🏙️ Ottawa

About

Profile

Ottawa is a city region in Canada. On this site its research comes from a government body, the National Research Council Canada, and a university, the University of Ottawa. The National Research Council Canada produces most of the region's works here and draws the large majority of their citations. The University of Ottawa adds a smaller set of works that run from the late 2010s into the 2020s.

Among the works here, both institutions work in the same field: the computational study of language. The papers ask how machines can read opinion, emotion and stance in text, especially in social media posts and customer reviews. They also ask how text can be moved between languages through machine translation. On this site, then, Ottawa is a hub for natural language processing, where a government laboratory and a university work on related problems and sometimes on the same papers.

History

The region's research on this site begins in 2002 with Thumbs Up or Thumbs Down? by Peter Turney at the National Research Council Canada. The paper treats review classification as a question of semantic orientation, meaning whether the words in a text lean positive or negative. It needs no labelled training examples. This unsupervised approach to opinion is the earliest line among the works here, and later Ottawa papers keep returning to it.

From 2010 the focus moves from overall positive or negative judgements to particular emotions. Emotions Evoked by Common Words and Phrases, from 2010, aims to give emotion analysis resources closer in size to the polarity lexicons that already existed. A year later, Data-Driven Response Generation in Social Media connects the council's work to statistical methods from machine translation and applies them to conversational text.

In the middle of the 2010s the work becomes organised around shared evaluation tasks. In 2014 the council entered a SemEval task on aspect-based sentiment. In 2016 and 2018 Ottawa researchers ran SemEval tasks on stance and on affect in tweets. Machine translation also grows as a strand, as the council's part in the 2017 Conference on Machine Translation findings shows.

From 2017 the University of Ottawa appears more strongly. It contributes new stance datasets and a dependency-based model for aspect-level sentiment. In 2023 it adds a group of papers on dialectal Arabic that bring sentiment analysis and translation together.

Research profile

The largest field among Ottawa's works here is sentiment and emotion analysis, and Turney's 2002 paper is its starting point. Two later papers build lexical resources. The first is the 2010 study of emotions evoked by words and phrases. The second is the 2018 paper Obtaining Reliable Human Ratings of Valence, Arousal, and Dominance, which presents the NRC VAD Lexicon. It uses crowdsourced judgements to score a large English vocabulary on three affective dimensions: valence, arousal and dominance. Both resources treat emotion as something that can be measured word by word and in degrees, and not only as present or absent.

Aspect-based sentiment is a closely related topic. NRC-Canada-2014: Detecting Aspects and Sentiment in Customer Reviews describes the council's system for the SemEval-2014 task. The task asks which aspects of a product or service a review discusses and how the writer feels about each one. At the University of Ottawa, Aspect-Level Sentiment Analysis Via Convolution over Dependency Tree tackles the same problem by running graph convolutions over the grammatical structure of a sentence. Its design builds on earlier relation-extraction models that stacked a graph convolutional network above an LSTM.

Stance detection is a second major strand. It asks whether a writer is for or against a given target, and it is where the two institutions work most closely together. SemEval-2016 Task 6: Detecting Stance in Tweets presents the first shared task on stance in tweets and is credited to both institutions. The University of Ottawa extends this line with A Dataset for Multi-Target Stance Detection. That dataset is built from election tweets and comes with a large pool of unlabelled tweets. A later political corpus, P-Stance, collects English tweets from the 2020 United States presidential race.

Emotion intensity in social media connects the two strands. SemEval-2018 Task 1: Affect in Tweets changes the question from whether an emotion is expressed to how strongly it is expressed, and poses it as a shared task.

Machine translation is the third field. At the National Research Council Canada it appears in the WMT17 findings and in the phrase-pair techniques used for response generation. At the University of Ottawa it takes the form of work on informal Arabic. OSN-MDAD is a translation dataset for multi-dialectal Arabic conversations. It starts from the observation that people posting in Arabic online tend to use regional dialects more than Modern Standard Arabic. Two companion papers develop "content localization", which carries the meaning of a message into dialectal Arabic instead of translating it word for word. Content-Localization based Neural Machine Translation applies the idea to translation itself. Content-Localization based System for Analyzing Sentiment and Hate Behaviors uses it to turn English sentiment and hate-speech datasets into material for Arabic.

Leading scientists

At the National Research Council Canada, Saif Mohammad appears first. Mohammad's work here covers emotion lexicons, the VAD ratings, and the SemEval tasks on stance and on affect in tweets. Together these make up much of the region's most cited research. Colin Cherry comes next, as one of the council's researchers in a body of work that reaches into machine translation and social media text. Svetlana Kiritchenko is first author of the 2014 aspect-based sentiment system for customer reviews and is connected to both Ottawa institutions on this site. Peter Turney wrote the 2002 paper on semantic orientation. George Foster, Xiaodan Zhu, Chi-kiu Lo and Roland Kuhn complete the council's group. Their work belongs to the council's translation and sentiment research, which includes its part in the WMT17 benchmarking campaign.

At the University of Ottawa, Parinaz Sobhani appears first and is lead author of the 2017 multi-target stance dataset. Richong Zhang and Diana Inkpen follow, among the university's researchers whose papers here cover stance and sentiment. Fatimah Alzamzami leads the 2023 papers on dialectal Arabic, content localization and translation. Abdulmotaleb El Saddik also appears at the university. After Svetlana Kiritchenko, who is named here as well, come Kai Sun, first author of the 2019 dependency-tree model for aspect-level sentiment, and John Healy.

Landmark works

The most cited of Ottawa's works in our library is Turney's 2002 paper on unsupervised review classification. It proposes a three-stage method that decides whether a review recommends its subject by looking at the semantic orientation of its phrases. Next comes the 2016 stance task, the first shared evaluation of stance in tweets. The 2018 Affect in Tweets task follows close behind. It was designed to capture how intense emotions are, and not only whether they are present.

Findings of the 2017 Conference on Machine Translation is another highly cited work. It reports on a yearly evaluation exercise that measures progress in translation systems and provides common test sets. The 2018 VAD lexicon paper, the 2010 emotion lexicon study and Ritter's 2011 work on response generation also draw many citations. Ritter's paper replaces noisy GIZA++ word alignments with a step that lists every candidate phrase pair and keeps those that pass a filter. The 2014 aspect-based sentiment system is also widely cited among the works here.

Labs and units

The region's research on this site is concentrated in its two member institutions. The National Research Council Canada is the government laboratory behind most of the lexicons, shared tasks and translation work. The University of Ottawa contributes stance datasets, neural models for aspect-level sentiment, and the Arabic content-localization research. Each institution has its own entry on this site, which describes its topics and units in more detail.

Collaborations

The clearest link between Ottawa's government laboratory and its university is the 2016 stance-detection task, which is credited to both. Svetlana Kiritchenko is connected to both institutions. The university's later stance datasets, from 2017 and 2021, continue the line of work that the shared task began. The council's research also feeds into international community efforts: the SemEval tasks it organised or entered and the WMT17 campaign. Through these, Ottawa's work is tied to wider communities of researchers in sentiment analysis and machine translation.

Timeline

The early 2000s are represented by Turney's 2002 work on semantic orientation, the region's starting point on this site.

From 2010 to 2014 the National Research Council Canada builds emotion lexicons. In 2011 it adapts translation methods to generating responses on social media, and in 2014 it enters the SemEval task on aspect-based sentiment.

Shared tasks dominate between 2016 and 2019. Stance in tweets comes in 2016. The WMT17 findings and the multi-target stance dataset follow in 2017, then affect in tweets and the VAD lexicon in 2018. The University of Ottawa's dependency-tree sentiment model appears in 2019.

The 2020s bring the P-Stance political corpus in 2021. In 2023 the University of Ottawa publishes its papers on dialectal Arabic, which combine translation with the analysis of sentiment and hate speech.