File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/73/c73-1005_metho.xml
Size: 27,987 bytes
Last Modified: 2025-10-06 14:11:04
<?xml version="1.0" standalone="yes"?> <Paper uid="C73-1005"> <Title>EBERHARD PAUSE A CLASS OF TRANSFOKMATIONAL RECOGNITION GRAMMARS</Title> <Section position="1" start_page="0" end_page="0" type="metho"> <SectionTitle> EBERHARD PAUSE A CLASS OF TRANSFOKMATIONAL RECOGNITION GRAMMARS </SectionTitle> <Paragraph position="0"> 1. PETmCK considers transformational grammars (T-grammars) of a special* form which essentially have the properties described by N.</Paragraph> <Paragraph position="1"> CHOMSKY (1965).</Paragraph> <Paragraph position="2"> a) The base grammar is context-free. One recursive element S is distinguished. The base trees in general have the form s s or in linear notation Consequently each base tree consists of a finite set of subtrees or kernel trees each of the form S(@ $1(...)@) where x is a string Over i I shall assume the reader somewhat familiar with the theory of generative transformational grammar. : -:</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="metho"> <SectionTitle> 46 EBERHARD PAUSE </SectionTitle> <Paragraph position="0"> V -{@}.~ A string in which maximally n kernel trees stand one above the other has the depth (of embedding) n.</Paragraph> <Paragraph position="1"> b) Transformational rules (T-rules) are singulary or binary. The singulary ones operate on a subtree, called constituent tree, whose initial node has the label St and whose terminal string contains no sentence boundary symbol @. The binary T-rules work on a constituent tree and the kernel tree (called the matrix tree) which dominates it directly.</Paragraph> <Paragraph position="2"> c) The transformational rules are linearly ordered and are applied cyclically in this order from bottom to top. That is, a cycle is completed after the singulary and binary rules have been applied to a constituent tree. (In the process the sentence boundary symbols of this subtree are removed).</Paragraph> <Paragraph position="3"> Now Petrick defines a class of T-grammars generating recursive languages by first stating the condition of the &quot;recoverability of deletions &quot;. That is, only * terminal node or a subtree which is identical with another subtree that remains in the resulting tree after the application of the T-rule can be deleted. He states further conditions so that there is an upper bound for the depth of embedding of a base tree underlying a sentence of the grammar. In his case that means, that a base tree that underlies a sentence of length n can have maximally depth n.</Paragraph> <Paragraph position="4"> A recognition grammar for a T-grammar of this class is constructed by Petrick roughly in the following way: Transformational derivations are generated, and starting with the trees that constitute these derivations he determines context-free rules, so-called &quot;auxiliary rules &quot;, reflecting the structure of the derived trees. By means of these auxiliary rules a given string can be attributed as much structure as is necessary for applying the &quot;inverse &quot; T-rules. If the string is a sentence of the given grammar one can obtain the underlying base tree(s) by this procedure.</Paragraph> <Paragraph position="5"> 2. The following problems arise in connection with the class of transformational grammars defined by Petrick: a) To my knowledge there is no finite procedure to determine whether a given T-grammar has the defining properties of the class or not, which represents a solution to the problem whether there exists V means the entire vocabulary, Y~ the terminal, ~ the nonterminal vocabulary of the base.</Paragraph> <Paragraph position="6"> A CLASS OF TRANSFORMATIONAL RECOGNITION GRAMMARS 47 in principle some recognition procedure for the grammar. There also remains the question of whether this problem is recursively solvable at all.</Paragraph> <Paragraph position="7"> b) In general there is no finite number of auxiliary rules implied by the derivations of a T-grammar of this form. For that reason Petrick can only construct auxiliary rules for finitely many derivations. These rules can then be used only for the analysis of sentences whose underlying base trees have maximally depth n (for a given n)whose length is therefore smaller than or equal to n. For sentences containing more than n words a new recognition grammar has to be constructed. There may also be sentences of length smaller than or equal to n which cannot even be analysed because the depth of their underlying base trees exceeds the specified boundary.</Paragraph> <Paragraph position="8"> c) The set of trees accepted by the auxiliary rules is larger than the one generated by the original grammar. As a consequence spurious trees result at the end of the analysis process and these mustbe discarded by an additional synthesis phase.</Paragraph> <Paragraph position="9"> These problems suggest that it might be reasonable to look at the following requirements when using a grammar for the recognition of sentences. It should belong to a (nontrivial) class of grammar having the following properties: a) All generated languages are recursive.</Paragraph> <Paragraph position="10"> b) It is decidable whether a given grammar belongs to the class or not.</Paragraph> <Paragraph position="11"> c) There should be a general procedure which costructs for each grammar of this type a practicable recognition grammar. That is, analysis with such a grammar should be as efficient as possible.</Paragraph> <Paragraph position="12"> I do not wish to claim that these requirements have all been met for the type of grammar that I worked with. 8 But I want to show now what results I obtained with a different approach to the matter and what difficulties one obviously faces when dealing with transformational recognition.</Paragraph> <Paragraph position="13"> 3. In what follows I will first of all illustrate a general method by which the set of base trees, given by some arbitrary context-free grammar can be decomposed into a finite set of kernel trees. This construction serves (in section 4) as a basis for the definition of a class of T-grammars that generate recursive languages. A hierarchy of types of</Paragraph> </Section> <Section position="3" start_page="0" end_page="0" type="metho"> <SectionTitle> 3 For details see my dissertation (E. PAUSE, 1972). 48 EBERHARD PAUSE </SectionTitle> <Paragraph position="0"> grammar will be characterized, included in this class, such that for each type membership of a given grammar is decidable. Section 5 deals with recognition grammars.</Paragraph> <Paragraph position="1"> The whole matter is discussed with respect to T-grammars of a generalized form: a) Every context-free grammar can serve as base grammar. b) Transformational rules are unordered. There is no distinction between singulary and binary rules. 4 We may assume that a context-free grammar P-~ (V, Y,, R, S) where V-lg = ~, the sentence symbol S in ~, is reduced and has a standard form. s That is, R contains only rules of the form (I) S ~ ~, (II) A ~ D1 ... D,, n _> 2, A in ~, D~ in l ~ - {S},otherwise,if S ~ ~. is in R e (III) A ~ a, a in \]g.</Paragraph> <Paragraph position="2"> There is no loss of generality on considering only context-free grammars of this form since for every context-free grammar an equivalent standard form grammar could effectively be given. A recursion sequence (of length k) of a standard form context-free grammar P is defined as a sequence of rules from R</Paragraph> <Paragraph position="4"> where A~ is in ~. u,, v, in V*, v Ao = A, and A~ :~ A i (for i C/: j) otherwise. Each symbol A i occuring in the sequence is a recursive element in P. Since there are only finitely many elements in * every recursion</Paragraph> </Section> <Section position="4" start_page="0" end_page="0" type="metho"> <SectionTitle> 4 In E. PAUSE (1972) I also investigate T-grammars that cdegntain generalized transfor- </SectionTitle> <Paragraph position="0"> mations of roughly the same form as discussed by N. CHOMSK'Y (1957). Concerning the form of the T-rules I refer the reader to the sample grammar in section 4.</Paragraph> <Paragraph position="1"> For details see for instance S. GrNSBURG (1966).</Paragraph> <Paragraph position="2"> e ~. denotes the empty string.</Paragraph> <Paragraph position="3"> V* denotes the free semigroup (generated by V).</Paragraph> <Paragraph position="5"> sequence of P can maximally have length m ~-/O/s and since! R is a finite set, the number of recursion sequences of P is finite. : Two recursion sequences are equivalent if they could be ideni~ified by cyclical permutation of the rules. This leads to the concept o~: the equivalence class or recursion cycle of a recursion sequence. Exactly k recursion sequence of length k which could be transformed into one another: constitute a recursion cycle. Therefore each recursion cycle is uniquely determined by specifying one recursion sequence belonging to it.</Paragraph> <Paragraph position="6"> Now all recursion cycles of a standard form context-free grammar P = (V, ~g, R, S), where t =/0/, could be enumerated in the following way: Let us consider all sequences of rules from R of the form</Paragraph> <Paragraph position="8"> r~,: Ak~l --> ukX, v, where r~, is in R, A s in * and Aj ~ A,~ (for-j ~ m), then after maximally k _~ t steps either a) X, is in ~g or b) X, is equal to Aj for some j smaller than k. In case (b) the sequence of rules rii ..... ri,, noted as above, obviously represents some recursion sequence of P. Clearly, there are only finitely many sequences as given above in P and it is easy to see that in all these sequences (which correspond in some sense to all derivations in P of length smaller than or equal to t) there is to be found at least one representative of each recursion cycle of P.</Paragraph> <Paragraph position="9"> Now equivalent recursion sequences are identified until there re-mains exacdy one representative for each cycle. Furthermore, one arbitrary recursive element of each recursion cycle is chosen. Let {C1, .... Cs} be the set of these symbols each of which is called a base symbol of P together with the sentence symbol S. Then consider all occurences of base symbols on the right hand side of the rules of P as terminal elements which could not further be expanded by some rule. Construct s /~/denotes the cardinality of ~.</Paragraph> <Paragraph position="11"> bols: you will obtain a finite set of trees B with terminal string either in ~,~&quot; or in {Ig u Z}* having the following properties: a) every base tree corresponding to some sentence derivation in :P can be uniquely decomposed into elements of B (in the obvious man.ner).</Paragraph> <Paragraph position="12"> b) If elements of B are embedded into one another by identifying each time some initial node with some terminal node having as label the same base symbol the resulting tree is always a subtree of a tree assC/,ciated with some sentence derivation in P.</Paragraph> <Paragraph position="13"> Hence the elements of B are kernel trees in almost the same sense as are those occuring in base trees considered by Petrick. I will not give here a proof of my main statements but will illustrate them by an example.</Paragraph> <Paragraph position="14"> Let P = (V, Z, R, S) where Ig ~- ( a, d, f, i, j, n, v, x, y }, =(C, D, E, F, F', H, L J, K, M, N, Q}, and R ={S~CF'EI, F'~XY, C~MJ, M~KI-t,K~DF, H~ QH, Q-~- AF, H~ NF, A ~ a,D~ d, F ~ f, E ~ v, I~ i,J~j, N~ n, X~ x, Y~ y}.</Paragraph> <Paragraph position="15"> Taken P as base grammar the base trees in general have the form s c ~ M-- ~u i Y\ ?\ &Y\ \\ !i\[ d f a f ... a f ,, f .I x \]' v i Now all rule sequences of length smaller than or equal to 12 = \[~\[, (of the form (~_)), starting with the sentence symbol S are A CLASS OF TRANSFORMATIONAL RECOGNITION GRAMMARS 51</Paragraph> <Paragraph position="17"> It can be immediately observed that there is only one recursion sequence; namely the rule H ~ QH in (2) which represents simultaneously the only recursion cycle of P. Now { H, S } is taken as the set of base symbols of P. Further following the given construction finally the kernel trees (a), (b), (c) result: (a) S(C(M(K(D(d)F(f))H)J(j))F'(X(x)Y(y))E(v)I(i)) (b) H(Q(A(a)F(f))H) &quot; (c) H(N(n)F(f)) It is easy to see that for instance taking the embedding sequence</Paragraph> <Paragraph position="19"> there is a tree associated with some sentence generation in P such that the sequence is a subtree of it. From the construction equally follows that each base tree can be decomposed into elements of B in the opposite manner as contructing trees using the members of B.</Paragraph> <Paragraph position="20"> Since there are three kernel trees in ~ embedded into one another we speak in exactly the same sense as above of the depth of a tree.</Paragraph> <Paragraph position="21"> Thus 0~ has depth 3. In the following we will also call every tree built up only of kernel trees a complex.</Paragraph> <Paragraph position="22"> 52 ~B~ vAus~ 4. In this section, I will first deal with some relevant implications of the foregoing construction and introduce the required terminology. Suppose, some T-rule ,r of a T-grammar G is applied to a base tree 0~: Let ~ be the resulting tree. Then ~ can also be decomposed into kernel trees, taking as cutting-points again those nodes labeled with base symbols. Some of the resulting kernel trees will perhaps have been deleted or will have been modified by the application of'r. In the latter case we get derived kernel trees. In any case, each tree that could be generated in G can be decomposed into (base or derived) kernel trees. I will call the set of all those kernel trees, thus obtained of all trees occuring in the derivations of G, the decomposition set B* of G. Suppose further that there are k terms in the structural description of-: which have been related to k nodes of 0~. Now the sequence of all and only those kernel trees (from left to right) in which these nodes lie will be called the characteristic domain of'r.</Paragraph> <Paragraph position="23"> The complex which is the smallest subtree of 0c such that it contains exactly the nodes, characterized by % (see the dashed line)is said to be a minimal complex of-~ (rel:/tive to 0t).</Paragraph> <Paragraph position="24"> Obviously only the minimal complex of a T-rule (relative to some tree) is relevant for the observation of the structural change produced by the rule, because all structure beyond it (above and below in the tree) is not directly concerned. To examine certain properties of transformational derivations, it could therefore be sufficient to construct derivations consisting of such complexes rather than of the whole trees. This can be done nearly in the following way: a) Relative to all T-rules of G, at first all possible minimal complexes of depth 1 are built up of the elements of B. The set of kernel trees, obtained after the decomposition of the complexes resulting from A CLASS OF TRANSFORMATIONAL RECOGNITION GRAMMARS 53 the application of the rules, yields together with set B set B1,1. This procedure can be repeated for minimal complexes over B of depth 2, 3, ... which gives the sets Bz,1, B3,,, ...</Paragraph> <Paragraph position="25"> b) Now continuations of the foregoing derivations (of length 1) could be constructed taking as basis each time the corresponding set Bi,z, Bi, z .... According to the resulting derivations of length 2, 3, ... the sets B1,2, B1,3, ..., B~,z, B~,3 .... of kernel trees will be obtained. By this procedure, which I cannot describe in detail here, the decomposition set B* of G can be enumerated as the union set of the sets Bp, q (p, q > 1).</Paragraph> <Paragraph position="26"> For our purposes, it is relevant to consider what happens with some kernel tree y or some node of a tree in the course of a derivation: performing some transformation, y is either deleted or there are some derived kernel trees (at least one) in the resulting tree which are the images of y, either identical or modified by the rule. Again the associated images of these in the following derived tree (which are also images of y) could be identified, and so on. This procedure could be carried out along the whole derivation starting with an element of B, called the origin, in the base tree. The same is valid for a node in a tree. It is either deleted or copied or only transferred by applying a T-rule, where in the latter cases it has some images (at least one) in the resulting tree 13 called the occurrences of the node in 13.</Paragraph> <Paragraph position="27"> Now let us consider some derivation in G. It starts with a base tree ~, and we are looking at some kernel tree y that is a subtree of 0c, and some derived tree 13 occurring in the derivation: a) All occurrences of terminal nodes of y in 13, labeled with terminal elements, Or occurrences of nodes in 13, inserted in y or an image of it by T-rules introducing new terminals (or morphemes), are called the rest-nodes of y in ~.</Paragraph> <Paragraph position="28"> b) Each terminal node of y labeled with a base symbol is called a base-node, and all occurrences of such nodes of y in 13 are said to be base-nodes of y in 13.</Paragraph> <Paragraph position="29"> After these remarks we are ready to outline the class of structurally bounded T-grammars. For a grammar G to be structurally bounded in particular the following conditions must hold: a) The condition of recoverability of deletions.</Paragraph> <Paragraph position="30"> b) If a rest-node of some kernel tree.~ in the characteristic domain of a T-rule $ relative to a tree 13 is deleted, the number of rest-nodes of 0~ in 13 must be greater than 1.</Paragraph> <Paragraph position="31"> c) If a subtree y is deleted by performing &quot;~, then each base-node</Paragraph> </Section> <Section position="5" start_page="0" end_page="0" type="metho"> <SectionTitle> $4 EBERHARD PAUSE </SectionTitle> <Paragraph position="0"> occurring in % and lying in some kernel tree of ~, may be an occurrence or a base-node of the origin of y.</Paragraph> <Paragraph position="1"> d) Let the number of terminal nodes of the origin of a kernel tree occurring in the characteristic domain of &quot;: be m. Then this subtree, following its changes under the rules up to ~ by considering its &quot;trace&quot; along its images in the given derivation, may not have occurred more than m-1 times in the characteristic domain of-r.</Paragraph> <Paragraph position="2"> For this class of grammars, say G1, the following theorem can be proved: (A) If G is structurally bounded, then the length of each sentence derived from some base tree ~ of depth n cannot be shorter than n, n the member of kernel trees contained in 0~ could maximally be ~; k ~-1 i=l (where k is a constant depending on G).</Paragraph> <Paragraph position="3"> Suppose G is structurally bounded, and m is the maximum of sub-trees which could be deleted on account of their identity by one T-rule of G. Lei.j be the number ofT-rules of G, and r the length of the longest terminal string of members in the set B of G, then let j.m.r = k. n ~--- 1: If0~ is a base tree of depth 1, then every tree ~ that could be derived from 0c may contain a rest-node of e. Thus the length of the terminal string of ~ is greater than or equal to n = 1, and the number of kernel trees in ~ is 1 = ~k ~-~.</Paragraph> <Paragraph position="4"> Now suppose (A) is true for all trees of depth h smaller than n (h greater than 1).</Paragraph> <Paragraph position="5"> Let ~ be a base tree of depth n. Without loss of generality, we may assume that 0~ has the form, where ~ is some kernel~tree, Y1, ..-, ~'~ subtrees maximally of depth n-l, and let us further assume that s does not exceed k. Then k-1 of these subtrees could be deleted maximally. Some rest node of ~ in a tree \[~ derived from 0~ must also survive. From this and by induction it A CLASS OF TRANSFORMATIONAL RECOGNITION GRAMMARS 55 follows that the length of the terminal string of \[3 must be greater or equal to n-1 + 1 = n. Hence the number of kernel trees in ~ has as maximum value k. (~ k iq) + 1 = 2f k iq.</Paragraph> <Paragraph position="7"> Obviously a decision procedure for the language generated by some grammar of G1 could be given. However, it can be proved that there is no recursive solution to the problem whether in general a given grammar belongs to G1 or not.</Paragraph> <Paragraph position="8"> In the sequel I shall go on to sketch some properties of grammars in G1 for which this problem is decidable.</Paragraph> <Paragraph position="9"> In considering the construc.tion of the decomposition set B* of a given T-grammar G and of the corresponding derivations of complexes (see page 52 and 53), there might exist some point where the following requirements are fulfilled: Suppose that all derivations starting with complexes up to a certain depth p have been constructed, that the length of these derivations has as yet reached some q, and that they are &quot;structurally bounded &quot;. Suppose further that by continuing the construction for some p' greater than p no new derivation (beginning with complexes over/3) could be started. Then G is structurally bounded and is said to be in class Gp,~, if the given derivations become periodical. That is, if they have the same continuations in the q + 1-th step as in the q- n + 1-th step (for some n), in the q + 2-th step as in the q-n + 2-th step, ..., in step q + n as in step q, and so on. This in particular means, that at those points in the derivations always the same minimal complexes appear periodically, and that therefore the decomposition set B* of G is finite. 9 Since there are grammars for which these properties hold for arbitrary _p and q, there exists an infinite hierarchy G2 of types of grammars Gp, q (p, q >1) in G1. For each p and q obviously membership of a given grammar in Gp,q could be determined in a finite number of &quot;steps.</Paragraph> <Paragraph position="10"> The following example should help in understanding these brief remarks: Let G be a T-grammar with base given by the ab0ve-mentioned context-free grammar P, and let G contain the following obligatory T-rules: There is great evidence in the assumption that every <~ kernel sentence * of a sentence of a natural language can undergo only a finite number of transformations.</Paragraph> </Section> <Section position="6" start_page="0" end_page="59" type="metho"> <SectionTitle> V 56 EBILRHAI~D PAUSE </SectionTitle> <Paragraph position="0"> rl: (~, y, v, i) 1, 2, 3-+ 3, 2, 01deg &quot;r2: (~, j, x, i, v) 1, 2, 3, 4 -+ 0,1, 3, 4 %: (K,H,F',v) 1,2,3,4-+ l+3,2,0,4 &quot;r4: (d,f, F', ~e)1, 2, 3,-+ 1,3,3, &quot;rs: (~, F', F', n,f, ~)1, 2, 3, 4, -+ l, 0, 3, 2 &quot;re: (M,F',F', Q, H, ~) 1, 2, 3, 4, ~ 1,3,3, &quot;rT: (M, a,f, F',~) 1, 2, 3-+ 1,3,3 Now the following derivations could be constructed:</Paragraph> <Paragraph position="2"> There is no difficulty in verifying that the continuations which could follow always start with the minimal complexes ~u and ~18. G is obviously structurally bounded and lies in G2,6. The following graph representing the order in the possible applications of the T-rules of G makes perhaps the periodicity of the derivations more transparent: grammar G of the class G m, namely G itself, using the general decision procedure for the language generated by G. This is, however, obviously not an efficient procedure but represents some general way of analysis by synthesis.</Paragraph> <Paragraph position="3"> I will now sketch a way of constructing recognition grammars for the class of grammars considered here which in most cases are more practicable. This will mainly be done by examining our sample grammar. I will finish with the discussion of some problems involved in the described procedure.</Paragraph> <Paragraph position="4"> Let us consider the T-grammar G given above: Analysis should start with some given string over V, say &quot;d j i n j i v &quot;. Now by reversing the generation process, the last transformation that has been applied in generating this sentence, &quot;rs: (~, F', F', n, f ~) 1, 2, 3, 4 --+ 1, 0, 3, 2, has now to be performed first. The inverse transformation &quot;:5 -1 (which can mechanically be constructed from ~5) would have the structural description (M, F', n, F', ~,1). However, there are no rules to attribute some structure to the given string such that there were nodes labeled with F which could be related to the corresponding terms in this inverse structural description. Base rules cannot do this work because the base structure has been modified by the applicafon of T-rules.</Paragraph> <Paragraph position="5"> Now by inspecting the derivations of complexes a term A of a structural description, could be ' expanded' in this way: the sequence of labels of the nodes which are dominated by a node related to A, such that the string of these labels is accepted by the inverse base rules, 1~ is substituted for A in .. This will be done, if possible (see below) and necessary, for each term and each T-rule relative to the different trees to which it has been applied in all derivations. In general, for each T-rule more than one 'expanded' rule results.</Paragraph> <Paragraph position="6"> ,s For each base rule of the form A-+ x, the rule x-~ A is the inverse. 58 ~B~.m~D I'AUS~.</Paragraph> <Paragraph position="7"> Since there are only finitely many different derived (minimal) complexes in the derivations of a grammar G of Gp,q, one obtains a grammar G' (called reversible) containing a finite set of expanded T-rules. G' is by construction equivalent to G. A recognition grammar G is obtained from G' by computing the corresponding inverse Trules. Then analysis will be performed by 'intermixed parsing ', that is, by alternating the application of base rules and T-rules. ~3 In order to prevent the expanded T-rules being applicable at some point in a derivation, where the original rules could not have been applied, auxiliary symbols are inserted in the rules to control their correct application.</Paragraph> <Paragraph position="8"> To illustrate these remarks, let us look at the expanded inverse T-rules of our sample grammar G: * i1: (~, i, v) 1, 2-+y, 2+1 .r~: (~,j d3,~, i, ~)) 1, 2, 3, 4-+1, x, 3, 4 .r~: (K,j,d,,~,i,H,v) l,...,6-+ l, 0, O,O, 5,2+d3,x+4+6 &quot;r\]~: (d,j, dsi, i,j, d52, i, ~) t -~&quot; (d,j,d~,i,j,c\[6~,i,~) t 1 ..... 7~ 1,0, O,f, 5, d4x, 7 ~742 * , , -~: (~,j, i, n,j, i, ~) 1, 2, 3, 4, 5~1, d5,1+2, 4+ d5,2+ +5+3, o,f &quot;ra~: (~,j,i,Q,j, dT,~,i,H,~)l .... ,7 ~ 1, de.1 + 2, 4 + d6,~ + -~. (C/~,a,j, ds,x,i,j, ds,2, i,~) l &quot;rT&quot;~i (C/~,a,j, d6,~,i,j, de,~,i,C/~) 1, ..., 7---~1, O,O,f, 5, d~x, 7 ~'72 * The construction should be clear in considering the original rules and the derivations given above. The symbols d~,,~ denote auxiliary symbols. For the rules % and &quot;rv two expanded rules have been constructed be-cause &quot;r4 and &quot;r7 precede the application of either the rule &quot;r6 or xe. The recognition procedure will be illustrated by the analysis of the sample sentence &quot; d j in j iv &quot;. Let t~ ~i, tt R denote the appliCation of the rule -r i or of the inverse base rules respectively. Notice that the inverse T-rules are not true inverse rules, since they apply to sequences of trees, so~alled &quot;terminal rest-trees&quot; x3 It can be shown that the set of context-semitive languages is properly included in the set of languages generated by grammars of G,. Furthermore, for each context-sensitive grammar a reversible T-grammar can effectively be given. A CLASS OF TRANSFORMATIONAL RECOGNITION GRAMMARS dj i nj i v d j ds,1 i j d~,2 i n f v dfj d4,~ i nfv</Paragraph> </Section> class="xml-element"></Paper>