Last things to try for training

[Dr. Juan Miguel Cejuela]

• Word embeddings with spacy lemmas and lower case
• 11 labels for ST model, BIEO for all3 model --> BIEO
• feature.possible_states=BOOL to True for CRF
• word embeddings with Gaussian representation (code is not easly runnable, not enough time to make it work)
• Hyperparameters for word embeddings

[Aleksandar Bojchevski]

Spacy WE

these are the results with spacy WE

P:0.7950    R:0.8090    F:0.8019    0   exact
P:0.3144    R:0.3333    F:0.3236    1   exact
P:0.3853    R:0.3925    F:0.3889    2   exact
P:0.7136    R:0.7300    F:0.7217    TOTAL   exact

P:0.8922    R:0.8948    F:0.8935    0   overlapping
P:0.7667    R:0.7863    F:0.7764    1   overlapping
P:0.7319    R:0.7372    F:0.7345    2   overlapping
P:0.8641    R:0.8699    F:0.8670    TOTAL   overlapping

vs baseline old WE

P:0.7914    R:0.8036    F:0.7975    0   exact
P:0.3007    R:0.3262    F:0.3129    1   exact
P:0.3611    R:0.3645    F:0.3628    2   exact
P:0.7065    R:0.7234    F:0.7148    TOTAL   exact

P:0.8918    R:0.8926    F:0.8922    0   overlapping
P:0.7644    R:0.7990    F:0.7813    1   overlapping
P:0.7552    R:0.7660    F:0.7606    2   overlapping
P:0.8640    R:0.8713    F:0.8676    TOTAL   overlapping

[Aleksandar Bojchevski]

Spacy WE without post-processing

P:0.7750    R:0.7556    F:0.7652    0   exact
P:0.2762    R:0.2821    F:0.2792    1   exact
P:0.3763    R:0.3271    F:0.3500    2   exact
P:0.6936    R:0.6767    F:0.6850    TOTAL   exact

P:0.8857    R:0.8633    F:0.8744    0   overlapping
P:0.7481    R:0.7870    F:0.7671    1   overlapping
P:0.7734    R:0.7279    F:0.7500    2   overlapping
P:0.8581    R:0.8448    F:0.8514    TOTAL   overlapping

old WE without post-processing

P:0.7749    R:0.7551    F:0.7649    0   exact
P:0.2560    R:0.2679    F:0.2618    1   exact
P:0.3444    R:0.2897    F:0.3147    2   exact
P:0.6885    R:0.6727    F:0.6805    TOTAL   exact

P:0.8872    R:0.8635    F:0.8752    0   overlapping
P:0.7476    R:0.8010    F:0.7734    1   overlapping
P:0.7939    R:0.7482    F:0.7704    2   overlapping
P:0.8595    R:0.8479    F:0.8536    TOTAL   overlapping

[Aleksandar Bojchevski]

Spacy WE just on nala

P:0.8317    R:0.7900    F:0.8103    0   exact
P:0.4064    R:0.3551    F:0.3791    1   exact
P:0.4545    R:0.3968    F:0.4237    2   exact
P:0.6750    R:0.6207    F:0.6467    TOTAL   exact

P:0.9361    R:0.8952    F:0.9152    0   overlapping
P:0.9370    R:0.8214    F:0.8754    1   overlapping
P:0.8831    R:0.8095    F:0.8447    2   overlapping
P:0.9312    R:0.8600    F:0.8942    TOTAL   overlapping

OLD WE just on nala

P:0.8168    R:0.7876    F:0.8019    0   exact
P:0.4180    R:0.3692    F:0.3921    1   exact
P:0.4815    R:0.4127    F:0.4444    2   exact
P:0.6723    R:0.6250    F:0.6478    TOTAL   exact

P:0.9308    R:0.9026    F:0.9165    0   overlapping
P:0.9333    R:0.8235    F:0.8750    1   overlapping
P:0.8919    R:0.7952    F:0.8408    2   overlapping
P:0.9280    R:0.8637    F:0.8947    TOTAL   overlapping

[Dr. Juan Miguel Cejuela]

Closing. Las things here: #58