Semantics: Distributional and Compositional. Dudes and PROPS

(I haven't posted any thing in a long while, the stuff is accumulating in a hazardous way. Today we had Gabi Stanovsky visiting and his talk was great, and it reminded me of posting this.)

There is by now a great deal of literature on the deep problem of unifying distributional semantics (in terms of vectors and cosine distances) and logical or compositional semantics (in terms of negation, conjunction, disjunction, implication, etc.) Because it is an interesting and very topical problem (several of the people involved have sold multi-million dollar companies, for example) several groups have tried to crack the problem, with different theories.

The vision paper, explaining why we need "distributional semantics" as well as "logical semantics" is Combining Symbolic and Distributional Models of Meaning,   by Pulman and Clark. only 4 pages and well worth reading!

Then I  made a list of a few other papers that caught my attention and that might indicate a way forward for what I want to do.  My list:
1. Combined Distributional and Logical Semantics, Lewis and Steedman, 2013.
2. Transforming Dependency Structures to Logical Forms for Semantic Parsing, Reddy et al, 2016.
3. Flexible Semantic Composition with DUDES, Cimiano, 2009.
4. Getting More Out Of Syntax with PROPS, Stanovsky et al, in arXiv on 4 March 2016.

These two last papers form a side trip from the main concern of merging distributional semantics and logical semantics, but are still about meanings. The DUDES is fairly short, old (2009) and the author seems to be more concerned with lexical resources nowadays. The PROPS paper is longer and seems much more useful to my goals. (also, isn't props a great name?)

The basic  ideas of the paper  seem to be:

1. NLP applications often rely on dependency trees to recognize major elements of the proposition structure of sentences.
2. many phenomena are not easily read out of dependency trees, often leading to ad-hoc heuristic post-processing or  information loss.
3. they suggest  PROPS – an output representation designed to explicitly and uniformly express much of the proposition structure which is implied from syntax.
4. they also provide an associated tool for extracting it from dependency trees (yay!!). 

(Project page at PropS -- Syntax Based Proposition Extraction, with online demo.
code in GitHub (GitHub - gabrielStanovsky/props: PropS offers an output representation designed to explicitly and uniformly express much… ) requires python and java 7.

Their desiderata:
a. uniformly represent propositions headed by different types of predicates, verbal or not.
b. canonicalize different syntactic constructions that correspond to the same proposition structure
c. decouple independent propositions while clearly marking proposition boundaries
d. "mask" non-core syntactic detail, yielding cleaner compact structures.
e. enable simple access to the represented propositions by a uniform graph traversal.

Their design principles: 
a.  Want to mask non-core syntactic detail:  
    - remove auxiliary words and instead encode their syntactic function as features; 
    - group atomic units (such as noun compounds) within a single node
b. Represent propositions in a uniform manner (verbal and adjectival)
c. Canonicalize and differentiate syntactic constructions: 
   - Unify the representation of propositions which are semantically equivalent;
    - Differentiate syntactically- similar, yet semantically-different, constructions.
d. Mark proposition boundaries
e. Propagate Relations: every relation which is inferable through parse tree traversal (for instance, through conjunctions) should be explicitly marked in the representation. 

Their output format:
1. similar to dependencies, BUT
2. Typed nodes: (1) Predicates, which evoke a proposition and 

   (2) Non-predicates, which can be either arguments or modifiers.

3. simplify the graph structure by allowing multi-word nodes (e.g., Barack Obama), versus having each node corresponding to a single word in dependency trees.
4. resulting structures are no longer limited to trees, but are DAGS.
5. a label set of 14 relations (compared with approximately 50 in Stanford dependencies) 

I need to check how Bridge/XLE deals with the pair: “The director who edited ‘Rear Window’ released Psycho” and
“Hitchcock, who edited ‘Rear Window’, released Psycho”. Need also to check and mark what  they call raising verbs?
They say [...]``we heuristically use a set of approximately 30 verbs which were found by (Chrupała and van Genabith, 2007) to frequently occur in raising constructions. For these verbs do not produce a proposition." Seems sensible to me and I don't think we did this in Bridge.


Evaluation:
MCTest corpus for machine comprehension (Richardson et al., 2013), composed of 500 short stories, each followed by 4 multiple choice questions. The MCTest comprehension task does not require extensive world knowledge. Focus on questions which are marked in the corpus as answerable from a single sentence in the story (905 questions followed by 3620 candidate answers). Richardson et al (2013) introduce a lexical matching algorithm, which they adapt to use either dependency or PROPS structures, both obtained using the Berkeley parser. (numbers show the progression expected, but still low).