Weakly Supervised Induction of Affective Events by Optimizing Semantic Consistency
Authors: Haibo Ding, Ellen Riloff
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our research investigates the prevalence of affective events in a personal story corpus, and introduces a weakly supervised method for large scale induction of affective events. We present an iterative learning framework that constructs a graph with nodes representing events and initializes their affective polarities with sentiment analysis tools as weak supervision. The learning algorithm iteratively refines the polarity values by optimizing semantic consistency across all events in the graph. Our model learns over 100,000 affective events and identifies their polarities more accurately than other methods. |
| Researcher Affiliation | Academia | Haibo Ding, Ellen Riloff School of Computing University of Utah Salt Lake City, UT 84112 {hbding, riloff}@cs.utah.edu |
| Pseudocode | Yes | Algorithm 1 Iterative Learning Algorithm 1: Input: W sim, W opp, W cmp, v0, c0 2: Output: v Rn |L| 3: while v has not converged do 4: Update vt using Eq. 2 5: Update ct using Eq. 3 6: end while 7: return vt |
| Open Source Code | No | The paper states: 'We plan to construct a lexicon of the affective events learned by the SC model and make it freely available to the research community.' This indicates a future plan for a lexicon, not current availability of the source code for the methodology. |
| Open Datasets | Yes | As the text corpus, we used the ICWSM 2009 and 2011 Spinn3r data sets1, which together contain over 177 million blog posts. ... 1http://www.icwsm.org/data/ |
| Dataset Splits | Yes | Of these 1,490 manually annotated events, we randomly selected 1,000 as our test set for evaluation and use the remaining 490 events as a development set for tuning parameters. |
| Hardware Specification | No | The paper does not mention any specific hardware used for running the experiments (e.g., GPU models, CPU types, or cloud computing resources with specifications). |
| Software Dependencies | Yes | We evaluated the performance of five existing affective lexicons: MPQA (Wilson, Wiebe, and Hoffmann 2005), Senti Word Net3.0 (Senti WN) (Baccianella, Esuli, and Sebastiani 2010)... |
| Experiment Setup | Yes | In our experiments, our full model uses the following values: β = 0.6, γ = 0.8, η = 0.1, which were selected on our development data. ... We used μ=0.1 in experiments based on the development set. ... We found that the following values achieved the best F1 scores on our development data and were therefore used throughout our experiments: λ=0 for MPQA, λ=0.25 for Conno Frame S, λ=0.3 for Conno Frame O, λ=0.4 for Conno WN, λ=0.5 for Senti WN, and λ=0.6 for +/-Effect WN. ... where α=0.7 based on the development set. |