app.py

import json
import os
from dataclasses import dataclass, field
from typing import List, Optional, Dict
from PIL import Image

import numpy as np
import pandas as pd
import streamlit as st
from fsspec.implementations.local import LocalFileSystem
from huggingface_hub import HfFileSystem

import streamlit.components.v1 as components

@dataclass
class Field:
    type: str
    title: str
    name: str = None
    mandatory: bool = True
    # if value of field is in the list of those values, makes following siblings mandatory
    following_mandatory_values: list = False
    help: Optional[str] = None
    children: Optional[List['Field']] = None
    other_params: Optional[Dict[str, object]] = field(default_factory=lambda: {})

# Function to get user ID from URL
def get_param_from_url(param):
    user_id = st.query_params.get(param, "")
    return user_id


########################################################################################
# CHANGE THE FOLLOWING VARIABLES ACCORDING TO YOUR NEEDS

# 'local' or 'hf'. hf is for Hugging Face file system but has limits on the number of access per hour
filesystem = 'hf'
# path to repo or local file system TODO rename
input_repo_path = 'datasets/emvecchi/annotation'
output_repo_path = 'datasets/emvecchi/annotation'
# filesystem = 'local'
# path to repo or local file system
# input_repo_path = '/data/mod-gen-eval-pref'
# output_repo_path = '/data/mod-gen-eval-pref'

to_annotate_file_name = 'to_annotate.csv'  # CSV file to annotate
COLS_TO_SAVE = ['comment_id','comment','confidence_score']

agreement_labels = ['strongly disagree', 'disagree', 'neither agree no disagree', 'agree', 'strongly agree']
quality_labels = ['very poor', 'poor', 'acceptable', 'good', 'very good']
priority_labels = ['not a priority', 'low priority', 'neutral', 'moderate priority', 'high priority']
yes_no_labels = ['yes','no']
yes_no_other_labels = ['yes','no','other']
default_labels = agreement_labels

function_choices = ['Broadening Discussion', 
                    'Improving Comment Quality', 
                    'Content Correction', 
                    'Keeping Discussion on Topic', 
                    'Organizing Discussion', 
                    'Policing', 
                    'Resolving Site Use Issues', 
                    'Social Functions', 
                    'Other (please specify)']
property_choices = ['appropriateness',
                    'clarity',
                    'constructiveness', 
                    'common good',
                    'effectiveness',
                    'emotion', 
                    'impact',
                    'overall quality', 
                    'proposal',
                    'Q for justification',
                    'storytelling', 
                    'rationality', 
                    'reasonableness', 
                    'reciprocity', 
                    'reference',
                    'respect',
                    'moderation behavior',
                    'Other (please specify)']
assistance_choices = ['Expand the breadth of moderator role',
                    'Reduce my own bias',
                    #'Assist with recall',
                    'Avoids me missing relevant instances',
                    'Improve speed of moderation tasks',
                    'Manage prioritization of comments to consider',
                    'Visualization of properties narrows down moderator contribution',
                    'Other (please specify)']
                    
default_choices = function_choices

consent_text = '''
## Consent Form

You will be asked to take part in a research study. Before you decide to take part in this study, it is important that you understand why the study is being done and what it involves. Please read the following information carefully.
________________________________________________________________________________________

Project title: Moderator Intervention Prediction\\
Researchers: E.M. Vecchi, N. Falk, I. Jundi, G. Lapesa\\
Institute: Institute for Machine Speech Processing (IMS)\\
University: University of Stuttgart\\
Contact: eva-maria.vecchi@ims.uni-stuttgart.de
_________________________________________________________________________________________

### Description of the research study

In this study, we investigate an approach to assist expert moderators in online discussion platforms by automatically identifying comments in need of moderation. The annotators' task is to evaluate whether a comment returned by our system are indeed requires moderator intervention, and assess the impact such a system would have on the task of moderation.

The intended use of the results of this study includes an analysis as well as processed versions of the collected data in the context of a publicly available scientific publication.

**Time required:** Your participation will take up to an estimated 8 hours. The time required may vary on an individual basis.

**Risks and benefits:** The risks to your participation in this online survey are those associated with basic computer tasks, including boredom, fatigue, mild stress, or breach of confidentiality. Some of the topics discussed in the online posts to be annotated may include violence, suicide or rape. The only benefit to you is the learning experience from participating in a research study. The benefit to society is the contribution to scientific knowledge

**Compensation:** You will be compensated for participating in this study. If you are interested, we will also be more than happy to share more information about our research with you.

**Voluntary participation:** Your participation in this study is voluntary. It is your decision whether or not to participate in this study. If you decide to participate in this study, you will be asked to confirm this consent form ("I agree."). Even after signing the consent form, you can withdraw from participation at any time and without giving any reason. Partial data will not be analysed.

**Confidentiality:** Your responses to this experiment will be anonymous. Please do not share any
Information that can be used to identify you. The researcher(s) will make every effort to maintain your confidentiality.

**Contact:** If at any time you have questions about this study or would like to report any adverse effects due to this study, please contact the researcher(s).

**Trigger Warning:** The texts included in this study are produced in an online debate forum and some topics that are discussed, how they are discussed, and user perspectives may be uncomfortable or sensitive. First, all texts included here do not represent the views of the researchers conducting the study. Secondly, we provide the option [described in detail in the guidelines provided in the next step] to avoid having to annotate any instance that is problematic or uncomfortable for the annotator without penalty of compensation.

### Consent:
Please indicate, in the box below, that you are at least 18 years old, have read and understood this consent form, are comfortable using the English language to complete the survey, and you agree to participate in this online research survey.

- *I am age 18 or older.*
- *I have read this consent form or had it read to me.*
- *I am comfortable using the English language to participate in this survey.*
- *I agree to participate in this research and I want to continue with the survey.*
'''
guidelines_text = 'Please read <a href="https://acrobat.adobe.com/id/urn:aaid:sc:EU:1a1347b0-3423-49ee-aa28-87679f8a69c0">the guidelines</a>'
study_code = 'CE552C7F'
failed_sanity_check_code = 'C17NFPW5'
redirect_url = f'https://app.prolific.com/submissions/complete?cfc={study_code}'

annotation_guidelines_fields: List[Field] = [
    Field(name="annotation_guidelines", type="radio", title="Did you read the guidelines?", mandatory=True,
          other_params={'labels': ['Yes, in detail, and I understand the study',
                                   'Yes, in detail, but still confused',
                                   'Yes, I skimmed it',
                                   'I will read it later',
                                   'No, not interested in reading them',
                                   'I can not open the link',
                                   ],
                        'accepted_values': [0]}),
]

intro_fields: List[Field] = [
    Field(type="container", title="**Introductory Questions**", children=[
        Field(name="intro_moderation_goals", type="textarea", title="As a moderator, what are your goals/objectives for the comment section?"),
        Field(name="intro_experience", type="textarea", title="What do you feel contributes to a good experience for the users/discussion?"),
        Field(name="intro_valuable_comment", type="textarea", title="What makes a comment or contribution valuable?"),
        Field(name="intro_bad_comment", type="textarea", title="What makes a comment or contribution of poor quality, unconstructive or detrimental to the discussion?"),
    ]),
]

end_fields: List[Field] = [
    Field(type="container", title="**Thank you for taking part in this study!**", children=[
        Field(name="email", type="textarea", title="Enter an email address where we can send you the voucher. **Without this, we cannot compensate you for your contribution to our research.**"),
    ]),
]


fields: List[Field] = [
    Field(name="topic", type="input_col", title="**Topic:**"),
    Field(type="expander", title="**Preceeding Comment:** *(expand)*", children=[
        Field(name="parent_comment", type="input_col", title=""),
    ]),
    Field(name="comment", type="input_col", title="**Comment:**"),
    Field(name="image_name", type="input_col", title=""),# "**Visualization of high contributing properties:**"),

    Field(type="container", title="**Need for Moderation**", children=[
        Field(name="to_moderate", type="y_n_radio",
                   title="Do feel this comment/discussion would benefit from moderator intervention?", mandatory=True),
        Field(name="priority_level", type="likert_radio",
                   title="With what level of **priority** would you need to interact with this comment?", other_params={'labels': priority_labels},
                     mandatory=True),
    ]),

    Field(type="container", title="**Moderation Function**", children=[
        Field(name="mod_function", type="multiselect",
                     title="What type of moderation function is needed here? *(Multiple selection possible)*",
                     mandatory=True),
        Field(name="mod_function_other", type="text", title="*If Other, please specify:*", mandatory=False),
    ]),

    Field(type="container", title="**Contributing properties**", children=[
        Field(name="relevant_properties", type="multiselect",
                title="Which property(s) is most impactful in your assessment? *(Multiple selection possible)*", 
                other_params={'choices': property_choices}, mandatory=True),
        Field(name="relevant_properties_other", type="text", title="*If Other, please specify:*", mandatory=False),
    ]),
    
    Field(type="container", title="**Moderator Assistance**", children=[
        Field(name="helpful", type="y_n_radio",
                title="If this comment/discussion was flagged to you, would it be helpful in your task of moderation?", mandatory=True, following_mandatory_values=[0]),
        Field(name="mod_assistance", type="multiselect",
                title="If yes, please motivate the benefit it would contribute to the task. *(Multiple selection possible)*", 
                other_params={'choices': assistance_choices}),
        Field(name="mod_assistance_other", type="text", title="*If Other, please specify:*", mandatory=False),
    ]),
    
    Field(type="container", title="**Other**", children=[
        Field(name="other_comments", type="text", title="Please provide any additional details or information: *(optional)*", mandatory=False),
    ]),
]
INPUT_FIELD_DEFAULT_VALUES = {'slider': 0,
                              'text': '',
                              'textarea': '',
                              'checkbox': False,
                              'radio': None,
                              'select_slider': 0,
                              'multiselect': None,
                              'likert_radio': None,
                              'y_n_radio': None}
SHOW_HELP_ICON = False
SHOW_VALIDATION_ERROR_MESSAGE = True

########################################################################################
if filesystem == 'hf':
    HF_TOKEN = os.environ.get("HF_TOKEN_WRITE2")
    print("is none?", HF_TOKEN is None)
    hf_fs = HfFileSystem(token=HF_TOKEN)
else:
    hf_fs = LocalFileSystem()

def get_start_index():
    if hf_fs.exists(output_repo_path + '/' + get_base_path()):
        files = hf_fs.ls(output_repo_path + '/' + get_base_path())
        return len(files) - 2
    else:
        return -3

def read_data():
    with hf_fs.open(input_repo_path + '/' + to_annotate_file_name) as f:
        return pd.read_csv(f)

def read_saved_data():
    _path = get_path()
    if hf_fs.exists(output_repo_path + '/' + _path):
        with hf_fs.open(output_repo_path + '/' + _path) as f:
            try:
                return json.load(f)
            except json.JSONDecodeError as e:
                print(e)
    return None
    

# Write a remote file
def save_data(data):
    if not hf_fs.exists(f"{output_repo_path}/{get_base_path()}"):
        hf_fs.mkdir(f"{output_repo_path}/{get_base_path()}")
    with hf_fs.open(f"{output_repo_path}/{get_path()}", "w") as f:
        f.write(json.dumps(data))


def get_base_path():
    return f"{st.session_state.user_id}"

def get_path():
    return f"{get_base_path()}/{st.session_state.current_index}.json"


def display_image(image_path):
    with hf_fs.open(image_path) as f:
        img = Image.open(f)
        st.image(img, caption='8 most contributing properties', use_column_width=True)
        
#################################### Streamlit App ####################################

# Function to navigate rows
def navigate(index_change):
    st.session_state.current_index += index_change
    # only works consistently if done before rerun
    js = '''
    <script>
            var body = window.parent.document.querySelector(".main");
            body.scrollTop = 0;
            window.scrollY = 0;
    </script>
    '''
    st.components.v1.html(js, height=0)
    # https://discuss.streamlit.io/t/click-twice-on-button-for-changing-state/45633/2

    # disable text input enter to submit
    # https://discuss.streamlit.io/t/text-input-how-to-disable-press-enter-to-apply/14457/6
    components.html(
        """
        <script>
        const inputs = window.parent.document.querySelectorAll('input');
        inputs.forEach(input => {
            input.addEventListener('keydown', function(event) {
                if (event.key === 'Enter') {
                    event.preventDefault();
                }
            });
        });
        </script>
        """,
        height=0
    )

    #st.experimental_rerun()
    st.rerun()

def show_field(f: Field, index: int, data_collected):
    if f.type not in INPUT_FIELD_DEFAULT_VALUES.keys():
        st.session_state.following_mandatory = False
        match f.type:
            case 'input_col':
                value = st.session_state.data.iloc[index][f.name]
                if value and value is not np.nan:
                    st.write(f.title)
                    if f.name == 'image_name':
                        display_image(os.path.join(input_repo_path, 'images', value))
                    else:
                        st.write(value)
            case 'markdown':
                st.markdown(f.title)
            case 'expander' | 'container':
                with (st.expander(f.title) if f.type == 'expander' else st.container(border=True)):
                    if f.type == 'container':
                        st.markdown(f.title)
                    for child in f.children:
                        show_field(child, index, data_collected)
            case 'skip_checkbox':
                st.checkbox(f.title, key=f.name, value=False)
    else:
        key = f.name + str(index)
        st.session_state.data_inputs_keys.append(f.name)
        value = st.session_state[key] if key in st.session_state else \
            (data_collected[f.name] if data_collected else INPUT_FIELD_DEFAULT_VALUES[f.type])
        if not SHOW_HELP_ICON:
            f.title = f'**{f.title}**\n\n{f.help}' if f.help else f.title

        validation_error = False

        # form is not displayed for first time
        if st.session_state.form_displayed == st.session_state.current_index:
            if f.mandatory or st.session_state.following_mandatory:
                if st.session_state[key] == INPUT_FIELD_DEFAULT_VALUES[f.type]:
                    st.session_state.valid = False
                    validation_error = True
                elif f.following_mandatory_values and st.session_state[key] in f.following_mandatory_values:
                    st.session_state.following_mandatory = True

        
            # check for any unaccepted values
            if (
                    (f.other_params.get('accepted_values') and
                     value not in f.other_params.get('accepted_values')) or
                    (f.other_params.get('accepted_values_per_sample') and
                     index in f.other_params.get('accepted_values_per_sample') and
                     value not in f.other_params.get('accepted_values_per_sample').get(index))
            ):
                st.session_state.unacceptable_response = True

        if f.mandatory or st.session_state.following_mandatory:
            f.title += " :red[* required!]" if (validation_error and not SHOW_VALIDATION_ERROR_MESSAGE) else' :red[*]'
        f.help = None

        
        match f.type:
            case 'checkbox':
                st.checkbox(f.title,
                            key=key,
                            value=value, help=f.help)
            case 'radio':
                st.radio(f.title,
                            ["yes","no"],
                            key=key,
                            help=f.help)
            case 'slider':
                st.slider(f.title,
                          min_value=0, max_value=6, step=1,
                          key=key,
                          value=value, help=f.help)
            case 'select_slider':
                labels = default_labels if not f.other_params.get('labels') else f.other_params.get('labels')
                st.select_slider(f.title,
                                 options=[0, 20, 40, 60, 80, 100],
                                 format_func=lambda x: labels[x // 20],
                                 key=key,
                                 value=value, help=f.help)
            case 'multiselect':
                choices = default_choices if not f.other_params.get('choices') else f.other_params.get('choices')
                st.multiselect(f.title,
                                options = choices,
                                format_func=lambda x: x,
                                key=key, max_selections=3,
                                default=value, help=f.help)
            case 'likert_radio':
                labels = default_labels if not f.other_params.get('labels') else f.other_params.get('labels')
                st.radio(f.title,
                         options=[0, 1, 2, 3, 4],
                         format_func=lambda x: labels[x],
                         key=key,
                         index=value, help=f.help, horizontal=True)
            case 'y_n_radio':
                labels = yes_no_labels if not f.other_params.get('labels') else f.other_params.get('labels')
                st.radio(f.title,
                         options=[0, 1],
                         format_func=lambda x: labels[x],
                         key=key,
                         index=value, help=f.help, horizontal=True)
            case 'text':
                st.text_input(f.title, key=key, value=value, max_chars=None)
            case 'textarea':
                st.text_area(f.title, key=key, value=value, max_chars=None)

        if validation_error:
            st.session_state.unacceptable_response = False
            st.error(f"Mandatory field")


def show_fields(fields: List[Field]):
    st.session_state.valid = True
    index = st.session_state.current_index
    data_collected = read_saved_data()
    st.session_state.data_inputs_keys = []
    st.session_state.following_mandatory = False

    for field in fields:
        show_field(field, index, data_collected)

    submitted = st.form_submit_button("Submit")
    if submitted:
        if 'unacceptable_response' in st.session_state and st.session_state.unacceptable_response:
            prep_and_save_data(index, ('skip' in st.session_state and st.session_state['skip']))
            st.rerun()
        skip_sample = ('skip' in st.session_state and st.session_state['skip'])
        if not skip_sample and not st.session_state.valid:
            st.error("Please fill in all mandatory fields")
            # st.rerun() # filed-out values are not shown otherwise
        else:
            with st.spinner(text="saving"):
                prep_and_save_data(index, skip_sample)
            st.success("Feedback submitted successfully!")
            navigate(1)

    st.session_state.form_displayed = st.session_state.current_index

def prep_and_save_data(index, skip_sample):
    save_data({
        'user_id': st.session_state.user_id,
        'index': st.session_state.current_index,
        **(st.session_state.data.iloc[index][COLS_TO_SAVE].to_dict() if index >= 0 else {}),
        **{k: st.session_state[k + str(index)] for k in st.session_state.data_inputs_keys},
        'skip': skip_sample
    })

# st.set_page_config(layout='wide')
# Title of the app
st.title("Moderator Intervention Prediction")


st.markdown(
    """<style>
div[data-testid="stMarkdownContainer"] > p {
    font-size: 1rem;
}
    section.main > div {max-width:60rem}
    </style>
    """, unsafe_allow_html=True)


def add_annotation_guidelines():
    st.write(f"username is {st.session_state.user_id}")
    st.markdown(
        "<details open><summary><b>Annotation Guidelines</b></summary>" + guidelines_text + "</details><br>"
        , unsafe_allow_html=True)


if 'unacceptable_response' in st.session_state and st.session_state.unacceptable_response:
    add_annotation_guidelines()
    st.error("You are not eligible for this study. Thank you for your time!" +
             ("" if st.session_state.current_index < 0 else
              #" You will receive a small compensation as explained in the guidelines. "
              "Please email eva-maria.vecchi@ims.uni-stuttgart.de for issues or questions."
             ))
    st.stop()


# Load the data to annotate
if 'data' not in st.session_state:
    st.session_state.data = read_data()

# user id
user_id_from_url = get_param_from_url("user_id")
if user_id_from_url:
    st.session_state.user_id = user_id_from_url

# current index
if 'current_index' not in st.session_state:
    start_index = get_start_index()
    st.session_state.current_index = start_index
    st.session_state.form_displayed = -3

if get_param_from_url('show_extra_fields'):
    fields += url_conditional_fields


def add_validated_submit(fields, message):
    st.session_state.form_displayed = st.session_state.current_index
    if st.form_submit_button("Submit"):
        if all(not x for x in fields):
            st.error(message)
        else:
            navigate(1)


def add_checked_submit():
    check = st.checkbox('I agree', key='consent')
    add_validated_submit([check], "Please agree to give your consent to proceed")

if 'go_to' not in st.session_state:
    st.session_state.go_to = 0
    
if st.session_state.current_index == -3:
    with st.form("data_form"):
        st.markdown(consent_text)
        add_checked_submit()

elif st.session_state.current_index == -2:
    if st.session_state.get('user_id'):
        navigate(1)
    else:
        with st.form("data_form"):
            st.session_state.user_id = st.text_input('User ID', value=user_id_from_url)
            add_validated_submit([st.session_state.user_id], "Please enter a valid user ID")


elif st.session_state.current_index == -1:
    add_annotation_guidelines()
    with st.form("intro_form"):
        show_fields(intro_fields)

elif st.session_state.current_index < len(st.session_state.data):
    add_annotation_guidelines()
    with st.form("data_form"+str(st.session_state.current_index)):
        show_fields(fields)

elif st.session_state.current_index == len(st.session_state.data):
    with st.form("intro_form"):
        show_fields(end_fields)
        
else:
    st.write(f"Thank you for taking part in this study!")

# Navigation buttons
if st.session_state.current_index > 0:
        if st.button("Previous"):
            navigate(-1)

if 0 <= st.session_state.current_index < len(st.session_state.data):
    st.write(f"Page {st.session_state.current_index + 1} out of {len(st.session_state.data)}")


def navigate_to_index():
    try:
        go_to_index = int(st.session_state.go_to)-1
        farthest_index = get_start_index()
        # Check if the index is within the valid range
        if 0 <= go_to_index < len(st.session_state.data):
            if go_to_index <= farthest_index:
                st.session_state.current_index = go_to_index
                st.experimental_rerun()
            else: 
                st.info(f"You have not annotated up to instance #{int(st.session_state.go_to)} yet. \nYour last annotated instance was #{farthest_index+1}.")
        else:
            st.info(f"{int(st.session_state.go_to)} is not in the range of data.")
    except ValueError:
        st.error("The page number must be an integer.")

col1, col2, col3 = st.columns([2, 2, 2])
with col2:
    with st.form(key="nav_form"):
        st.number_input("Go to Page: ", key="go_to", format="%.0f")
        if st.form_submit_button("Go"):
            navigate_to_index()

st.markdown(
    """<style>
    div[data-testid="InputInstructions"] {
        visibility: hidden;
    }
    </style>""", unsafe_allow_html=True
)