This is the repository of the Python (3.10+) implementation of EAD-CogBot (Expressive Abductive-Deductive Cognitive ChatBot) a framework able to instantiate chatbots for Abductive-Deductive reasoning on a first-order logic knowledge base, but endowed of the expressive power of LLMs. EAD-CogBot is built on top on the cognitive architecture AD-CASPAR based on Natural Language Processing, whose paper is published in Cognitive Systems Research.
This repository has been tested with the following packages versions:
- Phidias (release 1.3.4.alpha)
- SpaCy (2.2.4)
- Natural Language Toolkit (3.5)
- python-telegram-bot (12.8)
- pymongo (3.10.1)
- pandas (2.1.4)
> git clone https://github.com/corradosantoro/phidias
> pip install -r requirements.txt
> pip install .> pip install spacy
> python -m spacy download en_core_web_trffrom prompt:
> pip install nltkfrom python console:
> import nltk
> nltk.download('wordnet')
> nltk.download('omw-1.4')> pip install python-telegram-bot==12.8> pip install pymongo-
Install a new Mongodb community instance from here (a GUI Compass installation is also recommended from here), then create a new database named ad-caspar containing a collection clauses (the easier way is by using Compass). The url of the MongoDB server must be specified by changing the value of HOST (section LKB) in config.ini.
-
Create a new mongodb user in the Mongo shell (or Compass shell) as it follows:
> use ad-caspar
> db.createUser({
user: "root",
pwd: "example",
roles: [
{ role: "readWrite", db: "ead-cogbot" }
]
})In the case of using a mongoDB container, the latter can be accessed by the link: http://localhost:8087/ (user/password are set in config.ini).
> docker-compose -f mongo.yaml up> pip install pandas
> pip install openpyxlFollow the instructions reported here for the current system.
- Download Llama-2-7b-chat-hf (or 70b) from huggingface and copy it in a local folder (BASE_MODEL in [LLM] Section of config.ini). The other two adapters path must be set as well (ADAPTER_PATH1 and ADAPTER_PATH2 in [LLM] Section of config.ini). Both adapters finetuning's code is in the folder "finetuning" of this repository.
> pip install transformers==4.34.0
> pip install peft==0.4.0
> pip install sentencepiece==0.1.99
> pip install datasets==2.13.0
> pip install accelerate==0.23.0
> pip install bitsandbytes==0.41.1
> pip install trl==0.4.7
> pip install safetensors>=0.3.1
> pip install scipy> pip install huggingface_hubBefore going any further it is first necessary to create a new telegram bot by following the instruction in this page. The returned token must be put in TELEGRAM_TOKEN (AGENT Section) in config.ini.
> python ead-cogbot.py
PHIDIAS Release 1.3.4.alpha (deepcopy-->clone,micropython,py3)
Autonomous and Robotic Systems Laboratory
Department of Mathematics and Informatics
University of Catania, Italy (santoro@dmi.unict.it)
eShell: main >In section [LLM] of config.ini, the parameter MODE must be set as follows:
- AD (Abduction-Deduction with FOL-to_NL response)
- LLM (only Query/Answer LLM)
- DUAL (AD+LLM) -----> to be added soon!
Sentence can be processed also outside chatbot environment, by using proc(X) (X=sentence). In any case, assertions must end with "." and questions must end with "?". Every other text will be processed by the LLM.
eShell: main > proc("Joe Biden is the President of United States.")
Asserting definite clause into Fol Kb.
Is_VBZ(Joe_NNP_Biden_NNP(x1), Of_IN(President_NNP(x2), United_NNP_States_NNP(x3)))
eShell: main > hkb()
Is_VBZ(Joe_NNP_Biden_NNP(x1), Of_IN(President_NNP(x2), United_NNP_States_NNP(x3)))
1 clauses in Higher Knowledge Base
eShell: main > proc("Who is Joe Biden?")
Result: {v_27: x10, x11: Of_IN(President_NNP(v_28), United_NNP_States_NNP(v_29))}
Generating llm text from FOL....
The President of the United StatesWhen a question falls out defined hot topic, EAD-CogBot will begin the response with the custom prefix Well...I am not sure, but.... (which can be set with by parameter PREFIX_LLM_RESP in [LLM] Section of config.ini), and after it will attach generated text from LLM.
eShell: main > proc("Who is Barack Obama?")
Generating llm text....
Well...I am not sure, but....Barack Obama is a former President of the United States.eShell: main > proc("tell me a joke")
Generating llm text....
Why was the math book sad? Because it had too many problems.In case of LKB_USAGE = true in [LKB] section of config.ini, Low Knowledge Base can be inspected as follows:
eShell: main > lkb()
0 clauses in Low Knowledge Base
eShell: main >both High KB e Low KB can be emptied with the following commands:
eShell: main > chkb()
High Clauses kb initialized.
0 clauses deleted.
eShell: main > clkb()
Low Clauses kb initialized.
0 clauses deleted.
eShell: main >The Telegram chatbot must be started by shell command go() as follows:
eShell: main > go()
EAD-CogBot started! Bot is running in AD mode...To start a session you have to go to the telegram bot window and type the word "hello".
Assertions must end with "." and questions must end with "?".
After such interaction with the telegram bot, the two layers of the Clauses KB will be as it follows:
eShell: main > hkb()
eShell: main > Became_VBD(Barack_NNP_Obama_NNP(x1), Of_IN(President_NNP(x2), United_NNP_States_NNP(x3)))
1 clauses in High Knowledge Base
eShell: main > lkb()
Became_VBD(Barack_NNP_Obama_NNP(x1), Of_IN(President_NNP(x2), United_NNP_States_NNP(x3)))
['Became_VBD', 'Barack_NNP_Obama_NNP', 'Of_IN', 'President_NN', 'United_NNP_States_NNP']
Barack Obama became the president of United States.
1 clauses in Low Knowledge BaseQuerying chatbot on hot topic...
Querying chatbot on non-hot topic wll let LLama-2-chat to get an answer from its implicit knowledge...
This prototype gives the change to feed automatically the Clauses KB from file text, set with the parameter FILE_KB_NAME (AGENT section), by the means of the command feed() given in the phidias prompt. Three examples knowledge base of increasing size are available for testing purposes: west25.txt, west104.txt, west303.txt (inside kbs folder). The command feed() must be executed after chatbot awakening with the word hello at startup.
> feed()A syntax for a single input argument X is also available, for instance:
> feed("Colonel West is American")This prototype gives the change to export Low Clauses KB content (clauses and corresponding sentences) into a excel file whose name must be set in FILE_EXPORT_LKB_NAME (AGENT section), with the command expt() given in the phidias prompt.
> expt()A detailed overview of how the wh-questions are treated is provided here. In case of LKB usage, after a bot reboot the result will be slightly different because the High Clauses KB will be empty and must be populated getting clauses from the Low Clauses KB, taking in account of a confidence level about the presence of the lemmatized labels in the clauses. Such a confidence level, depending of the domain can be changed by modifying the value of MIN_CONFIDENCE (LKB Section) in config.ini. The first query will get a result form the Low KB (From LKB: True), while the second one from the High KB (From HKB: True); thats because the content of the High KB is preserved during the session, otherwise it can be emptied after a query by changing the value of EMPTY_HKB_AFTER_REASONING (LKB Section) in config.ini.
In order to test the Nested Reasoning, where knowledge is combined to obtain human-like reasoning, you must be sure some parameters in config.ini are as it follows:
(achieving all useful clause via generalization-based knowledge base expansion)
Section [NL_TO_FOL]:
- ASSIGN_RULES_ADMITTED = true
- CONDITIONAL_WORDS = WHEN
- DEAJECT_NOM = false
Section [REASONING]
- NESTED_REASONING = true
Section [POS]
- OBJ_JJ_TO_NOUN = true
Section [GEN]
- GEN_PREP = true
- GEN_ADJ = true
- GEN_ADV = true
(achieving all useful clause via deajectival-nominalization-based knowledge base expansion)
Section [NL_TO_FOL]:
- ASSIGN_RULES_ADMITTED = true
- CONDITIONAL_WORDS = WHEN
- DEAJECT_NOM = true
Section [REASONING]
- NESTED_REASONING = true
Section [POS]
- OBJ_JJ_TO_NOUN = true
Section [GEN]
- GEN_PREP = true
- GEN_ADJ = false
- GEN_ADV = true
ASSIGN_RULES_ADMITTED ar used for creating logic implication starting from a copular verbs (be, present tense), while CONDITIONAL_WORDS (when, if, while, etc.) are those for what we want a logic implication will be asserted.
It is well-known that natural language can be ambiguous, subject to interpretation about the semantic role of each lexical parts. For such a reason out-of-common sense utterance might lead to unexpected logical forms, due to the dataset the dependency parser has been trained on. Still, as reported here, the model used for dependency parsing has an accuracy of 0.95, which means that some missful/wrong dependecy classification is expected, especially for longer sentences. Beyond that, the following are known issues related to the code in this repository:
- Anaphora resolution/coreferentiators are not included yet in this code. So it is recommended to not use sentence containing pronoms, otherwise any abductive/deductive operations won't be successful. For this purpose, the integration of tools such as neuralcoref is planned. Coders might include such a tool in their own fork of this repository.
- Sentence containing singles quotation marks (') are still not well managed. So, it is recommended to not use it, and, in such a case, to rephrase utterances differently.
- Occasional crashes during parsing of text may occur, especially during conversion from natural language into logical forms/definite clauses. In this case, rephrasing/reducing utterances is recommended.
- IMPORTANT: all production rules are designed starting from a toy domain. Practical use involves a global review of all parsing rules and modules.