SA-aa-S: A Final Year Project

Transcript

1. Sentiment Analysis as a Service A case study of Twitter Analytics Lochemem Bruno Michael United States International University - Africa
2. Part I: Background and Problem
3. Background Sentiment Analysis and related text-analysis techniques are available in many contexts Renowned projects like Google Cloud NLP and MonkeyLearn are popular SaaS options Custom NLP toolkits written in Python, Java, and Julia are also prominent The SA-aa-S project combines aspects of said Sentiment Insight acquisition techniques
4. Problem Statement A great deal of opinions are pulsed through the internet - via Social Media - with great regularity There exists a need to classify - via appropriate annotation - said opinions The vastness of diverse text patterns warrants additional text analyses (collocation, keyword extraction)
5. Part II: The Solution
6. What is your appraisal of this?
7. What the algorithm says Sentiment Positive Negative Neutral Compound mostly positive 0.247 0.097 0.657 0.8776 Compound score factors in intensity of all words
8. A Simple Project A SA-aa-S project which offers Sentiment, Collocation, and Keyword analyses of Twitter text The app polls Twitter for text data; analyzes said data and conveys chart visuals to users authenticated via the platform
9. Envisaged Benefits Automated text data consumption via daemon Quantification of word and emoticon intensity Algorithmic grading of phraseme relevance Algorithmic grading of collocations
10. Development Challenges Difficulty writing a fault-tolerant long-running daemon Implementing synchronous Twitter OAauth in asynhronous environment Finding a compatible, comprehensive text analysis tool Writing performant code at all app tiers
11. Salient Philosophies Functional Programming function jwtEncode(array $claims): string { $ret = _\compose( _\partial(_\extend, JWT_DEFAULT_CLAIMS), _\partialRight( Firebase\\JWT\\JWT::encode , JWT_SECRET ) ); return $ret($claims); }
12. More stuff Asynchronous programming for potent I/O ... public function searchTweets( array $opts = [] ): PromiseInterface { $ret = _\partialRight(self::trimTweet, false); return $ret($this->_search( searchTweets , _\extend( $opts, [ include_entities => false, result_type => mixed ]) )); } ...
13. Even more stuff State management for efficient data manipulation (Flux/Redux pattern) export let oauthLoginReducer = ( state = false, action ) => { switch (action.type) { case OAUTH_LOGIN_SUCCESS: return action.result case OAUTH_LOGOUT_SUCCESS: { const { result } = action if (has(result, [ token ]) && result.token) return false } default: return state } }
14. More complicated? Virtual DOM-based UI rendering const LoadingPane: React.FC<LoadingPaneType> = ({ comp, tabId }) => ( <TabPane tabId={tabId}> <Loading classes="w-5 mx-auto" /> {comp} </TabPane> )
15. How about some map-reduce? Tweet data tracking //map function (doc) { if (doc.created_at) { emit(new Date(doc.created_at).toDateString(), 1) } } //reduce function (keys, values, rereduce) { ... if (rereduce) { ret.total += values[acc].total ... } else { ret.total = sum(values) ... } ... }
16. NLP, Anyone? Collocations via PMI function collocation(string $text): array { $ret = (new TokensDocument(_tokenize($text))) ... ->applyTransformation(new PunctuationFilter( PUNCTUATION_LIST )) ->applyTransformation(new StopWordsFilter( STOPWORDS_LIST )) ... return (new CollocationFinder($ret->toArray(), 3)) ->getCollocationsByPmi(); }
17. Future Additions Parts of Speech Tagging Entity Tagging Developer Portal Pricing tiers