Please enable JavaScript.

Coggle requires JavaScript to display documents.

Deep Parsing in Watson (Evaluation (evaluated for parse correctness on two…

- - - - Each tree node N
        is centered on a headword, which is surrounded by its left and right modifiers, which are, in turn, tree nodes
      - Each modifier M of N fills a slot in N. The slot shows the grammatical role of M in N
      - the modifier structure is the surface structure of the parse analysis
    - - complement slots - determined by the properties of their headword (e.g verb which take subjects and objects)
        
        can name grammatical roles
        
        can name logical arguments of word senses
      - adjunct slots - determined by the part of speech of the headword sense (e.g. verbs can take an adverb)
        
        can typically modify any node of the category (POS) with which they are associated
    - - 5 main ingredients
        
        (3) The (logical or deep) argument frame of the node
        
        consists of the list of complement slots of the word sense, each slot being associated with its filler node (or nil if it has no filler
        e.g. derive(17, u, 15, 12) which translates into pairs (subj: nil, obj: ph15, comp:ph12)
        
        the slot-fillers represent the logical arguments, e.g. for verb the first argument is the logical subject, i.e. for passive forms the logical subject should be represented, not the grammatical one
        
        the set of word-sense predications form the core of the SG deep structure since they are close to logical predications
        
        (4) The features of the node
        
        can be morphosyntactic or semantic features
        
        the first feature listed is the POS
        
        Most of the features come from those of the headword sense, as obtained from morpholexical analysis of the headword
        
        some may be added during syntactic analysis when the node acquires modifier
        
        semantic features contain deep structure information
        
        (2) The ID of the node
        
        integer that is the word number of the headword in the segment
        
        for multiwords - the ID is the word number of the head of the multiword
        
        in the word-sense predication display, the first argument is the ID
        e.g. derive(17, u, 15, 12)
        
        (5) The (surface) modifier structure for the node
        
        a node N has two associated lists - for its left modifiers (premodifiers) and its right modifiers (postmodifiers), where each modifier node is paired with the slot it fills in N
        
        The slot shown closest to the headword of a node N is the slot that N fills in its mother node
        
        For each complement slot S, the slot option used for S is shown in parentheses after S
        
        (1) The headword of the node
        
        the parse data stores several versions:
        a) the form as it occurs in the text
        b) lemma
        c) the SG word sense
        
        normally a single-word token, but can also be a multiword, a punctuation, or a special node (e.g. quotation node)
        
        if represented as a predicate followed by arguments - words-sense predication - which is the main vehicle for showing the deep structure of the parse
        e.g. derive(17, u, 15, 12)
      - deep and surface structure
        
        deep
        
        The core of the SG deep structure is the set of word-sense predications described in (3) above,
        
        Deep structure information also exists in the semantic features of nodes
        
        some morphosyntactic features, such as tense and number, matter for logical form
        
        adjunct slots appearing in (5) can also be of relevance to deep structure e.g., determiners may produce quantifiers in logical form
        
        surface
        
        The core of the surface structure lies in the headword information (1) and (2), the morphosyntactic features, and the surface modifier structure (5)
  - - - Index word in lemma form (talk)
      - a sequence of sense frames (3 in the example)
        
        can specify any of the 7 items, all of which are optional except the first one
        
        word-sense name
        
        numerical score
        
        may reward or penalize uses of the current sense frame
        
        features - both semantic and syntactic
        
        ~ 160 semantic types, mainly high-level (e.g. physical object; substance; abstraction; types for time, location, and measures; collection; living being; human etc.)
        
        subject area test
        
        if provided can allow/exclude the current sense or give it a numerical score
        
        complement slot frame
        
        for verbs initial subject slot is implied
        
        a slot has an associated list of possible (slot)
        options, which are symbols that name constraints on the possible fillers of the slot
        
        most slot are optional, obligatory slots are marked
        
        generalized support verb construction
        
        using a special keyword to store the use of support verbs under the noun (e.g. the support verb "make" stored under "reference" for the structure "make a reference")
        
        POS - 15 possible tags (v)
      - Example:
        talk < v (obj n (p about)) (comp (p to with))
        < v obj1 (comp1 (p into))
        < n nsubj (nobj n (p about)) (ncomp (p to with))
    - - Noun-verb correspondences
        
        E.g. verb defer has indicated noun-forms deferral, deference,
        etc.
      - Chunk lexicons
        
        Handle multiword entries
        
        proper nouns derived from Wikipedia and the Unified Medical Language System
      - Augmentation of ESG base lexicon by WordNet
        
        increasing the number of entries
        
        marking the augmented lexicon systematically with semantic types
        
        new entries consist of proper nouns and certain multiword common nouns
      - lexical answer types (LAT) reward features
        
        Aid in identification of answer types
      - Matching noun frames with verb frames
        
        encoding verb-related nouns in a way that properly corresponds to verb frames
        
        e.g. encode a relationship between “celebration and
        “celebrate”
        
        can help in relation extraction or in matching questions to answers
  - - - a modifier phrase fills a slot in an adjacent matrix phrase (on the left or right)
      - slots tried are those from the available (complement) slots list (ASL) of the matrix and from the adjunct slots associated with the POS of the matrix
      - all slots of the matrix are tried, nondeterministically; for each such slot, all of its options are tried, normally in an if-then-else way
      - a numerical scoring system, where each (intermediate or final) phrase is assigned a parse score
        
        the final parses of a segment are numerically ranked
        
        during parsing, intermediate phrase analyses can be pruned away from the parse space (chart)
      - On the basis of (partial) capitalization and possible quoting, SG can form quote nodes, which are proper noun nodes, but show an internal syntactic structure underneath