MTM. With LFs, different paradigmatic and syntagmatic relations between lexemes are described exhaustively. Such functions can be represented as ‘F(L₁) = L₂’, in which L₁ is called the key of LF, and L₂ its value (Nasr et al. Reference Nasr, Owen and Kittredge1998). For instance, in the function ‘Magn (sleep _v) = deeply, heavily, like a dog’, the key word is ‘sleep _v’, and the values ‘deeply, heavily, like a dog’ serve as intensifiers to the key word. That is to say, other words, such as ‘greatly, very’, which can stand for a modifier with the ‘very’ or ‘to a high degree’ meaning either, however, cannot satisfy the value of the ‘Magn (sleep _v)’ function.

According to former works on ECD (Mel'čuk Reference Mel'čuk, Embleton and Embleton1998, Reference Mel'čuk2000; Milićević Reference Milićević2006), we can draw the following conclusions for its major properties: (1) it strongly emphasizes lexicon; (2) it is semantics-based; (3) it is a synthesis dictionary; (4) it centers around restricted co-occurrences between syntactic and lexical of LUs; (5) it is an integral dictionary describing all information necessary to use a given LU in all possible contexts; and (6) it is a formalized dictionary. Thanks to ECD, not only can the machines transfer the entries from a source language into a particular target language as precisely as possible, but the language learners can also obtain useful language knowledge about the corresponding language, e.g. specific collocations, idioms, etc.

Although empirical MT models are ‘most effective for obtaining cost-effective, robust, and good quality translations in the shortest delay possible’ (Mel'čuk and Wanner Reference Mel'čuk and Wanner2008: 103), they ignore the syntactic and semantic properties of their operands. Most such models depend on statistical data like word co-occurrence from examples, or corpora, but leaves out traditional linguistic data. Thus, the qualities of translation results are always not high. On the contrary, MTT concentrates on the description of all linguistic information and models the language production process into a stratificational system. The minute description of linguistic information provides NLP scholars abundant knowledge about natural languages. The stratificational viewing angle of languages also does them a favor by learning the construction specifics of human natural languages. Till now MTT has succeeded in some engineering text-generation and MT systems, e.g. FoG, LFS, PedPro, ETAP-1, ETAP-2, etc. For an NLP scholar, it would be a worthy attempt to listen to the voice of MTT.