is Section is dedicated to the VALLEX-twins, VALLEX 1.0 and PDT-VALLEX. Both lexicons have been elaborated within the FGD framework and their common features deeply affected the data structure of Swe-VALLEX, being its main source of inspiration. ough they have a lot in common, there are some differences between them, which do not allow for common description. A comparison of the two lexicons will be given right at the beginning of this section:
9.5.1 Differences between VALLEX and PDT-VALLEX
VALLEX has been built as a machine-readable lexicon for public use. PDT-VALLEX has been developed during the annotation of PDT (for details see [53]) as a supporting annotation tool. VALLEX includes only verbs, while PDT-VALLEX includes also nouns and adjectives.
Being independent of the corpus annotation work, VALLEX has been built pro- ceeding from lemma to lemma. A verb’s uses were investigated in text corpora and compared to information in existing lexicons, as a result of which deep valency frames
9.5. VALLEX, PDT-VALLEX 111 of the given verb were defined. e deep valency frame slot fillers were then completed with surface frame slot filler constraints, and the entire frame was enriched with ex- amples.
Unlike Vallex, PDT-VALLEX-frames were created in parallel with the data an- notation. When the annotated corpus did not contain a given reading of a verb, it simply remained missing in the lexicon. As more and more data had been processed, the PDT-VALLEX-frames were altered, in order to match the data in the best possi- ble way and to secure the highest possible consistency of annotation. As new valency frames of certain types of nouns and adjectives were arising at the later stages of PDT- annotation, they were entered into the lexicon.
9.5.2
Structure of VALLEX
e Czech valency lexicon describes the valency behaviour of each lexeme (verb, noun, adjective or adverb) in form of valency frames, which roughly correspond to senses. Like rolesets in the PropBank-Lexicon, the valency frames primarily rely on syntactic criteria though the syntactic criteria are sometimes modified to take account of the semantics of the given verb (see below).
A valency frame in the strict sense consists of inner participants and obligatory free modifications [120]. Free modifications are prototypically optional and do not belong to the valency frame in the strict sense, though some frames require a free modification as an obligatory slot in the frame (e.g. direction in verbs of movement, see Section 6.3.2). Both the obligatory and the optional inner participants belong to the valency frame in the strict sense. Like the free modifications, the inner par- ticipants have semantic labels according to the cognitive roles they typically perform: ACT (Actor), PAT (Patient), ADDR(Addressee), ORIG (Origin), and EFF (Effect). However, if a verb only has one inner participant, it is automatically labeled with ACT. A two-participant verb always has an ACT and a PAT (see Section 6.3.4).
PDT-VALLEX and VALLEX are very similar in structure: each lexeme corre- sponds to one entry. e entry is divided into valency frames. A valency frame is modeled as a sequence of frame slots. Each frame slot corresponds to one comple- mentation of the verb in question. Each slot is assigned a functor according to its semantic relation to the governing verb. Each slot includes an enumeration of its surface forms. Each frame is supported by at least one example sentence.
PDT-VALLEX notes only the valency frames in the strict sense (i.e. obligatory or optional inner participants and obligatory free modifications), while VALLEX also lists optional free modifications typical of the given frame. When delimiting the re- spective valency frames, syntactic as well as semantic criteria are adopted. erefore a verb can have two valency frames with identical distribution of functors. Lopatková [88] notes that “the change in morphemic realization signalizes the possibility of dif- ferent meanings; on the other hand, particular complementation in a valency frame
112CHAPTER 9. EXAMPLES OF VALENCY LEXICONS AND COLLOCATIONAL LEXICONS can have morphemic variants (if the meaning is ‘sufficiently close’)”.
Additional syntactic information is attached to each frame. It concerns mainly reflexivity and reciprocity conditions, verb control, aspectual characteristics of the given frame and its aspectual counterparts. Besides that, information on possible diatheses is added. Each frame is also classified on the scale of idiomaticity (values ‘primary usage’ – ‘secondary usage’ – ‘idiomatic usage’). Frames of ‘frozen collocations and idioms” [88] differ from ordinary frames in that they contain one slot with the functor CPHR or DPHR. DPHR is reserved “for the dependent parts of collocations with which the complementation is lexically limited to a single word (or to a restricted set of words) and the collocation cannot be syntactically analyzed. e CPHR functor is mainly used for marking predicate nouns in LVCs.
Some frames had been tentatively sorted into semantic-syntactic classes, out of which already 15 classes arose to be of help during consistency checking in VALLEX 1.0. e idea of Levin-like sorting of the verbs into classes based on the verbs’ syntac- tic and semantic behavior significantly affected the recent activities around VALLEX. Žabokrtský [166] suggested a new structure of VALLEX, which would capture reg- ular syntactic alternations. His point of departure was the fact that even very subtle changes of meaning, triggered by syntactic shifts, require their own frames. Conse- quently, the lexicon grows bigger than could be intuitively expected. e.g. each meme- ber of the pair of Levin’s example sentences would have different functors assigned, and thus it would require its own frame:
1. e sun.
ACT
radiates heat.PAT
2. Heat.
ACT
radiates from the sun.DIR1
.e drawbacks are obvious:
• ough the shift ‘ACT – DIR1 and PAT – ACT’ possibly applies to many verbs, it will not be explicitely indicated. ere will be no evidence that this shift is a regularly occurring phenomenon.
• Two frames instead of one common increases the space requirements of the lexicon.
• Every manually created frame can contain inconsistencies and/or errors. erefore, Žabokrtský suggests putting up a list of alternations to observe in clus- ters of lexical units. (In case of VALLEX 1.0 it means in clusters of the current frames). Whenever a pair of alternation realizations was found, only one member of the pair per alternation would be selected (in a more or less arbitrary way) to be explicitly stated in the lexicon as basic lexical unit (BLU). e other pair member would become its de-
9.6. FRAMENET 113