I am training my model which has following data but I am failing at intent classification. can anyone suggest why its happening and suggest better parameter for intent classifier to improve dection.
| intent | Examples of intents |
|---|---|
| budget_details | 13 |
| goodbye | 24 |
| affirm | 29 |
| deny | 36 |
| greet | 42 |
| product_submission | 2776 |
| request_waiver | 2976 |
| customer_details | 3011 |
| product_details | 4913 |
| product_information | 26516 |
Hi @shubham
this is a bit vague to be honest. Could you please provide some verbose information?
-vv flag and post the output here?Besides it seems a bit unbalanced where “a bit” means a lot. Is your classifier working but you are not satisfied with its performance? If so, please post your pipeline, such that we can analyze the problem properly!
Regards Julian
Hey @shubham, may I see the confusion matrix produced by rasa test nlu? Essentially, I want to see how exactly your intent classifier fails.
language: en
pipeline:
name: “SpacyNLP” model: “en_core_web_lg”
name: “SpacyTokenizer”
name: “SpacyFeaturizer”
name: “RegexFeaturizer”
name: DucklingHTTPExtractor url: http://localhost:8000 dimensions:
name: “CRFEntityExtractor” features: [[“low”, “title”,“pos”, “pos2”], [“bias”, “low”, “prefix5”, “prefix2”, “suffix5”,“digit”, “suffix3”,“suffix2”,“upper”, “title” ,“pattern”], [“low”, “title”, “upper”, “pos”, “pos2”]] BILOU_flag: true max_iterations: 50 L1_c: 0.1 L2_c: 0.1
name: “EntitySynonymMapper”
name: CountVectorsFeaturizer
OOV_token: OOV
token_pattern: (?u)\b\w+\b
name: “EmbeddingIntentClassifier”
“hidden_layers_sizes_a”: [256, 128]
“hidden_layers_sizes_b”:
“batch_size”: 10 “epochs”: 300
“embed_dim”: 20
“mu_pos”: 0.8
“mu_neg”: -0.4
“similarity_type”: “cosine”
“num_neg”: 20
“use_max_sim_neg”: true
“random_seed”: 50
“C2”: 0.002
“C_emb”: 0.9
“droprate”: 0.2
“intent_tokenization_flag”: true
“intent_split_symbol”: “+”
“evaluate_every_num_epochs”: 10
“evaluate_on_num_examples”: 1000
policies:
Hey @shubham, thanks for posting the pipeline, I think it will come handy a bit later.
For now, can, please, you explain what is not working and ideally provide the kind of information (either error messages or the intent confusion matrix, etc.) that would help me or @JulianGerhard help you, as we explained earlier? Thanks.