# Note 0705

## SRL
### Current idea
- capture data regularities encoded in dependency syntax
- some SRL relations occurs in specfic syntactic structure
- v-predicate: SUBJ dependency : Noun $\rightarrow$ v-predicate: A0 semantic : Noun
- a simple biaffine classifier does not encode such tendencies explicitly
- dependency syntax encodes a sort of prior complementing simple biaffine parser
- encode the prior with a mixture model
- mixture model component can encode prior for each SDP pattern
- SDP pattern inducing similar semantic relations can be grouped together
- we effectively have several clusters, each of which encodes one kind of SRL tendency
### dependency-based SRL
#### Status: success
- ~0.6-1 F1 improvement over baseline
- Greatest improvement comes from separating (0, 1) SDP pattern
- Mixture model also recovers a similar cluster as i setted
- (0,1)
- (1, 0) (1, 1)
- (2, 0)
- and others
#### other improvements
- Encoding information in shortest dependency path via LSTM
- SDP pattern shown above can be considered as counting
- LSTM can learn counting -> possibly generalize SDP patterns to a boarder range of syntactic features
- Also, we can use go for the LSTM over SDP idea
- !!! However, judging from score only, using LSTM seems to be worse

####
### span-based SRL
#### Status: fail
- almost no effect, or even worse than baseline
- The problem is that the strong regularities of dependency-SRL is lost in span-based SRL (when taking it as a sequence-tagging task)
- dependency pattern only show clue to head words
- span-based SRL requires tagging for the whole constituent (including the head and other words for that constituent)

0:DOCUMENT 1:PART 2:INDEX 3:WORD 4:POS 5:PARSE 6:LEMMA 7:FRAME 8:SENSE 9:SPEAKER 10:NE 11-N:ARGS N:COREF
bc/cctv/00/cctv_0001   0   0           This    DT  (TOP(S(NP*         -    -   -   Speaker#1        *   (ARG2*   (61
bc/cctv/00/cctv_0001   0   1            map    NN           *)        -    -   -   Speaker#1        *        *)   61)
bc/cctv/00/cctv_0001   0   2      reflected   VBD        (VP*    reflect  01   1   Speaker#1        *      (V*)    -
bc/cctv/00/cctv_0001   0   3            the    DT        (NP*         -    -   -   Speaker#1        *   (ARG1*     -
bc/cctv/00/cctv_0001   0   4       European    JJ           *         -    -   -   Speaker#1    (NORP)       *     -
bc/cctv/00/cctv_0001   0   5    battlefield    NN           *         -    -   -   Speaker#1        *        *     -
bc/cctv/00/cctv_0001   0   6      situation    NN           *))       -    -   -   Speaker#1        *        *)    -
bc/cctv/00/cctv_0001   0   7              .     .           *))       -    -   -   Speaker#1        *        *     -

- A better decoding method for span-based task will be needed
- [span-selection](https://aclanthology.org/D18-1191.pdf)
- chart-filling (maybe joint-constituency-spanSRL parsing?)
- ...